history table refactoring

Reviewed-on: https://git.deuxfleurs.fr/Deuxfleurs/garage/pulls/462

Cargo.lock

@@ -1080,7 +1080,6 @@ dependencies = [
  "parse_duration",
  "prometheus",
  "rand 0.8.5",
- "rmp-serde",
  "serde",
  "serde_bytes",
  "serde_json",
@@ -1156,7 +1155,6 @@ dependencies = [
  "hex",
  "opentelemetry",
  "rand 0.8.5",
- "rmp-serde",
  "serde",
  "serde_bytes",
  "tokio",
@@ -1200,7 +1198,6 @@ dependencies = [
  "netapp",
  "opentelemetry",
  "rand 0.8.5",
- "rmp-serde",
  "serde",
  "serde_bytes",
  "tokio",
@@ -1229,7 +1226,6 @@ dependencies = [
  "pnet_datalink",
  "rand 0.8.5",
  "reqwest",
- "rmp-serde",
  "schemars",
  "serde",
  "serde_bytes",
@@ -1243,6 +1239,7 @@ dependencies = [
 name = "garage_table"
 version = "0.8.1"
 dependencies = [
+ "arc-swap",
  "async-trait",
  "bytes",
  "futures",
@@ -1254,7 +1251,6 @@ dependencies = [
  "hexdump",
  "opentelemetry",
  "rand 0.8.5",
- "rmp-serde",
  "serde",
  "serde_bytes",
  "tokio",
@@ -1276,6 +1272,7 @@ dependencies = [
  "garage_db",
  "git-version",
  "hex",
+ "hexdump",
  "http",
  "hyper",
  "lazy_static",

Cargo.nix

@@ -32,7 +32,7 @@ args@{
   ignoreLockHash,
 }:
 let
-  nixifiedLockHash = "463114c4544bfa9b442a43afc6b39eb588f5720825c7a246ba9188c4bdb52944";
+  nixifiedLockHash = "b6aeefc112eb232904b24398f4e5da776c8ee2c13d427a26dbdf1732205d4fc9";
   workspaceSrc = if args.workspaceSrc == null then ./. else args.workspaceSrc;
   currentLockHash = builtins.hashFile "sha256" (workspaceSrc + /Cargo.lock);
   lockHashIgnored = if ignoreLockHash
@@ -1503,7 +1503,7 @@ in
       (lib.optional (rootFeatures' ? "garage/bundled-libs" || rootFeatures' ? "garage/default") "bundled-libs")
       (lib.optional (rootFeatures' ? "garage/consul-discovery") "consul-discovery")
       (lib.optional (rootFeatures' ? "garage/default") "default")
-      (lib.optional (rootFeatures' ? "garage/k2v") "k2v")
+      (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/k2v") "k2v")
       (lib.optional (rootFeatures' ? "garage/kubernetes-discovery") "kubernetes-discovery")
       (lib.optional (rootFeatures' ? "garage/lmdb") "lmdb")
       (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/metrics") "metrics")
@@ -1539,7 +1539,6 @@ in
       parse_duration = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".parse_duration."2.1.1" { inherit profileName; }).out;
       ${ if rootFeatures' ? "garage/default" || rootFeatures' ? "garage/metrics" || rootFeatures' ? "garage/prometheus" then "prometheus" else null } = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".prometheus."0.13.0" { inherit profileName; }).out;
       rand = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rand."0.8.5" { inherit profileName; }).out;
-      rmp_serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rmp-serde."0.15.5" { inherit profileName; }).out;
       serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde."1.0.137" { inherit profileName; }).out;
       serde_bytes = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde_bytes."0.11.5" { inherit profileName; }).out;
       structopt = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".structopt."0.3.26" { inherit profileName; }).out;
@@ -1569,7 +1568,7 @@ in
     registry = "unknown";
     src = fetchCrateLocal (workspaceSrc + "/src/api");
     features = builtins.concatLists [
-      (lib.optional (rootFeatures' ? "garage/k2v" || rootFeatures' ? "garage_api/k2v") "k2v")
+      (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/k2v" || rootFeatures' ? "garage_api/k2v") "k2v")
       (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/metrics" || rootFeatures' ? "garage_api/metrics") "metrics")
       (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/metrics" || rootFeatures' ? "garage_api/metrics" || rootFeatures' ? "garage_api/opentelemetry-prometheus") "opentelemetry-prometheus")
       (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/metrics" || rootFeatures' ? "garage_api/metrics" || rootFeatures' ? "garage_api/prometheus") "prometheus")
@@ -1639,7 +1638,6 @@ in
       hex = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".hex."0.4.3" { inherit profileName; }).out;
       opentelemetry = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".opentelemetry."0.17.0" { inherit profileName; }).out;
       rand = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rand."0.8.5" { inherit profileName; }).out;
-      rmp_serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rmp-serde."0.15.5" { inherit profileName; }).out;
       serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde."1.0.137" { inherit profileName; }).out;
       serde_bytes = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde_bytes."0.11.5" { inherit profileName; }).out;
       tokio = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".tokio."1.17.0" { inherit profileName; }).out;
@@ -1688,7 +1686,7 @@ in
     src = fetchCrateLocal (workspaceSrc + "/src/model");
     features = builtins.concatLists [
       [ "default" ]
-      (lib.optional (rootFeatures' ? "garage/k2v" || rootFeatures' ? "garage_api/k2v" || rootFeatures' ? "garage_model/k2v") "k2v")
+      (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/k2v" || rootFeatures' ? "garage_api/k2v" || rootFeatures' ? "garage_model/k2v") "k2v")
       (lib.optional (rootFeatures' ? "garage/lmdb" || rootFeatures' ? "garage_model/lmdb") "lmdb")
       [ "sled" ]
       (lib.optional (rootFeatures' ? "garage/sqlite" || rootFeatures' ? "garage_model/sqlite") "sqlite")
@@ -1710,7 +1708,6 @@ in
       netapp = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".netapp."0.5.2" { inherit profileName; }).out;
       opentelemetry = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".opentelemetry."0.17.0" { inherit profileName; }).out;
       rand = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rand."0.8.5" { inherit profileName; }).out;
-      rmp_serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rmp-serde."0.15.5" { inherit profileName; }).out;
       serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde."1.0.137" { inherit profileName; }).out;
       serde_bytes = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde_bytes."0.11.5" { inherit profileName; }).out;
       tokio = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".tokio."1.17.0" { inherit profileName; }).out;
@@ -1752,7 +1749,6 @@ in
       pnet_datalink = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".pnet_datalink."0.28.0" { inherit profileName; }).out;
       rand = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rand."0.8.5" { inherit profileName; }).out;
       ${ if rootFeatures' ? "garage/consul-discovery" || rootFeatures' ? "garage_rpc/consul-discovery" || rootFeatures' ? "garage_rpc/reqwest" then "reqwest" else null } = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".reqwest."0.11.12" { inherit profileName; }).out;
-      rmp_serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rmp-serde."0.15.5" { inherit profileName; }).out;
       ${ if rootFeatures' ? "garage/kubernetes-discovery" || rootFeatures' ? "garage_rpc/kubernetes-discovery" || rootFeatures' ? "garage_rpc/schemars" then "schemars" else null } = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".schemars."0.8.8" { inherit profileName; }).out;
       serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde."1.0.137" { inherit profileName; }).out;
       serde_bytes = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde_bytes."0.11.5" { inherit profileName; }).out;
@@ -1769,6 +1765,7 @@ in
     registry = "unknown";
     src = fetchCrateLocal (workspaceSrc + "/src/table");
     dependencies = {
+      arc_swap = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".arc-swap."1.5.0" { inherit profileName; }).out;
       async_trait = (buildRustPackages."registry+https://github.com/rust-lang/crates.io-index".async-trait."0.1.52" { profileName = "__noProfile"; }).out;
       bytes = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".bytes."1.2.0" { inherit profileName; }).out;
       futures = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".futures."0.3.21" { inherit profileName; }).out;
@@ -1780,7 +1777,6 @@ in
       hexdump = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".hexdump."0.1.1" { inherit profileName; }).out;
       opentelemetry = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".opentelemetry."0.17.0" { inherit profileName; }).out;
       rand = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rand."0.8.5" { inherit profileName; }).out;
-      rmp_serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".rmp-serde."0.15.5" { inherit profileName; }).out;
       serde = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde."1.0.137" { inherit profileName; }).out;
       serde_bytes = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".serde_bytes."0.11.5" { inherit profileName; }).out;
       tokio = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".tokio."1.17.0" { inherit profileName; }).out;
@@ -1794,7 +1790,7 @@ in
     registry = "unknown";
     src = fetchCrateLocal (workspaceSrc + "/src/util");
     features = builtins.concatLists [
-      (lib.optional (rootFeatures' ? "garage/k2v" || rootFeatures' ? "garage_api/k2v" || rootFeatures' ? "garage_model/k2v" || rootFeatures' ? "garage_util/k2v") "k2v")
+      (lib.optional (rootFeatures' ? "garage/default" || rootFeatures' ? "garage/k2v" || rootFeatures' ? "garage_api/k2v" || rootFeatures' ? "garage_model/k2v" || rootFeatures' ? "garage_util/k2v") "k2v")
     ];
     dependencies = {
       arc_swap = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".arc-swap."1.5.0" { inherit profileName; }).out;
@@ -1808,6 +1804,7 @@ in
       garage_db = (rustPackages."unknown".garage_db."0.8.1" { inherit profileName; }).out;
       git_version = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".git-version."0.3.5" { inherit profileName; }).out;
       hex = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".hex."0.4.3" { inherit profileName; }).out;
+      hexdump = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".hexdump."0.1.1" { inherit profileName; }).out;
       http = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".http."0.2.8" { inherit profileName; }).out;
       hyper = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".hyper."0.14.18" { inherit profileName; }).out;
       lazy_static = (rustPackages."registry+https://github.com/rust-lang/crates.io-index".lazy_static."1.4.0" { inherit profileName; }).out;

flake.lock

@@ -10,17 +10,17 @@
         "rust-overlay": "rust-overlay"
       },
       "locked": {
-        "lastModified": 1666087781,
+        "lastModified": 1673262828,
-        "narHash": "sha256-trKVdjMZ8mNkGfLcY5LsJJGtdV3xJDZnMVrkFjErlcs=",
+        "narHash": "sha256-pDqno5/2ghQDt4LjVt5ZUMV9pTSA5rGGdz6Skf2rBwc=",
         "owner": "Alexis211",
         "repo": "cargo2nix",
-        "rev": "a7a61179b66054904ef6a195d8da736eaaa06c36",
+        "rev": "505caa32110d42ee03bd68b47031142eff9c827b",
         "type": "github"
       },
       "original": {
         "owner": "Alexis211",
         "repo": "cargo2nix",
-        "rev": "a7a61179b66054904ef6a195d8da736eaaa06c36",
+        "rev": "505caa32110d42ee03bd68b47031142eff9c827b",
         "type": "github"
       }
     },
@@ -57,17 +57,17 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1665657542,
+        "lastModified": 1673261889,
-        "narHash": "sha256-mojxNyzbvmp8NtVtxqiHGhRfjCALLfk9i/Uup68Y5q8=",
+        "narHash": "sha256-7trMsi0z7EfYNC/Nc5EtulvChBHQAo376XRICyWr89Q=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "a3073c49bc0163fea6a121c276f526837672b555",
+        "rev": "baed728abe983508cabc99d05cccc164fe748744",
         "type": "github"
       },
       "original": {
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "a3073c49bc0163fea6a121c276f526837672b555",
+        "rev": "baed728abe983508cabc99d05cccc164fe748744",
         "type": "github"
       }
     },

flake.nix

@@ -1,10 +1,10 @@
 {
   description = "Garage, an S3-compatible distributed object store for self-hosted deployments";
 
-  inputs.nixpkgs.url = "github:NixOS/nixpkgs/a3073c49bc0163fea6a121c276f526837672b555";
+  inputs.nixpkgs.url = "github:NixOS/nixpkgs/baed728abe983508cabc99d05cccc164fe748744";
   inputs.cargo2nix = {
     # As of 2022-10-18: two small patches over unstable branch, one for clippy and one to fix feature detection
-    url = "github:Alexis211/cargo2nix/a7a61179b66054904ef6a195d8da736eaaa06c36";
+    url = "github:Alexis211/cargo2nix/505caa32110d42ee03bd68b47031142eff9c827b";
     inputs.nixpkgs.follows = "nixpkgs";
   };
 

nix/common.nix

@@ -3,15 +3,15 @@ rec {
    * Fixed dependencies
    */
   pkgsSrc = fetchTarball {
-    # As of 2022-10-13
+    # As of 2023-01-09
-    url = "https://github.com/NixOS/nixpkgs/archive/a3073c49bc0163fea6a121c276f526837672b555.zip";
+    url = "https://github.com/NixOS/nixpkgs/archive/baed728abe983508cabc99d05cccc164fe748744.zip";
-    sha256 = "1bz632psfbpmicyzjb8b4265y50shylccvfm6ry6mgnv5hvz324s";
+    sha256 = "1m7kmcjhnj3lx7xqs0nh262c4nxs5n8p7k9g6kc4gv1k5nrcrnpf";
   };
   cargo2nixSrc = fetchGit {
     # As of 2022-10-18: two small patches over unstable branch, one for clippy and one to fix feature detection
     url = "https://github.com/Alexis211/cargo2nix";
     ref = "custom_unstable";
-    rev = "a7a61179b66054904ef6a195d8da736eaaa06c36";
+    rev = "505caa32110d42ee03bd68b47031142eff9c827b";
   };
 
   /*

nix/compile.nix

@@ -42,7 +42,7 @@ let
   */
   toolchainOptions =
     if target == null || target == "x86_64-unknown-linux-musl" || target == "aarch64-unknown-linux-musl" then {
-      rustVersion = "1.63.0";
+      rustVersion = "1.65.0";
       extraRustComponents = [ "clippy" ];
     } else {
       rustToolchain = pkgs.symlinkJoin {

nix/kaniko.nix

@@ -7,7 +7,7 @@ pkgs.buildGoModule rec {
     owner = "GoogleContainerTools";
     repo = "kaniko";
     rev = "v${version}";
-    sha256 = "1fnclr556avxay6pvgw5ya3xbxfnf2gv4njq2hr4fd6fcjyslq5h";
+    sha256 = "TXgzO/NfLXVo5a7yyO3XYSk+9H1CwMF+vwbRx3kchQ8=";
   };
 
   vendorSha256 = null;

src/api/k2v/item.rs

@@ -211,7 +211,7 @@ pub async fn handle_poll_item(
 	let item = garage
 		.k2v
 		.rpc
-		.poll(
+		.poll_item(
 			bucket_id,
 			partition_key,
 			sort_key,

src/api/router_macros.rs

@@ -145,6 +145,7 @@ macro_rules! generateQueryParameters {
     ) => {
         #[derive(Debug)]
         #[allow(non_camel_case_types)]
+        #[allow(clippy::upper_case_acronyms)]
         enum Keyword {
             EMPTY,
             $( $kw_name, )*

src/api/s3/put.rs

@@ -119,6 +119,17 @@ pub(crate) async fn save_stream<S: Stream<Item = Result<Bytes, Error>> + Unpin>(
 		return Ok((version_uuid, data_md5sum_hex));
 	}
 
+	// The following consists in many steps that can each fail.
+	// Keep track that some cleanup will be needed if things fail
+	// before everything is finished (cleanup is done using the Drop trait).
+	let mut interrupted_cleanup = InterruptedCleanup(Some((
+		garage.clone(),
+		bucket.id,
+		key.into(),
+		version_uuid,
+		version_timestamp,
+	)));
+
 	// Write version identifier in object table so that we have a trace
 	// that we are uploading something
 	let mut object_version = ObjectVersion {
@@ -139,44 +150,27 @@ pub(crate) async fn save_stream<S: Stream<Item = Result<Bytes, Error>> + Unpin>(
 	// Transfer data and verify checksum
 	let first_block_hash = async_blake2sum(first_block.clone()).await;
 
-	let tx_result = (|| async {
+	let (total_size, data_md5sum, data_sha256sum) = read_and_put_blocks(
-		let (total_size, data_md5sum, data_sha256sum) = read_and_put_blocks(
+		&garage,
-			&garage,
+		&version,
-			&version,
+		1,
-			1,
+		first_block,
-			first_block,
+		first_block_hash,
-			first_block_hash,
+		&mut chunker,
-			&mut chunker,
+	)
-		)
+	.await?;
-		.await?;
 
-		ensure_checksum_matches(
+	ensure_checksum_matches(
-			data_md5sum.as_slice(),
+		data_md5sum.as_slice(),
-			data_sha256sum,
+		data_sha256sum,
-			content_md5.as_deref(),
+		content_md5.as_deref(),
-			content_sha256,
+		content_sha256,
-		)?;
+	)?;
 
-		check_quotas(&garage, bucket, key, total_size).await?;
+	check_quotas(&garage, bucket, key, total_size).await?;
-
-		Ok((total_size, data_md5sum))
-	})()
-	.await;
-
-	// If something went wrong, clean up
-	let (total_size, md5sum_arr) = match tx_result {
-		Ok(rv) => rv,
-		Err(e) => {
-			// Mark object as aborted, this will free the blocks further down
-			object_version.state = ObjectVersionState::Aborted;
-			let object = Object::new(bucket.id, key.into(), vec![object_version.clone()]);
-			garage.object_table.insert(&object).await?;
-			return Err(e);
-		}
-	};
 
 	// Save final object state, marked as Complete
-	let md5sum_hex = hex::encode(md5sum_arr);
+	let md5sum_hex = hex::encode(data_md5sum);
 	object_version.state = ObjectVersionState::Complete(ObjectVersionData::FirstBlock(
 		ObjectVersionMeta {
 			headers,
@@ -188,6 +182,10 @@ pub(crate) async fn save_stream<S: Stream<Item = Result<Bytes, Error>> + Unpin>(
 	let object = Object::new(bucket.id, key.into(), vec![object_version]);
 	garage.object_table.insert(&object).await?;
 
+	// We were not interrupted, everything went fine.
+	// We won't have to clean up on drop.
+	interrupted_cleanup.cancel();
+
 	Ok((version_uuid, md5sum_hex))
 }
 
@@ -426,6 +424,33 @@ pub fn put_response(version_uuid: Uuid, md5sum_hex: String) -> Response<Body> {
 		.unwrap()
 }
 
+struct InterruptedCleanup(Option<(Arc<Garage>, Uuid, String, Uuid, u64)>);
+
+impl InterruptedCleanup {
+	fn cancel(&mut self) {
+		drop(self.0.take());
+	}
+}
+impl Drop for InterruptedCleanup {
+	fn drop(&mut self) {
+		if let Some((garage, bucket_id, key, version_uuid, version_ts)) = self.0.take() {
+			tokio::spawn(async move {
+				let object_version = ObjectVersion {
+					uuid: version_uuid,
+					timestamp: version_ts,
+					state: ObjectVersionState::Aborted,
+				};
+				let object = Object::new(bucket_id, key, vec![object_version]);
+				if let Err(e) = garage.object_table.insert(&object).await {
+					warn!("Cannot cleanup after aborted PutObject: {}", e);
+				}
+			});
+		}
+	}
+}
+
+// ----
+
 pub async fn handle_create_multipart_upload(
 	garage: Arc<Garage>,
 	req: &Request<Body>,

src/block/Cargo.toml

@@ -31,7 +31,6 @@ rand = "0.8"
 async-compression = { version = "0.3", features = ["tokio", "zstd"] }
 zstd = { version = "0.9", default-features = false }
 
-rmp-serde = "0.15"
 serde = { version = "1.0", default-features = false, features = ["derive", "rc"] }
 serde_bytes = "0.11"
 

src/block/manager.rs

@@ -3,8 +3,10 @@ use std::pin::Pin;
 use std::sync::Arc;
 use std::time::Duration;
 
+use arc_swap::ArcSwapOption;
 use async_trait::async_trait;
 use bytes::Bytes;
+use rand::prelude::*;
 use serde::{Deserialize, Serialize};
 
 use futures::Stream;
@@ -22,9 +24,12 @@ use garage_rpc::rpc_helper::netapp::stream::{stream_asyncread, ByteStream};
 
 use garage_db as db;
 
+use garage_util::background::{vars, BackgroundRunner};
 use garage_util::data::*;
 use garage_util::error::*;
 use garage_util::metrics::RecordDuration;
+use garage_util::persister::PersisterShared;
+use garage_util::time::msec_to_rfc3339;
 
 use garage_rpc::rpc_helper::OrderTag;
 use garage_rpc::system::System;
@@ -87,7 +92,8 @@ pub struct BlockManager {
 
 	pub(crate) metrics: BlockManagerMetrics,
 
-	tx_scrub_command: mpsc::Sender<ScrubWorkerCommand>,
+	pub scrub_persister: PersisterShared<ScrubWorkerPersisted>,
+	tx_scrub_command: ArcSwapOption<mpsc::Sender<ScrubWorkerCommand>>,
 }
 
 #[derive(Serialize, Deserialize, Clone, Debug)]
@@ -126,7 +132,7 @@ impl BlockManager {
 		let metrics =
 			BlockManagerMetrics::new(rc.rc.clone(), resync.queue.clone(), resync.errors.clone());
 
-		let (scrub_tx, scrub_rx) = mpsc::channel(1);
+		let scrub_persister = PersisterShared::new(&system.metadata_dir, "scrub_info");
 
 		let block_manager = Arc::new(Self {
 			replication,
@@ -138,21 +144,46 @@ impl BlockManager {
 			system,
 			endpoint,
 			metrics,
-			tx_scrub_command: scrub_tx,
+			scrub_persister,
+			tx_scrub_command: ArcSwapOption::new(None),
 		});
 		block_manager.endpoint.set_handler(block_manager.clone());
 
+		block_manager
+	}
+
+	pub fn spawn_workers(self: &Arc<Self>, bg: &BackgroundRunner) {
 		// Spawn a bunch of resync workers
 		for index in 0..MAX_RESYNC_WORKERS {
-			let worker = ResyncWorker::new(index, block_manager.clone());
+			let worker = ResyncWorker::new(index, self.clone());
-			block_manager.system.background.spawn_worker(worker);
+			bg.spawn_worker(worker);
 		}
 
 		// Spawn scrub worker
-		let scrub_worker = ScrubWorker::new(block_manager.clone(), scrub_rx);
+		let (scrub_tx, scrub_rx) = mpsc::channel(1);
-		block_manager.system.background.spawn_worker(scrub_worker);
+		self.tx_scrub_command.store(Some(Arc::new(scrub_tx)));
+		bg.spawn_worker(ScrubWorker::new(
+			self.clone(),
+			scrub_rx,
+			self.scrub_persister.clone(),
+		));
+	}
 
-		block_manager
+	pub fn register_bg_vars(&self, vars: &mut vars::BgVars) {
+		self.resync.register_bg_vars(vars);
+
+		vars.register_rw(
+			&self.scrub_persister,
+			"scrub-tranquility",
+			|p| p.get_with(|x| x.tranquility),
+			|p, tranquility| p.set_with(|x| x.tranquility = tranquility),
+		);
+		vars.register_ro(&self.scrub_persister, "scrub-last-completed", |p| {
+			p.get_with(|x| msec_to_rfc3339(x.time_last_complete_scrub))
+		});
+		vars.register_ro(&self.scrub_persister, "scrub-corruptions_detected", |p| {
+			p.get_with(|x| x.corruptions_detected)
+		});
 	}
 
 	/// Ask nodes that might have a (possibly compressed) block for it
@@ -325,8 +356,11 @@ impl BlockManager {
 	}
 
 	/// Send command to start/stop/manager scrub worker
-	pub async fn send_scrub_command(&self, cmd: ScrubWorkerCommand) {
+	pub async fn send_scrub_command(&self, cmd: ScrubWorkerCommand) -> Result<(), Error> {
-		let _ = self.tx_scrub_command.send(cmd).await;
+		let tx = self.tx_scrub_command.load();
+		let tx = tx.as_ref().ok_or_message("scrub worker is not running")?;
+		tx.send(cmd).await.ok_or_message("send error")?;
+		Ok(())
 	}
 
 	/// Get the reference count of a block
@@ -643,14 +677,21 @@ impl BlockManagerLocked {
 			}
 		};
 
-		let mut path2 = path.clone();
+		let mut path_tmp = path.clone();
-		path2.set_extension("tmp");
+		let tmp_extension = format!("tmp{}", hex::encode(thread_rng().gen::<[u8; 4]>()));
-		let mut f = fs::File::create(&path2).await?;
+		path_tmp.set_extension(tmp_extension);
+
+		let mut delete_on_drop = DeleteOnDrop(Some(path_tmp.clone()));
+
+		let mut f = fs::File::create(&path_tmp).await?;
 		f.write_all(data).await?;
 		f.sync_all().await?;
 		drop(f);
 
-		fs::rename(path2, path).await?;
+		fs::rename(path_tmp, path).await?;
+
+		delete_on_drop.cancel();
+
 		if let Some(to_delete) = to_delete {
 			fs::remove_file(to_delete).await?;
 		}
@@ -716,3 +757,23 @@ async fn read_stream_to_end(mut stream: ByteStream) -> Result<Bytes, Error> {
 		.concat()
 		.into())
 }
+
+struct DeleteOnDrop(Option<PathBuf>);
+
+impl DeleteOnDrop {
+	fn cancel(&mut self) {
+		drop(self.0.take());
+	}
+}
+
+impl Drop for DeleteOnDrop {
+	fn drop(&mut self) {
+		if let Some(path) = self.0.take() {
+			tokio::spawn(async move {
+				if let Err(e) = fs::remove_file(&path).await {
+					debug!("DeleteOnDrop failed for {}: {}", path.display(), e);
+				}
+			});
+		}
+	}
+}

src/block/repair.rs

@@ -13,7 +13,7 @@ use tokio::sync::watch;
 use garage_util::background::*;
 use garage_util::data::*;
 use garage_util::error::*;
-use garage_util::persister::Persister;
+use garage_util::persister::PersisterShared;
 use garage_util::time::*;
 use garage_util::tranquilizer::Tranquilizer;
 
@@ -148,7 +148,7 @@ impl Worker for RepairWorker {
 		}
 	}
 
-	async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
 		unreachable!()
 	}
 }
@@ -168,15 +168,24 @@ pub struct ScrubWorker {
 	work: ScrubWorkerState,
 	tranquilizer: Tranquilizer,
 
-	persister: Persister<ScrubWorkerPersisted>,
+	persister: PersisterShared<ScrubWorkerPersisted>,
-	persisted: ScrubWorkerPersisted,
 }
 
 #[derive(Serialize, Deserialize)]
-struct ScrubWorkerPersisted {
+pub struct ScrubWorkerPersisted {
-	tranquility: u32,
+	pub tranquility: u32,
-	time_last_complete_scrub: u64,
+	pub(crate) time_last_complete_scrub: u64,
-	corruptions_detected: u64,
+	pub(crate) corruptions_detected: u64,
+}
+impl garage_util::migrate::InitialFormat for ScrubWorkerPersisted {}
+impl Default for ScrubWorkerPersisted {
+	fn default() -> Self {
+		ScrubWorkerPersisted {
+			time_last_complete_scrub: 0,
+			tranquility: INITIAL_SCRUB_TRANQUILITY,
+			corruptions_detected: 0,
+		}
+	}
 }
 
 enum ScrubWorkerState {
@@ -197,27 +206,20 @@ pub enum ScrubWorkerCommand {
 	Pause(Duration),
 	Resume,
 	Cancel,
-	SetTranquility(u32),
 }
 
 impl ScrubWorker {
-	pub fn new(manager: Arc<BlockManager>, rx_cmd: mpsc::Receiver<ScrubWorkerCommand>) -> Self {
+	pub(crate) fn new(
-		let persister = Persister::new(&manager.system.metadata_dir, "scrub_info");
+		manager: Arc<BlockManager>,
-		let persisted = match persister.load() {
+		rx_cmd: mpsc::Receiver<ScrubWorkerCommand>,
-			Ok(v) => v,
+		persister: PersisterShared<ScrubWorkerPersisted>,
-			Err(_) => ScrubWorkerPersisted {
+	) -> Self {
-				time_last_complete_scrub: 0,
-				tranquility: INITIAL_SCRUB_TRANQUILITY,
-				corruptions_detected: 0,
-			},
-		};
 		Self {
 			manager,
 			rx_cmd,
 			work: ScrubWorkerState::Finished,
 			tranquilizer: Tranquilizer::new(30),
 			persister,
-			persisted,
 		}
 	}
 
@@ -266,12 +268,6 @@ impl ScrubWorker {
 					}
 				}
 			}
-			ScrubWorkerCommand::SetTranquility(t) => {
-				self.persisted.tranquility = t;
-				if let Err(e) = self.persister.save_async(&self.persisted).await {
-					error!("Could not save new tranquilitiy value: {}", e);
-				}
-			}
 		}
 	}
 }
@@ -283,9 +279,18 @@ impl Worker for ScrubWorker {
 	}
 
 	fn status(&self) -> WorkerStatus {
+		let (corruptions_detected, tranquility, time_last_complete_scrub) =
+			self.persister.get_with(|p| {
+				(
+					p.corruptions_detected,
+					p.tranquility,
+					p.time_last_complete_scrub,
+				)
+			});
+
 		let mut s = WorkerStatus {
-			persistent_errors: Some(self.persisted.corruptions_detected),
+			persistent_errors: Some(corruptions_detected),
-			tranquility: Some(self.persisted.tranquility),
+			tranquility: Some(tranquility),
 			..Default::default()
 		};
 		match &self.work {
@@ -299,7 +304,7 @@ impl Worker for ScrubWorker {
 			ScrubWorkerState::Finished => {
 				s.freeform = vec![format!(
 					"Last scrub completed at {}",
-					msec_to_rfc3339(self.persisted.time_last_complete_scrub)
+					msec_to_rfc3339(time_last_complete_scrub)
 				)];
 			}
 		}
@@ -320,18 +325,17 @@ impl Worker for ScrubWorker {
 					match self.manager.read_block(&hash).await {
 						Err(Error::CorruptData(_)) => {
 							error!("Found corrupt data block during scrub: {:?}", hash);
-							self.persisted.corruptions_detected += 1;
+							self.persister.set_with(|p| p.corruptions_detected += 1)?;
-							self.persister.save_async(&self.persisted).await?;
 						}
 						Err(e) => return Err(e),
 						_ => (),
 					};
 					Ok(self
 						.tranquilizer
-						.tranquilize_worker(self.persisted.tranquility))
+						.tranquilize_worker(self.persister.get_with(|p| p.tranquility)))
 				} else {
-					self.persisted.time_last_complete_scrub = now_msec();
+					self.persister
-					self.persister.save_async(&self.persisted).await?;
+						.set_with(|p| p.time_last_complete_scrub = now_msec())?;
 					self.work = ScrubWorkerState::Finished;
 					self.tranquilizer.clear();
 					Ok(WorkerState::Idle)
@@ -341,12 +345,13 @@ impl Worker for ScrubWorker {
 		}
 	}
 
-	async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
 		let (wait_until, command) = match &self.work {
 			ScrubWorkerState::Running(_) => return WorkerState::Busy,
 			ScrubWorkerState::Paused(_, resume_time) => (*resume_time, ScrubWorkerCommand::Resume),
 			ScrubWorkerState::Finished => (
-				self.persisted.time_last_complete_scrub + SCRUB_INTERVAL.as_millis() as u64,
+				self.persister.get_with(|p| p.time_last_complete_scrub)
+					+ SCRUB_INTERVAL.as_millis() as u64,
 				ScrubWorkerCommand::Start,
 			),
 		};

src/block/resync.rs

@@ -3,7 +3,6 @@ use std::convert::TryInto;
 use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
-use arc_swap::ArcSwap;
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 
@@ -22,7 +21,7 @@ use garage_util::background::*;
 use garage_util::data::*;
 use garage_util::error::*;
 use garage_util::metrics::RecordDuration;
-use garage_util::persister::Persister;
+use garage_util::persister::PersisterShared;
 use garage_util::time::*;
 use garage_util::tranquilizer::Tranquilizer;
 
@@ -49,13 +48,12 @@ const INITIAL_RESYNC_TRANQUILITY: u32 = 2;
 
 pub struct BlockResyncManager {
 	pub(crate) queue: CountedTree,
-	pub(crate) notify: Notify,
+	pub(crate) notify: Arc<Notify>,
 	pub(crate) errors: CountedTree,
 
 	busy_set: BusySet,
 
-	persister: Persister<ResyncPersistedConfig>,
+	persister: PersisterShared<ResyncPersistedConfig>,
-	persisted: ArcSwap<ResyncPersistedConfig>,
 }
 
 #[derive(Serialize, Deserialize, Clone, Copy)]
@@ -63,6 +61,15 @@ struct ResyncPersistedConfig {
 	n_workers: usize,
 	tranquility: u32,
 }
+impl garage_util::migrate::InitialFormat for ResyncPersistedConfig {}
+impl Default for ResyncPersistedConfig {
+	fn default() -> Self {
+		ResyncPersistedConfig {
+			n_workers: 1,
+			tranquility: INITIAL_RESYNC_TRANQUILITY,
+		}
+	}
+}
 
 enum ResyncIterResult {
 	BusyDidSomething,
@@ -90,22 +97,14 @@ impl BlockResyncManager {
 			.expect("Unable to open block_local_resync_errors tree");
 		let errors = CountedTree::new(errors).expect("Could not count block_local_resync_errors");
 
-		let persister = Persister::new(&system.metadata_dir, "resync_cfg");
+		let persister = PersisterShared::new(&system.metadata_dir, "resync_cfg");
-		let persisted = match persister.load() {
-			Ok(v) => v,
-			Err(_) => ResyncPersistedConfig {
-				n_workers: 1,
-				tranquility: INITIAL_RESYNC_TRANQUILITY,
-			},
-		};
 
 		Self {
 			queue,
-			notify: Notify::new(),
+			notify: Arc::new(Notify::new()),
 			errors,
 			busy_set: Arc::new(Mutex::new(HashSet::new())),
 			persister,
-			persisted: ArcSwap::new(Arc::new(persisted)),
 		}
 	}
 
@@ -141,6 +140,38 @@ impl BlockResyncManager {
 		)))
 	}
 
+	pub fn register_bg_vars(&self, vars: &mut vars::BgVars) {
+		let notify = self.notify.clone();
+		vars.register_rw(
+			&self.persister,
+			"resync-worker-count",
+			|p| p.get_with(|x| x.n_workers),
+			move |p, n_workers| {
+				if !(1..=MAX_RESYNC_WORKERS).contains(&n_workers) {
+					return Err(Error::Message(format!(
+						"Invalid number of resync workers, must be between 1 and {}",
+						MAX_RESYNC_WORKERS
+					)));
+				}
+				p.set_with(|x| x.n_workers = n_workers)?;
+				notify.notify_waiters();
+				Ok(())
+			},
+		);
+
+		let notify = self.notify.clone();
+		vars.register_rw(
+			&self.persister,
+			"resync-tranquility",
+			|p| p.get_with(|x| x.tranquility),
+			move |p, tranquility| {
+				p.set_with(|x| x.tranquility = tranquility)?;
+				notify.notify_waiters();
+				Ok(())
+			},
+		);
+	}
+
 	// ---- Resync loop ----
 
 	// This part manages a queue of blocks that need to be
@@ -435,33 +466,6 @@ impl BlockResyncManager {
 
 		Ok(())
 	}
-
-	async fn update_persisted(
-		&self,
-		update: impl Fn(&mut ResyncPersistedConfig),
-	) -> Result<(), Error> {
-		let mut cfg: ResyncPersistedConfig = *self.persisted.load().as_ref();
-		update(&mut cfg);
-		self.persister.save_async(&cfg).await?;
-		self.persisted.store(Arc::new(cfg));
-		self.notify.notify_waiters();
-		Ok(())
-	}
-
-	pub async fn set_n_workers(&self, n_workers: usize) -> Result<(), Error> {
-		if !(1..=MAX_RESYNC_WORKERS).contains(&n_workers) {
-			return Err(Error::Message(format!(
-				"Invalid number of resync workers, must be between 1 and {}",
-				MAX_RESYNC_WORKERS
-			)));
-		}
-		self.update_persisted(|cfg| cfg.n_workers = n_workers).await
-	}
-
-	pub async fn set_tranquility(&self, tranquility: u32) -> Result<(), Error> {
-		self.update_persisted(|cfg| cfg.tranquility = tranquility)
-			.await
-	}
 }
 
 impl Drop for BusyBlock {
@@ -476,15 +480,18 @@ pub(crate) struct ResyncWorker {
 	manager: Arc<BlockManager>,
 	tranquilizer: Tranquilizer,
 	next_delay: Duration,
+	persister: PersisterShared<ResyncPersistedConfig>,
 }
 
 impl ResyncWorker {
 	pub(crate) fn new(index: usize, manager: Arc<BlockManager>) -> Self {
+		let persister = manager.resync.persister.clone();
 		Self {
 			index,
 			manager,
 			tranquilizer: Tranquilizer::new(30),
 			next_delay: Duration::from_secs(10),
+			persister,
 		}
 	}
 }
@@ -496,9 +503,9 @@ impl Worker for ResyncWorker {
 	}
 
 	fn status(&self) -> WorkerStatus {
-		let persisted = self.manager.resync.persisted.load();
+		let (n_workers, tranquility) = self.persister.get_with(|x| (x.n_workers, x.tranquility));
 
-		if self.index >= persisted.n_workers {
+		if self.index >= n_workers {
 			return WorkerStatus {
 				freeform: vec!["This worker is currently disabled".into()],
 				..Default::default()
@@ -507,22 +514,24 @@ impl Worker for ResyncWorker {
 
 		WorkerStatus {
 			queue_length: Some(self.manager.resync.queue_len().unwrap_or(0) as u64),
-			tranquility: Some(persisted.tranquility),
+			tranquility: Some(tranquility),
 			persistent_errors: Some(self.manager.resync.errors_len().unwrap_or(0) as u64),
 			..Default::default()
 		}
 	}
 
 	async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
-		if self.index >= self.manager.resync.persisted.load().n_workers {
+		let (n_workers, tranquility) = self.persister.get_with(|x| (x.n_workers, x.tranquility));
+
+		if self.index >= n_workers {
 			return Ok(WorkerState::Idle);
 		}
 
 		self.tranquilizer.reset();
 		match self.manager.resync.resync_iter(&self.manager).await {
-			Ok(ResyncIterResult::BusyDidSomething) => Ok(self
+			Ok(ResyncIterResult::BusyDidSomething) => {
-				.tranquilizer
+				Ok(self.tranquilizer.tranquilize_worker(tranquility))
-				.tranquilize_worker(self.manager.resync.persisted.load().tranquility)),
+			}
 			Ok(ResyncIterResult::BusyDidNothing) => Ok(WorkerState::Busy),
 			Ok(ResyncIterResult::IdleFor(delay)) => {
 				self.next_delay = delay;
@@ -540,8 +549,8 @@ impl Worker for ResyncWorker {
 		}
 	}
 
-	async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
-		while self.index >= self.manager.resync.persisted.load().n_workers {
+		while self.index >= self.persister.get_with(|x| x.n_workers) {
 			self.manager.resync.notify.notified().await
 		}
 

src/garage/Cargo.toml

@@ -42,7 +42,6 @@ rand = "0.8"
 async-trait = "0.1.7"
 sodiumoxide = { version = "0.2.5-0", package = "kuska-sodiumoxide" }
 
-rmp-serde = "0.15"
 serde = { version = "1.0", default-features = false, features = ["derive", "rc"] }
 serde_bytes = "0.11"
 structopt = { version = "0.3", default-features = false }
@@ -74,7 +73,7 @@ base64 = "0.13"
 
 
 [features]
-default = [ "bundled-libs", "metrics", "sled" ]
+default = [ "bundled-libs", "metrics", "sled", "k2v" ]
 
 k2v = [ "garage_util/k2v", "garage_api/k2v" ]
 

src/garage/admin.rs

@@ -5,6 +5,7 @@ use std::sync::Arc;
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 
+use garage_util::background::BackgroundRunner;
 use garage_util::crdt::*;
 use garage_util::data::*;
 use garage_util::error::Error as GarageError;
@@ -17,7 +18,6 @@ use garage_table::*;
 use garage_rpc::*;
 
 use garage_block::manager::BlockResyncErrorInfo;
-use garage_block::repair::ScrubWorkerCommand;
 
 use garage_model::bucket_alias_table::*;
 use garage_model::bucket_table::*;
@@ -59,6 +59,7 @@ pub enum AdminRpc {
 		HashMap<usize, garage_util::background::WorkerInfo>,
 		WorkerListOpt,
 	),
+	WorkerVars(Vec<(Uuid, String, String)>),
 	WorkerInfo(usize, garage_util::background::WorkerInfo),
 	BlockErrorList(Vec<BlockResyncErrorInfo>),
 	BlockInfo {
@@ -74,13 +75,18 @@ impl Rpc for AdminRpc {
 
 pub struct AdminRpcHandler {
 	garage: Arc<Garage>,
+	background: Arc<BackgroundRunner>,
 	endpoint: Arc<Endpoint<AdminRpc, Self>>,
 }
 
 impl AdminRpcHandler {
-	pub fn new(garage: Arc<Garage>) -> Arc<Self> {
+	pub fn new(garage: Arc<Garage>, background: Arc<BackgroundRunner>) -> Arc<Self> {
 		let endpoint = garage.system.netapp.endpoint(ADMIN_RPC_PATH.into());
-		let admin = Arc::new(Self { garage, endpoint });
+		let admin = Arc::new(Self {
+			garage,
+			background,
+			endpoint,
+		});
 		admin.endpoint.set_handler(admin.clone());
 		admin
 	}
@@ -759,7 +765,7 @@ impl AdminRpcHandler {
 				)))
 			}
 		} else {
-			launch_online_repair(self.garage.clone(), opt).await;
+			launch_online_repair(&self.garage, &self.background, opt).await?;
 			Ok(AdminRpc::Ok(format!(
 				"Repair launched on {:?}",
 				self.garage.system.id
@@ -925,12 +931,11 @@ impl AdminRpcHandler {
 	async fn handle_worker_cmd(&self, cmd: &WorkerOperation) -> Result<AdminRpc, Error> {
 		match cmd {
 			WorkerOperation::List { opt } => {
-				let workers = self.garage.background.get_worker_info();
+				let workers = self.background.get_worker_info();
 				Ok(AdminRpc::WorkerList(workers, *opt))
 			}
 			WorkerOperation::Info { tid } => {
 				let info = self
-					.garage
 					.background
 					.get_worker_info()
 					.get(tid)
@@ -938,32 +943,101 @@ impl AdminRpcHandler {
 					.clone();
 				Ok(AdminRpc::WorkerInfo(*tid, info))
 			}
-			WorkerOperation::Set { opt } => match opt {
+			WorkerOperation::Get {
-				WorkerSetCmd::ScrubTranquility { tranquility } => {
+				all_nodes,
-					let scrub_command = ScrubWorkerCommand::SetTranquility(*tranquility);
+				variable,
-					self.garage
+			} => self.handle_get_var(*all_nodes, variable).await,
-						.block_manager
+			WorkerOperation::Set {
-						.send_scrub_command(scrub_command)
+				all_nodes,
-						.await;
+				variable,
-					Ok(AdminRpc::Ok("Scrub tranquility updated".into()))
+				value,
+			} => self.handle_set_var(*all_nodes, variable, value).await,
+		}
+	}
+
+	async fn handle_get_var(
+		&self,
+		all_nodes: bool,
+		variable: &Option<String>,
+	) -> Result<AdminRpc, Error> {
+		if all_nodes {
+			let mut ret = vec![];
+			let ring = self.garage.system.ring.borrow().clone();
+			for node in ring.layout.node_ids().iter() {
+				let node = (*node).into();
+				match self
+					.endpoint
+					.call(
+						&node,
+						AdminRpc::Worker(WorkerOperation::Get {
+							all_nodes: false,
+							variable: variable.clone(),
+						}),
+						PRIO_NORMAL,
+					)
+					.await??
+				{
+					AdminRpc::WorkerVars(v) => ret.extend(v),
+					m => return Err(GarageError::unexpected_rpc_message(m).into()),
 				}
-				WorkerSetCmd::ResyncWorkerCount { worker_count } => {
+			}
-					self.garage
+			Ok(AdminRpc::WorkerVars(ret))
-						.block_manager
+		} else {
-						.resync
+			#[allow(clippy::collapsible_else_if)]
-						.set_n_workers(*worker_count)
+			if let Some(v) = variable {
-						.await?;
+				Ok(AdminRpc::WorkerVars(vec![(
-					Ok(AdminRpc::Ok("Number of resync workers updated".into()))
+					self.garage.system.id,
+					v.clone(),
+					self.garage.bg_vars.get(v)?,
+				)]))
+			} else {
+				let mut vars = self.garage.bg_vars.get_all();
+				vars.sort();
+				Ok(AdminRpc::WorkerVars(
+					vars.into_iter()
+						.map(|(k, v)| (self.garage.system.id, k.to_string(), v))
+						.collect(),
+				))
+			}
+		}
+	}
+
+	async fn handle_set_var(
+		&self,
+		all_nodes: bool,
+		variable: &str,
+		value: &str,
+	) -> Result<AdminRpc, Error> {
+		if all_nodes {
+			let mut ret = vec![];
+			let ring = self.garage.system.ring.borrow().clone();
+			for node in ring.layout.node_ids().iter() {
+				let node = (*node).into();
+				match self
+					.endpoint
+					.call(
+						&node,
+						AdminRpc::Worker(WorkerOperation::Set {
+							all_nodes: false,
+							variable: variable.to_string(),
+							value: value.to_string(),
+						}),
+						PRIO_NORMAL,
+					)
+					.await??
+				{
+					AdminRpc::WorkerVars(v) => ret.extend(v),
+					m => return Err(GarageError::unexpected_rpc_message(m).into()),
 				}
-				WorkerSetCmd::ResyncTranquility { tranquility } => {
+			}
-					self.garage
+			Ok(AdminRpc::WorkerVars(ret))
-						.block_manager
+		} else {
-						.resync
+			self.garage.bg_vars.set(variable, value)?;
-						.set_tranquility(*tranquility)
+			Ok(AdminRpc::WorkerVars(vec![(
-						.await?;
+				self.garage.system.id,
-					Ok(AdminRpc::Ok("Resync tranquility updated".into()))
+				variable.to_string(),
-				}
+				value.to_string(),
-			},
+			)]))
 		}
 	}
 

src/garage/cli/cmd.rs

@@ -191,6 +191,9 @@ pub async fn cmd_admin(
 		AdminRpc::WorkerList(wi, wlo) => {
 			print_worker_list(wi, wlo);
 		}
+		AdminRpc::WorkerVars(wv) => {
+			print_worker_vars(wv);
+		}
 		AdminRpc::WorkerInfo(tid, wi) => {
 			print_worker_info(tid, wi);
 		}

src/garage/cli/structs.rs

@@ -517,11 +517,25 @@ pub enum WorkerOperation {
 	/// Get detailed information about a worker
 	#[structopt(name = "info", version = garage_version())]
 	Info { tid: usize },
+	/// Get worker parameter
+	#[structopt(name = "get", version = garage_version())]
+	Get {
+		/// Gather variable values from all nodes
+		#[structopt(short = "a", long = "all-nodes")]
+		all_nodes: bool,
+		/// Variable name to get, or none to get all variables
+		variable: Option<String>,
+	},
 	/// Set worker parameter
 	#[structopt(name = "set", version = garage_version())]
 	Set {
-		#[structopt(subcommand)]
+		/// Set variable values on all nodes
-		opt: WorkerSetCmd,
+		#[structopt(short = "a", long = "all-nodes")]
+		all_nodes: bool,
+		/// Variable node to set
+		variable: String,
+		/// Value to set the variable to
+		value: String,
 	},
 }
 
@@ -535,19 +549,6 @@ pub struct WorkerListOpt {
 	pub errors: bool,
 }
 
-#[derive(Serialize, Deserialize, StructOpt, Debug, Eq, PartialEq, Clone)]
-pub enum WorkerSetCmd {
-	/// Set tranquility of scrub operations
-	#[structopt(name = "scrub-tranquility", version = garage_version())]
-	ScrubTranquility { tranquility: u32 },
-	/// Set number of concurrent block resync workers
-	#[structopt(name = "resync-worker-count", version = garage_version())]
-	ResyncWorkerCount { worker_count: usize },
-	/// Set tranquility of block resync operations
-	#[structopt(name = "resync-tranquility", version = garage_version())]
-	ResyncTranquility { tranquility: u32 },
-}
-
 #[derive(Serialize, Deserialize, StructOpt, Debug, Eq, PartialEq, Clone)]
 pub enum BlockOperation {
 	/// List all blocks that currently have a resync error

src/garage/cli/util.rs

@@ -357,6 +357,14 @@ pub fn print_worker_info(tid: usize, info: WorkerInfo) {
 	format_table(table);
 }
 
+pub fn print_worker_vars(wv: Vec<(Uuid, String, String)>) {
+	let table = wv
+		.into_iter()
+		.map(|(n, k, v)| format!("{:?}\t{}\t{}", n, k, v))
+		.collect::<Vec<_>>();
+	format_table(table);
+}
+
 pub fn print_block_error_list(el: Vec<BlockResyncErrorInfo>) {
 	let now = now_msec();
 	let tf = timeago::Formatter::new();

src/garage/repair/offline.rs

@@ -1,8 +1,5 @@
 use std::path::PathBuf;
 
-use tokio::sync::watch;
-
-use garage_util::background::*;
 use garage_util::config::*;
 use garage_util::error::*;
 
@@ -20,12 +17,8 @@ pub async fn offline_repair(config_file: PathBuf, opt: OfflineRepairOpt) -> Resu
 	info!("Loading configuration...");
 	let config = read_config(config_file)?;
 
-	info!("Initializing background runner...");
-	let (done_tx, done_rx) = watch::channel(false);
-	let (background, await_background_done) = BackgroundRunner::new(16, done_rx);
-
 	info!("Initializing Garage main data store...");
-	let garage = Garage::new(config.clone(), background)?;
+	let garage = Garage::new(config)?;
 
 	info!("Launching repair operation...");
 	match opt.what {
@@ -43,13 +36,7 @@ pub async fn offline_repair(config_file: PathBuf, opt: OfflineRepairOpt) -> Resu
 		}
 	}
 
-	info!("Repair operation finished, shutting down Garage internals...");
+	info!("Repair operation finished, shutting down...");
-	done_tx.send(true).unwrap();
-	drop(garage);
-
-	await_background_done.await?;
-
-	info!("Cleaning up...");
 
 	Ok(())
 }

src/garage/repair/online.rs

@@ -12,38 +12,37 @@ use garage_model::s3::version_table::*;
 use garage_table::*;
 use garage_util::background::*;
 use garage_util::error::Error;
+use garage_util::migrate::Migrate;
 
 use crate::*;
 
-pub async fn launch_online_repair(garage: Arc<Garage>, opt: RepairOpt) {
+pub async fn launch_online_repair(
+	garage: &Arc<Garage>,
+	bg: &BackgroundRunner,
+	opt: RepairOpt,
+) -> Result<(), Error> {
 	match opt.what {
 		RepairWhat::Tables => {
 			info!("Launching a full sync of tables");
-			garage.bucket_table.syncer.add_full_sync();
+			garage.bucket_table.syncer.add_full_sync()?;
-			garage.object_table.syncer.add_full_sync();
+			garage.object_table.syncer.add_full_sync()?;
-			garage.version_table.syncer.add_full_sync();
+			garage.version_table.syncer.add_full_sync()?;
-			garage.block_ref_table.syncer.add_full_sync();
+			garage.block_ref_table.syncer.add_full_sync()?;
-			garage.key_table.syncer.add_full_sync();
+			garage.key_table.syncer.add_full_sync()?;
 		}
 		RepairWhat::Versions => {
 			info!("Repairing the versions table");
-			garage
+			bg.spawn_worker(RepairVersionsWorker::new(garage.clone()));
-				.background
-				.spawn_worker(RepairVersionsWorker::new(garage.clone()));
 		}
 		RepairWhat::BlockRefs => {
 			info!("Repairing the block refs table");
-			garage
+			bg.spawn_worker(RepairBlockrefsWorker::new(garage.clone()));
-				.background
-				.spawn_worker(RepairBlockrefsWorker::new(garage.clone()));
 		}
 		RepairWhat::Blocks => {
 			info!("Repairing the stored blocks");
-			garage
+			bg.spawn_worker(garage_block::repair::RepairWorker::new(
-				.background
+				garage.block_manager.clone(),
-				.spawn_worker(garage_block::repair::RepairWorker::new(
+			));
-					garage.block_manager.clone(),
-				));
 		}
 		RepairWhat::Scrub { cmd } => {
 			let cmd = match cmd {
@@ -52,13 +51,18 @@ pub async fn launch_online_repair(garage: Arc<Garage>, opt: RepairOpt) {
 				ScrubCmd::Resume => ScrubWorkerCommand::Resume,
 				ScrubCmd::Cancel => ScrubWorkerCommand::Cancel,
 				ScrubCmd::SetTranquility { tranquility } => {
-					ScrubWorkerCommand::SetTranquility(tranquility)
+					garage
+						.block_manager
+						.scrub_persister
+						.set_with(|x| x.tranquility = tranquility)?;
+					return Ok(());
 				}
 			};
 			info!("Sending command to scrub worker: {:?}", cmd);
-			garage.block_manager.send_scrub_command(cmd).await;
+			garage.block_manager.send_scrub_command(cmd).await?;
 		}
 	}
+	Ok(())
 }
 
 // ----
@@ -93,20 +97,15 @@ impl Worker for RepairVersionsWorker {
 	}
 
 	async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
-		let item_bytes = match self.garage.version_table.data.store.get_gt(&self.pos)? {
+		let (item_bytes, next_pos) = match self.garage.version_table.data.store.get_gt(&self.pos)? {
-			Some((k, v)) => {
+			Some((k, v)) => (v, k),
-				self.pos = k;
-				v
-			}
 			None => {
 				info!("repair_versions: finished, done {}", self.counter);
 				return Ok(WorkerState::Done);
 			}
 		};
 
-		self.counter += 1;
+		let version = Version::decode(&item_bytes).ok_or_message("Cannot decode Version")?;
-
-		let version = rmp_serde::decode::from_read_ref::<_, Version>(&item_bytes)?;
 		if !version.deleted.get() {
 			let object = self
 				.garage
@@ -134,10 +133,13 @@ impl Worker for RepairVersionsWorker {
 			}
 		}
 
+		self.counter += 1;
+		self.pos = next_pos;
+
 		Ok(WorkerState::Busy)
 	}
 
-	async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
 		unreachable!()
 	}
 }
@@ -174,20 +176,16 @@ impl Worker for RepairBlockrefsWorker {
 	}
 
 	async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
-		let item_bytes = match self.garage.block_ref_table.data.store.get_gt(&self.pos)? {
+		let (item_bytes, next_pos) =
-			Some((k, v)) => {
+			match self.garage.block_ref_table.data.store.get_gt(&self.pos)? {
-				self.pos = k;
+				Some((k, v)) => (v, k),
-				v
+				None => {
-			}
+					info!("repair_block_ref: finished, done {}", self.counter);
-			None => {
+					return Ok(WorkerState::Done);
-				info!("repair_block_ref: finished, done {}", self.counter);
+				}
-				return Ok(WorkerState::Done);
+			};
-			}
-		};
 
-		self.counter += 1;
+		let block_ref = BlockRef::decode(&item_bytes).ok_or_message("Cannot decode BlockRef")?;
-
-		let block_ref = rmp_serde::decode::from_read_ref::<_, BlockRef>(&item_bytes)?;
 		if !block_ref.deleted.get() {
 			let version = self
 				.garage
@@ -212,10 +210,13 @@ impl Worker for RepairBlockrefsWorker {
 			}
 		}
 
+		self.counter += 1;
+		self.pos = next_pos;
+
 		Ok(WorkerState::Busy)
 	}
 
-	async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
 		unreachable!()
 	}
 }

src/garage/server.rs

@@ -35,12 +35,15 @@ pub async fn run_server(config_file: PathBuf) -> Result<(), Error> {
 	#[cfg(feature = "metrics")]
 	let metrics_exporter = opentelemetry_prometheus::exporter().init();
 
+	info!("Initializing Garage main data store...");
+	let garage = Garage::new(config.clone())?;
+
 	info!("Initializing background runner...");
 	let watch_cancel = watch_shutdown_signal();
-	let (background, await_background_done) = BackgroundRunner::new(16, watch_cancel.clone());
+	let (background, await_background_done) = BackgroundRunner::new(watch_cancel.clone());
 
-	info!("Initializing Garage main data store...");
+	info!("Spawning Garage workers...");
-	let garage = Garage::new(config.clone(), background)?;
+	garage.spawn_workers(&background);
 
 	if config.admin.trace_sink.is_some() {
 		info!("Initialize tracing...");
@@ -63,7 +66,7 @@ pub async fn run_server(config_file: PathBuf) -> Result<(), Error> {
 	let run_system = tokio::spawn(garage.system.clone().run(watch_cancel.clone()));
 
 	info!("Create admin RPC handler...");
-	AdminRpcHandler::new(garage.clone());
+	AdminRpcHandler::new(garage.clone(), background.clone());
 
 	// ---- Launch public-facing API servers ----
 

src/model/Cargo.toml

@@ -30,7 +30,6 @@ tracing = "0.1.30"
 rand = "0.8"
 zstd = { version = "0.9", default-features = false }
 
-rmp-serde = "0.15"
 serde = { version = "1.0", default-features = false, features = ["derive", "rc"] }
 serde_bytes = "0.11"
 

src/model/bucket_alias_table.rs

@@ -1,18 +1,26 @@
-use serde::{Deserialize, Serialize};
-
 use garage_util::data::*;
 
 use garage_table::crdt::*;
 use garage_table::*;
 
-/// The bucket alias table holds the names given to buckets
+mod v08 {
-/// in the global namespace.
+	use garage_util::crdt;
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	use garage_util::data::Uuid;
-pub struct BucketAlias {
+	use serde::{Deserialize, Serialize};
-	name: String,
+
-	pub state: crdt::Lww<Option<Uuid>>,
+	/// The bucket alias table holds the names given to buckets
+	/// in the global namespace.
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct BucketAlias {
+		pub(super) name: String,
+		pub state: crdt::Lww<Option<Uuid>>,
+	}
+
+	impl garage_util::migrate::InitialFormat for BucketAlias {}
 }
 
+pub use v08::*;
+
 impl BucketAlias {
 	pub fn new(name: String, ts: u64, bucket_id: Option<Uuid>) -> Option<Self> {
 		if !is_valid_bucket_name(&name) {

src/model/bucket_table.rs

@@ -1,5 +1,3 @@
-use serde::{Deserialize, Serialize};
-
 use garage_table::crdt::*;
 use garage_table::*;
 use garage_util::data::*;
@@ -7,71 +5,82 @@ use garage_util::time::*;
 
 use crate::permission::BucketKeyPerm;
 
-/// A bucket is a collection of objects
+mod v08 {
-///
+	use crate::permission::BucketKeyPerm;
-/// Its parameters are not directly accessible as:
+	use garage_util::crdt;
-///  - It must be possible to merge paramaters, hence the use of a LWW CRDT.
+	use garage_util::data::Uuid;
-///  - A bucket has 2 states, Present or Deleted and parameters make sense only if present.
+	use serde::{Deserialize, Serialize};
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+
-pub struct Bucket {
+	/// A bucket is a collection of objects
-	/// ID of the bucket
+	///
-	pub id: Uuid,
+	/// Its parameters are not directly accessible as:
-	/// State, and configuration if not deleted, of the bucket
+	///  - It must be possible to merge paramaters, hence the use of a LWW CRDT.
-	pub state: crdt::Deletable<BucketParams>,
+	///  - A bucket has 2 states, Present or Deleted and parameters make sense only if present.
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct Bucket {
+		/// ID of the bucket
+		pub id: Uuid,
+		/// State, and configuration if not deleted, of the bucket
+		pub state: crdt::Deletable<BucketParams>,
+	}
+
+	/// Configuration for a bucket
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct BucketParams {
+		/// Bucket's creation date
+		pub creation_date: u64,
+		/// Map of key with access to the bucket, and what kind of access they give
+		pub authorized_keys: crdt::Map<String, BucketKeyPerm>,
+
+		/// Map of aliases that are or have been given to this bucket
+		/// in the global namespace
+		/// (not authoritative: this is just used as an indication to
+		/// map back to aliases when doing ListBuckets)
+		pub aliases: crdt::LwwMap<String, bool>,
+		/// Map of aliases that are or have been given to this bucket
+		/// in namespaces local to keys
+		/// key = (access key id, alias name)
+		pub local_aliases: crdt::LwwMap<(String, String), bool>,
+
+		/// Whether this bucket is allowed for website access
+		/// (under all of its global alias names),
+		/// and if so, the website configuration XML document
+		pub website_config: crdt::Lww<Option<WebsiteConfig>>,
+		/// CORS rules
+		pub cors_config: crdt::Lww<Option<Vec<CorsRule>>>,
+		/// Bucket quotas
+		#[serde(default)]
+		pub quotas: crdt::Lww<BucketQuotas>,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct WebsiteConfig {
+		pub index_document: String,
+		pub error_document: Option<String>,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct CorsRule {
+		pub id: Option<String>,
+		pub max_age_seconds: Option<u64>,
+		pub allow_origins: Vec<String>,
+		pub allow_methods: Vec<String>,
+		pub allow_headers: Vec<String>,
+		pub expose_headers: Vec<String>,
+	}
+
+	#[derive(Default, PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
+	pub struct BucketQuotas {
+		/// Maximum size in bytes (bucket size = sum of sizes of objects in the bucket)
+		pub max_size: Option<u64>,
+		/// Maximum number of non-deleted objects in the bucket
+		pub max_objects: Option<u64>,
+	}
+
+	impl garage_util::migrate::InitialFormat for Bucket {}
 }
 
-/// Configuration for a bucket
+pub use v08::*;
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct BucketParams {
-	/// Bucket's creation date
-	pub creation_date: u64,
-	/// Map of key with access to the bucket, and what kind of access they give
-	pub authorized_keys: crdt::Map<String, BucketKeyPerm>,
-
-	/// Map of aliases that are or have been given to this bucket
-	/// in the global namespace
-	/// (not authoritative: this is just used as an indication to
-	/// map back to aliases when doing ListBuckets)
-	pub aliases: crdt::LwwMap<String, bool>,
-	/// Map of aliases that are or have been given to this bucket
-	/// in namespaces local to keys
-	/// key = (access key id, alias name)
-	pub local_aliases: crdt::LwwMap<(String, String), bool>,
-
-	/// Whether this bucket is allowed for website access
-	/// (under all of its global alias names),
-	/// and if so, the website configuration XML document
-	pub website_config: crdt::Lww<Option<WebsiteConfig>>,
-	/// CORS rules
-	pub cors_config: crdt::Lww<Option<Vec<CorsRule>>>,
-	/// Bucket quotas
-	#[serde(default)]
-	pub quotas: crdt::Lww<BucketQuotas>,
-}
-
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct WebsiteConfig {
-	pub index_document: String,
-	pub error_document: Option<String>,
-}
-
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct CorsRule {
-	pub id: Option<String>,
-	pub max_age_seconds: Option<u64>,
-	pub allow_origins: Vec<String>,
-	pub allow_methods: Vec<String>,
-	pub allow_headers: Vec<String>,
-	pub expose_headers: Vec<String>,
-}
-
-#[derive(Default, PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub struct BucketQuotas {
-	/// Maximum size in bytes (bucket size = sum of sizes of objects in the bucket)
-	pub max_size: Option<u64>,
-	/// Maximum number of non-deleted objects in the bucket
-	pub max_objects: Option<u64>,
-}
 
 impl AutoCrdt for BucketQuotas {
 	const WARN_IF_DIFFERENT: bool = true;

src/model/garage.rs

@@ -27,20 +27,20 @@ use crate::index_counter::*;
 use crate::key_table::*;
 
 #[cfg(feature = "k2v")]
-use crate::k2v::{item_table::*, poll::*, rpc::*};
+use crate::k2v::{history_table::*, item_table::*, sub::*, rpc::*};
 
 /// An entire Garage full of data
 pub struct Garage {
 	/// The parsed configuration Garage is running
 	pub config: Config,
+	/// The set of background variables that can be viewed/modified at runtime
+	pub bg_vars: vars::BgVars,
 
 	/// The replication mode of this cluster
 	pub replication_mode: ReplicationMode,
 
 	/// The local database
 	pub db: db::Db,
-	/// A background job runner
-	pub background: Arc<BackgroundRunner>,
 	/// The membership manager
 	pub system: Arc<System>,
 	/// The block manager
@@ -70,6 +70,8 @@ pub struct Garage {
 pub struct GarageK2V {
 	/// Table containing K2V items
 	pub item_table: Arc<Table<K2VItemTable, TableShardedReplication>>,
+	/// Table containing K2V modification history
+	pub history_table: Arc<Table<K2VHistoryTable, TableShardedReplication>>,
 	/// Indexing table containing K2V item counters
 	pub counter_table: Arc<IndexCounter<K2VItem>>,
 	/// K2V RPC handler
@@ -78,7 +80,7 @@ pub struct GarageK2V {
 
 impl Garage {
 	/// Create and run garage
-	pub fn new(config: Config, background: Arc<BackgroundRunner>) -> Result<Arc<Self>, Error> {
+	pub fn new(config: Config) -> Result<Arc<Self>, Error> {
 		// Create meta dir and data dir if they don't exist already
 		std::fs::create_dir_all(&config.metadata_dir)
 			.ok_or_message("Unable to create Garage metadata directory")?;
@@ -167,7 +169,7 @@ impl Garage {
 			.expect("Invalid replication_mode in config file.");
 
 		info!("Initialize membership management system...");
-		let system = System::new(network_key, background.clone(), replication_mode, &config)?;
+		let system = System::new(network_key, replication_mode, &config)?;
 
 		let data_rep_param = TableShardedReplication {
 			system: system.clone(),
@@ -225,7 +227,6 @@ impl Garage {
 		info!("Initialize version_table...");
 		let version_table = Table::new(
 			VersionTable {
-				background: background.clone(),
 				block_ref_table: block_ref_table.clone(),
 			},
 			meta_rep_param.clone(),
@@ -240,7 +241,6 @@ impl Garage {
 		#[allow(clippy::redundant_clone)]
 		let object_table = Table::new(
 			ObjectTable {
-				background: background.clone(),
 				version_table: version_table.clone(),
 				object_counter_table: object_counter_table.clone(),
 			},
@@ -253,12 +253,16 @@ impl Garage {
 		#[cfg(feature = "k2v")]
 		let k2v = GarageK2V::new(system.clone(), &db, meta_rep_param);
 
+		// Initialize bg vars
+		let mut bg_vars = vars::BgVars::new();
+		block_manager.register_bg_vars(&mut bg_vars);
+
 		// -- done --
 		Ok(Arc::new(Self {
 			config,
+			bg_vars,
 			replication_mode,
 			db,
-			background,
 			system,
 			block_manager,
 			bucket_table,
@@ -273,6 +277,22 @@ impl Garage {
 		}))
 	}
 
+	pub fn spawn_workers(&self, bg: &BackgroundRunner) {
+		self.block_manager.spawn_workers(bg);
+
+		self.bucket_table.spawn_workers(bg);
+		self.bucket_alias_table.spawn_workers(bg);
+		self.key_table.spawn_workers(bg);
+
+		self.object_table.spawn_workers(bg);
+		self.object_counter_table.spawn_workers(bg);
+		self.version_table.spawn_workers(bg);
+		self.block_ref_table.spawn_workers(bg);
+
+		#[cfg(feature = "k2v")]
+		self.k2v.spawn_workers(bg);
+	}
+
 	pub fn bucket_helper(&self) -> helper::bucket::BucketHelper {
 		helper::bucket::BucketHelper(self)
 	}
@@ -287,24 +307,49 @@ impl GarageK2V {
 	fn new(system: Arc<System>, db: &db::Db, meta_rep_param: TableShardedReplication) -> Self {
 		info!("Initialize K2V counter table...");
 		let counter_table = IndexCounter::new(system.clone(), meta_rep_param.clone(), db);
+
 		info!("Initialize K2V subscription manager...");
 		let subscriptions = Arc::new(SubscriptionManager::new());
+
 		info!("Initialize K2V item table...");
 		let item_table = Table::new(
 			K2VItemTable {
 				counter_table: counter_table.clone(),
 				subscriptions: subscriptions.clone(),
 			},
+			meta_rep_param.clone(),
+			system.clone(),
+			db,
+		);
+		info!("Initialize K2V history table...");
+		let history_table = Table::new(
+			K2VHistoryTable {
+				subscriptions: subscriptions.clone(),
+			},
 			meta_rep_param,
 			system.clone(),
 			db,
 		);
-		let rpc = K2VRpcHandler::new(system, item_table.clone(), subscriptions);
+
+		info!("Initialize K2V RPC handler...");
+		let rpc = K2VRpcHandler::new(
+			system,
+			db,
+			item_table.clone(),
+			history_table.clone(),
+			subscriptions,
+		);
 
 		Self {
 			item_table,
+			history_table,
 			counter_table,
 			rpc,
 		}
 	}
+
+	pub fn spawn_workers(&self, bg: &BackgroundRunner) {
+		self.item_table.spawn_workers(bg);
+		self.counter_table.spawn_workers(bg);
+	}
 }

src/model/index_counter.rs

@@ -1,19 +1,18 @@
 use core::ops::Bound;
-use std::collections::{hash_map, BTreeMap, HashMap};
+use std::collections::{BTreeMap, HashMap};
 use std::marker::PhantomData;
 use std::sync::Arc;
 
-use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
-use tokio::sync::{mpsc, watch};
 
 use garage_db as db;
 
 use garage_rpc::ring::Ring;
 use garage_rpc::system::System;
-use garage_util::background::*;
+use garage_util::background::BackgroundRunner;
 use garage_util::data::*;
 use garage_util::error::*;
+use garage_util::migrate::Migrate;
 use garage_util::time::*;
 
 use garage_table::crdt::*;
@@ -31,14 +30,44 @@ pub trait CountedItem: Clone + PartialEq + Send + Sync + 'static {
 	fn counts(&self) -> Vec<(&'static str, i64)>;
 }
 
-/// A counter entry in the global table
+mod v08 {
-#[derive(Clone, PartialEq, Debug, Serialize, Deserialize)]
+	use super::CountedItem;
-pub struct CounterEntry<T: CountedItem> {
+	use garage_util::data::Uuid;
-	pub pk: T::CP,
+	use serde::{Deserialize, Serialize};
-	pub sk: T::CS,
+	use std::collections::BTreeMap;
-	pub values: BTreeMap<String, CounterValue>,
+
+	// ---- Global part (the table everyone queries) ----
+
+	/// A counter entry in the global table
+	#[derive(Clone, PartialEq, Debug, Serialize, Deserialize)]
+	pub struct CounterEntry<T: CountedItem> {
+		pub pk: T::CP,
+		pub sk: T::CS,
+		pub values: BTreeMap<String, CounterValue>,
+	}
+
+	/// A counter entry in the global table
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct CounterValue {
+		pub node_values: BTreeMap<Uuid, (u64, i64)>,
+	}
+
+	impl<T: CountedItem> garage_util::migrate::InitialFormat for CounterEntry<T> {}
+
+	// ---- Local part (the counter we maintain transactionnaly on each node) ----
+
+	#[derive(PartialEq, Clone, Debug, Serialize, Deserialize)]
+	pub(super) struct LocalCounterEntry<T: CountedItem> {
+		pub(super) pk: T::CP,
+		pub(super) sk: T::CS,
+		pub(super) values: BTreeMap<String, (u64, i64)>,
+	}
+
+	impl<T: CountedItem> garage_util::migrate::InitialFormat for LocalCounterEntry<T> {}
 }
 
+pub use v08::*;
+
 impl<T: CountedItem> Entry<T::CP, T::CS> for CounterEntry<T> {
 	fn partition_key(&self) -> &T::CP {
 		&self.pk
@@ -80,12 +109,6 @@ impl<T: CountedItem> CounterEntry<T> {
 	}
 }
 
-/// A counter entry in the global table
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct CounterValue {
-	pub node_values: BTreeMap<Uuid, (u64, i64)>,
-}
-
 impl<T: CountedItem> Crdt for CounterEntry<T> {
 	fn merge(&mut self, other: &Self) {
 		for (name, e2) in other.values.iter() {
@@ -142,7 +165,6 @@ impl<T: CountedItem> TableSchema for CounterTable<T> {
 pub struct IndexCounter<T: CountedItem> {
 	this_node: Uuid,
 	local_counter: db::Tree,
-	propagate_tx: mpsc::UnboundedSender<(T::CP, T::CS, LocalCounterEntry<T>)>,
 	pub table: Arc<Table<CounterTable<T>, TableShardedReplication>>,
 }
 
@@ -152,16 +174,11 @@ impl<T: CountedItem> IndexCounter<T> {
 		replication: TableShardedReplication,
 		db: &db::Db,
 	) -> Arc<Self> {
-		let background = system.background.clone();
+		Arc::new(Self {
-
-		let (propagate_tx, propagate_rx) = mpsc::unbounded_channel();
-
-		let this = Arc::new(Self {
 			this_node: system.id,
 			local_counter: db
 				.open_tree(format!("local_counter_v2:{}", T::COUNTER_TABLE_NAME))
 				.expect("Unable to open local counter tree"),
-			propagate_tx,
 			table: Table::new(
 				CounterTable {
 					_phantom_t: Default::default(),
@@ -170,16 +187,11 @@ impl<T: CountedItem> IndexCounter<T> {
 				system,
 				db,
 			),
-		});
+		})
+	}
 
-		background.spawn_worker(IndexPropagatorWorker {
+	pub fn spawn_workers(&self, bg: &BackgroundRunner) {
-			index_counter: this.clone(),
+		self.table.spawn_workers(bg);
-			propagate_rx,
-			buf: HashMap::new(),
-			errors: 0,
-		});
-
-		this
 	}
 
 	pub fn count(
@@ -208,11 +220,9 @@ impl<T: CountedItem> IndexCounter<T> {
 		let tree_key = self.table.data.tree_key(pk, sk);
 
 		let mut entry = match tx.get(&self.local_counter, &tree_key[..])? {
-			Some(old_bytes) => {
+			Some(old_bytes) => LocalCounterEntry::<T>::decode(&old_bytes)
-				rmp_serde::decode::from_read_ref::<_, LocalCounterEntry<T>>(&old_bytes)
+				.ok_or_message("Cannot decode local counter entry")
-					.map_err(Error::RmpDecode)
+				.map_err(db::TxError::Abort)?,
-					.map_err(db::TxError::Abort)?
-			}
 			None => LocalCounterEntry {
 				pk: pk.clone(),
 				sk: sk.clone(),
@@ -227,17 +237,14 @@ impl<T: CountedItem> IndexCounter<T> {
 			ent.1 += *inc;
 		}
 
-		let new_entry_bytes = rmp_to_vec_all_named(&entry)
+		let new_entry_bytes = entry
+			.encode()
 			.map_err(Error::RmpEncode)
 			.map_err(db::TxError::Abort)?;
 		tx.insert(&self.local_counter, &tree_key[..], new_entry_bytes)?;
 
-		if let Err(e) = self.propagate_tx.send((pk.clone(), sk.clone(), entry)) {
+		let dist_entry = entry.into_counter_entry(self.this_node);
-			error!(
+		self.table.queue_insert(tx, &dist_entry)?;
-				"Could not propagate updated counter values, failed to send to channel: {}",
-				e
-			);
-		}
 
 		Ok(())
 	}
@@ -250,23 +257,6 @@ impl<T: CountedItem> IndexCounter<T> {
 		TS: TableSchema<E = T>,
 		TR: TableReplication,
 	{
-		let save_counter_entry = |entry: CounterEntry<T>| -> Result<(), Error> {
-			let entry_k = self
-				.table
-				.data
-				.tree_key(entry.partition_key(), entry.sort_key());
-			self.table
-				.data
-				.update_entry_with(&entry_k, |ent| match ent {
-					Some(mut ent) => {
-						ent.merge(&entry);
-						ent
-					}
-					None => entry.clone(),
-				})?;
-			Ok(())
-		};
-
 		// 1. Set all old local counters to zero
 		let now = now_msec();
 		let mut next_start: Option<Vec<u8>> = None;
@@ -289,20 +279,22 @@ impl<T: CountedItem> IndexCounter<T> {
 
 			info!("zeroing old counters... ({})", hex::encode(&batch[0].0));
 			for (local_counter_k, local_counter) in batch {
-				let mut local_counter =
+				let mut local_counter = LocalCounterEntry::<T>::decode(&local_counter)
-					rmp_serde::decode::from_read_ref::<_, LocalCounterEntry<T>>(&local_counter)?;
+					.ok_or_message("Cannot decode local counter entry")?;
 
 				for (_, tv) in local_counter.values.iter_mut() {
 					tv.0 = std::cmp::max(tv.0 + 1, now);
 					tv.1 = 0;
 				}
 
-				let local_counter_bytes = rmp_to_vec_all_named(&local_counter)?;
+				let local_counter_bytes = local_counter.encode()?;
 				self.local_counter
 					.insert(&local_counter_k, &local_counter_bytes)?;
 
 				let counter_entry = local_counter.into_counter_entry(self.this_node);
-				save_counter_entry(counter_entry)?;
+				self.local_counter
+					.db()
+					.transaction(|mut tx| self.table.queue_insert(&mut tx, &counter_entry))?;
 
 				next_start = Some(local_counter_k);
 			}
@@ -343,9 +335,8 @@ impl<T: CountedItem> IndexCounter<T> {
 				let local_counter_key = self.table.data.tree_key(pk, sk);
 				let mut local_counter = match self.local_counter.get(&local_counter_key)? {
 					Some(old_bytes) => {
-						let ent = rmp_serde::decode::from_read_ref::<_, LocalCounterEntry<T>>(
+						let ent = LocalCounterEntry::<T>::decode(&old_bytes)
-							&old_bytes,
+							.ok_or_message("Cannot decode local counter entry")?;
-						)?;
 						assert!(ent.pk == *pk);
 						assert!(ent.sk == *sk);
 						ent
@@ -362,12 +353,14 @@ impl<T: CountedItem> IndexCounter<T> {
 					tv.1 += v;
 				}
 
-				let local_counter_bytes = rmp_to_vec_all_named(&local_counter)?;
+				let local_counter_bytes = local_counter.encode()?;
 				self.local_counter
 					.insert(&local_counter_key, local_counter_bytes)?;
 
 				let counter_entry = local_counter.into_counter_entry(self.this_node);
-				save_counter_entry(counter_entry)?;
+				self.local_counter
+					.db()
+					.transaction(|mut tx| self.table.queue_insert(&mut tx, &counter_entry))?;
 
 				next_start = Some(counted_entry_k);
 			}
@@ -378,103 +371,7 @@ impl<T: CountedItem> IndexCounter<T> {
 	}
 }
 
-struct IndexPropagatorWorker<T: CountedItem> {
+// ----
-	index_counter: Arc<IndexCounter<T>>,
-	propagate_rx: mpsc::UnboundedReceiver<(T::CP, T::CS, LocalCounterEntry<T>)>,
-
-	buf: HashMap<Vec<u8>, CounterEntry<T>>,
-	errors: usize,
-}
-
-impl<T: CountedItem> IndexPropagatorWorker<T> {
-	fn add_ent(&mut self, pk: T::CP, sk: T::CS, counters: LocalCounterEntry<T>) {
-		let tree_key = self.index_counter.table.data.tree_key(&pk, &sk);
-		let dist_entry = counters.into_counter_entry(self.index_counter.this_node);
-		match self.buf.entry(tree_key) {
-			hash_map::Entry::Vacant(e) => {
-				e.insert(dist_entry);
-			}
-			hash_map::Entry::Occupied(mut e) => {
-				e.get_mut().merge(&dist_entry);
-			}
-		}
-	}
-}
-
-#[async_trait]
-impl<T: CountedItem> Worker for IndexPropagatorWorker<T> {
-	fn name(&self) -> String {
-		format!("{} counter", T::COUNTER_TABLE_NAME)
-	}
-
-	fn status(&self) -> WorkerStatus {
-		WorkerStatus {
-			queue_length: Some(self.buf.len() as u64),
-			..Default::default()
-		}
-	}
-
-	async fn work(&mut self, must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
-		// This loop batches updates to counters to be sent all at once.
-		// They are sent once the propagate_rx channel has been emptied (or is closed).
-		let closed = loop {
-			match self.propagate_rx.try_recv() {
-				Ok((pk, sk, counters)) => {
-					self.add_ent(pk, sk, counters);
-				}
-				Err(mpsc::error::TryRecvError::Empty) => break false,
-				Err(mpsc::error::TryRecvError::Disconnected) => break true,
-			}
-		};
-
-		if !self.buf.is_empty() {
-			let entries_k = self.buf.keys().take(100).cloned().collect::<Vec<_>>();
-			let entries = entries_k.iter().map(|k| self.buf.get(k).unwrap());
-			if let Err(e) = self.index_counter.table.insert_many(entries).await {
-				self.errors += 1;
-				if self.errors >= 2 && *must_exit.borrow() {
-					error!("({}) Could not propagate {} counter values: {}, these counters will not be updated correctly.", T::COUNTER_TABLE_NAME, self.buf.len(), e);
-					return Ok(WorkerState::Done);
-				}
-				// Propagate error up to worker manager, it will log it, increment a counter,
-				// and sleep for a certain delay (with exponential backoff), waiting for
-				// things to go back to normal
-				return Err(e);
-			} else {
-				for k in entries_k {
-					self.buf.remove(&k);
-				}
-				self.errors = 0;
-			}
-
-			return Ok(WorkerState::Busy);
-		} else if closed {
-			return Ok(WorkerState::Done);
-		} else {
-			return Ok(WorkerState::Idle);
-		}
-	}
-
-	async fn wait_for_work(&mut self, _must_exit: &watch::Receiver<bool>) -> WorkerState {
-		match self.propagate_rx.recv().await {
-			Some((pk, sk, counters)) => {
-				self.add_ent(pk, sk, counters);
-				WorkerState::Busy
-			}
-			None => match self.buf.is_empty() {
-				false => WorkerState::Busy,
-				true => WorkerState::Done,
-			},
-		}
-	}
-}
-
-#[derive(PartialEq, Clone, Debug, Serialize, Deserialize)]
-struct LocalCounterEntry<T: CountedItem> {
-	pk: T::CP,
-	sk: T::CS,
-	values: BTreeMap<String, (u64, i64)>,
-}
 
 impl<T: CountedItem> LocalCounterEntry<T> {
 	fn into_counter_entry(self, this_node: Uuid) -> CounterEntry<T> {

src/model/k2v/history_table.rs

@@ -0,0 +1,107 @@
+use std::sync::Arc;
+
+use garage_db as db;
+
+use garage_table::crdt::*;
+use garage_table::*;
+
+use crate::k2v::sub::*;
+
+mod v08 {
+	use crate::k2v::causality::K2VNodeId;
+	pub use crate::k2v::item_table::v08::{DvvsValue, K2VItem, K2VItemPartition};
+	use garage_util::crdt;
+	use serde::{Deserialize, Serialize};
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct K2VHistoryEntry {
+		// Partition key: the partition key of ins_item
+
+		/// The inserted item
+		pub ins_item: K2VItem,
+
+		/// Sort key: the node ID and its local counter
+		pub node_counter: K2VHistorySortKey,
+
+		/// The value of the node's local counter before this entry was updated
+		pub prev_counter: u64,
+		/// The timesamp of the update (!= counter, counters are incremented
+		/// by one, timestamps are real clock timestamps)
+		pub timestamp: u64,
+
+		/// Mark this history entry for deletion
+		pub deleted: crdt::Bool,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct K2VHistorySortKey {
+		pub node: K2VNodeId,
+		pub counter: u64,
+	}
+
+	impl garage_util::migrate::InitialFormat for K2VHistoryEntry {
+		const VERSION_MARKER: &'static [u8] = b"Gk2vhe08";
+	}
+}
+
+pub use v08::*;
+
+impl Crdt for K2VHistoryEntry {
+	fn merge(&mut self, other: &Self) {
+		self.ins_item.merge(&other.ins_item);
+		self.deleted.merge(&other.deleted);
+	}
+}
+
+impl SortKey for K2VHistorySortKey {
+	type B<'a> = [u8; 16];
+
+	fn sort_key(&self) -> [u8; 16] {
+		let mut ret = [0u8; 16];
+		ret[0..8].copy_from_slice(&u64::to_be_bytes(self.node));
+		ret[8..16].copy_from_slice(&u64::to_be_bytes(self.counter));
+		ret
+	}
+}
+
+impl Entry<K2VItemPartition, K2VHistorySortKey> for K2VHistoryEntry {
+	fn partition_key(&self) -> &K2VItemPartition {
+		&self.ins_item.partition
+	}
+	fn sort_key(&self) -> &K2VHistorySortKey {
+		&self.node_counter
+	}
+	fn is_tombstone(&self) -> bool {
+		self.deleted.get()
+	}
+}
+
+pub struct K2VHistoryTable {
+	pub(crate) subscriptions: Arc<SubscriptionManager>,
+}
+
+impl TableSchema for K2VHistoryTable {
+	const TABLE_NAME: &'static str = "k2v_history";
+
+	type P = K2VItemPartition;
+	type S = K2VHistorySortKey;
+	type E = K2VHistoryEntry;
+	type Filter = DeletedFilter;
+
+	fn updated(
+		&self,
+		_tx: &mut db::Transaction,
+		_old: Option<&Self::E>,
+		new: Option<&Self::E>,
+	) -> db::TxOpResult<()> {
+		if let Some(new_ent) = new {
+			self.subscriptions.notify_range(new_ent);
+		}
+
+		Ok(())
+	}
+
+	fn matches_filter(entry: &Self::E, filter: &Self::Filter) -> bool {
+		filter.apply(entry.deleted.get())
+	}
+}

src/model/k2v/item_table.rs

@@ -1,7 +1,8 @@
-use serde::{Deserialize, Serialize};
 use std::collections::BTreeMap;
 use std::sync::Arc;
 
+use serde::{Deserialize, Serialize};
+
 use garage_db as db;
 use garage_util::data::*;
 
@@ -10,38 +11,49 @@ use garage_table::*;
 
 use crate::index_counter::*;
 use crate::k2v::causality::*;
-use crate::k2v::poll::*;
+use crate::k2v::sub::*;
 
 pub const ENTRIES: &str = "entries";
 pub const CONFLICTS: &str = "conflicts";
 pub const VALUES: &str = "values";
 pub const BYTES: &str = "bytes";
 
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+pub(super) mod v08 {
-pub struct K2VItem {
+	use crate::k2v::causality::K2VNodeId;
-	pub partition: K2VItemPartition,
+	use garage_util::data::Uuid;
-	pub sort_key: String,
+	use serde::{Deserialize, Serialize};
+	use std::collections::BTreeMap;
 
-	items: BTreeMap<K2VNodeId, DvvsEntry>,
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct K2VItem {
+		pub partition: K2VItemPartition,
+		pub sort_key: String,
+
+		pub(super) items: BTreeMap<K2VNodeId, DvvsEntry>,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize, Hash)]
+	pub struct K2VItemPartition {
+		pub bucket_id: Uuid,
+		pub partition_key: String,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct DvvsEntry {
+		pub(super) t_discard: u64,
+		pub(super) values: Vec<(u64, DvvsValue)>,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub enum DvvsValue {
+		Value(#[serde(with = "serde_bytes")] Vec<u8>),
+		Deleted,
+	}
+
+	impl garage_util::migrate::InitialFormat for K2VItem {}
 }
 
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize, Hash)]
+pub use v08::*;
-pub struct K2VItemPartition {
-	pub bucket_id: Uuid,
-	pub partition_key: String,
-}
-
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-struct DvvsEntry {
-	t_discard: u64,
-	values: Vec<(u64, DvvsValue)>,
-}
-
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub enum DvvsValue {
-	Value(#[serde(with = "serde_bytes")] Vec<u8>),
-	Deleted,
-}
 
 impl K2VItem {
 	/// Creates a new K2VItem when no previous entry existed in the db
@@ -61,7 +73,8 @@ impl K2VItem {
 		this_node: Uuid,
 		context: &Option<CausalContext>,
 		new_value: DvvsValue,
-	) {
+		node_counter: u64,
+	) -> u64 {
 		if let Some(context) = context {
 			for (node, t_discard) in context.vector_clock.iter() {
 				if let Some(e) = self.items.get_mut(node) {
@@ -86,7 +99,9 @@ impl K2VItem {
 			values: vec![],
 		});
 		let t_prev = e.max_time();
-		e.values.push((t_prev + 1, new_value));
+		let t_new = std::cmp::max(node_counter + 1, t_prev + 1);
+		e.values.push((t_new, new_value));
+		t_new
 	}
 
 	/// Extract the causality context of a K2V Item
@@ -225,7 +240,7 @@ impl TableSchema for K2VItemTable {
 
 		// 2. Notify
 		if let Some(new_ent) = new {
-			self.subscriptions.notify(new_ent);
+			self.subscriptions.notify_item(new_ent);
 		}
 
 		Ok(())

src/model/k2v/mod.rs

@@ -1,6 +1,8 @@
 pub mod causality;
 
+pub mod history_table;
 pub mod item_table;
 
-pub mod poll;
+pub(crate) mod sub;
+
 pub mod rpc;

src/model/k2v/poll.rs

@@ -1,50 +0,0 @@
-use std::collections::HashMap;
-use std::sync::Mutex;
-
-use serde::{Deserialize, Serialize};
-use tokio::sync::broadcast;
-
-use crate::k2v::item_table::*;
-
-#[derive(Debug, Hash, Clone, PartialEq, Eq, Serialize, Deserialize)]
-pub struct PollKey {
-	pub partition: K2VItemPartition,
-	pub sort_key: String,
-}
-
-#[derive(Default)]
-pub struct SubscriptionManager {
-	subscriptions: Mutex<HashMap<PollKey, broadcast::Sender<K2VItem>>>,
-}
-
-impl SubscriptionManager {
-	pub fn new() -> Self {
-		Self::default()
-	}
-
-	pub fn subscribe(&self, key: &PollKey) -> broadcast::Receiver<K2VItem> {
-		let mut subs = self.subscriptions.lock().unwrap();
-		if let Some(s) = subs.get(key) {
-			s.subscribe()
-		} else {
-			let (tx, rx) = broadcast::channel(8);
-			subs.insert(key.clone(), tx);
-			rx
-		}
-	}
-
-	pub fn notify(&self, item: &K2VItem) {
-		let key = PollKey {
-			partition: item.partition.clone(),
-			sort_key: item.sort_key.clone(),
-		};
-		let mut subs = self.subscriptions.lock().unwrap();
-		if let Some(s) = subs.get(&key) {
-			if s.send(item.clone()).is_err() {
-				// no more subscribers, remove channel from here
-				// (we will re-create it later if we need to subscribe again)
-				subs.remove(&key);
-			}
-		}
-	}
-}

src/model/k2v/rpc.rs

@@ -6,6 +6,7 @@
 //! mean the vector clock gets much larger than needed).
 
 use std::collections::HashMap;
+use std::convert::TryInto;
 use std::sync::Arc;
 use std::time::Duration;
 
@@ -15,9 +16,12 @@ use futures::StreamExt;
 use serde::{Deserialize, Serialize};
 use tokio::select;
 
+use garage_db as db;
+
 use garage_util::crdt::*;
 use garage_util::data::*;
 use garage_util::error::*;
+use garage_util::time::*;
 
 use garage_rpc::system::System;
 use garage_rpc::*;
@@ -26,8 +30,9 @@ use garage_table::replication::{TableReplication, TableShardedReplication};
 use garage_table::{PartitionKey, Table};
 
 use crate::k2v::causality::*;
+use crate::k2v::history_table::*;
 use crate::k2v::item_table::*;
-use crate::k2v::poll::*;
+use crate::k2v::sub::*;
 
 /// RPC messages for K2V
 #[derive(Debug, Serialize, Deserialize)]
@@ -59,6 +64,8 @@ impl Rpc for K2VRpc {
 pub struct K2VRpcHandler {
 	system: Arc<System>,
 	item_table: Arc<Table<K2VItemTable, TableShardedReplication>>,
+	history_table: Arc<Table<K2VHistoryTable, TableShardedReplication>>,
+	local_counter_tree: db::Tree,
 	endpoint: Arc<Endpoint<K2VRpc, Self>>,
 	subscriptions: Arc<SubscriptionManager>,
 }
@@ -66,14 +73,21 @@ pub struct K2VRpcHandler {
 impl K2VRpcHandler {
 	pub fn new(
 		system: Arc<System>,
+		db: &db::Db,
 		item_table: Arc<Table<K2VItemTable, TableShardedReplication>>,
+		history_table: Arc<Table<K2VHistoryTable, TableShardedReplication>>,
 		subscriptions: Arc<SubscriptionManager>,
 	) -> Arc<Self> {
+		let local_counter_tree = db
+			.open_tree("k2v_local_counter")
+			.expect("Unable to open DB tree for k2v local counter");
 		let endpoint = system.netapp.endpoint("garage_model/k2v/Rpc".to_string());
 
 		let rpc_handler = Arc::new(Self {
 			system,
 			item_table,
+			history_table,
+			local_counter_tree,
 			endpoint,
 			subscriptions,
 		});
@@ -181,7 +195,7 @@ impl K2VRpcHandler {
 		Ok(())
 	}
 
-	pub async fn poll(
+	pub async fn poll_item(
 		&self,
 		bucket_id: Uuid,
 		partition_key: String,
@@ -273,14 +287,18 @@ impl K2VRpcHandler {
 	}
 
 	fn local_insert(&self, item: &InsertedItem) -> Result<Option<K2VItem>, Error> {
-		let tree_key = self
+		let now = now_msec();
-			.item_table
-			.data
-			.tree_key(&item.partition, &item.sort_key);
 
 		self.item_table
 			.data
-			.update_entry_with(&tree_key[..], |ent| {
+			.update_entry_with(&item.partition, &item.sort_key, |tx, ent| {
+				let local_counter_key = item.partition.hash();
+				let old_local_counter = tx
+					.get(&self.local_counter_tree, &local_counter_key)?
+					.and_then(|x| x.try_into().ok())
+					.map(u64::from_be_bytes)
+					.unwrap_or_default();
+
 				let mut ent = ent.unwrap_or_else(|| {
 					K2VItem::new(
 						item.partition.bucket_id,
@@ -288,13 +306,37 @@ impl K2VRpcHandler {
 						item.sort_key.clone(),
 					)
 				});
-				ent.update(self.system.id, &item.causal_context, item.value.clone());
+				let new_local_counter = ent.update(
-				ent
+					self.system.id,
+					&item.causal_context,
+					item.value.clone(),
+					old_local_counter,
+				);
+
+				tx.insert(
+					&self.local_counter_tree,
+					&local_counter_key,
+					u64::to_be_bytes(new_local_counter),
+				)?;
+
+				let hist_entry = K2VHistoryEntry {
+					ins_item: ent.clone(),
+					node_counter: K2VHistorySortKey {
+						node: make_node_id(self.system.id),
+						counter: new_local_counter,
+					},
+					prev_counter: old_local_counter,
+					timestamp: now,
+					deleted: false.into(),
+				};
+				self.history_table.queue_insert(tx, &hist_entry)?;
+
+				Ok(ent)
 			})
 	}
 
-	async fn handle_poll(&self, key: &PollKey, ct: &CausalContext) -> Result<K2VItem, Error> {
+	async fn handle_poll_item(&self, key: &PollKey, ct: &CausalContext) -> Result<K2VItem, Error> {
-		let mut chan = self.subscriptions.subscribe(key);
+		let mut chan = self.subscriptions.subscribe_item(key);
 
 		let mut value = self
 			.item_table
@@ -331,7 +373,7 @@ impl EndpointHandler<K2VRpc> for K2VRpcHandler {
 			} => {
 				let delay = tokio::time::sleep(Duration::from_millis(*timeout_msec));
 				select! {
-					ret = self.handle_poll(key, causal_context) => ret.map(Some).map(K2VRpc::PollItemResponse),
+					ret = self.handle_poll_item(key, causal_context) => ret.map(Some).map(K2VRpc::PollItemResponse),
 					_ = delay => Ok(K2VRpc::PollItemResponse(None)),
 				}
 			}

src/model/k2v/sub.rs

@@ -0,0 +1,115 @@
+use std::collections::HashMap;
+use std::sync::Mutex;
+
+use serde::{Deserialize, Serialize};
+use tokio::sync::broadcast;
+
+use crate::k2v::history_table::*;
+use crate::k2v::item_table::*;
+
+#[derive(Debug, Hash, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct PollKey {
+	pub partition: K2VItemPartition,
+	pub sort_key: String,
+}
+
+#[derive(Debug, Hash, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct PollRange {
+	pub partition: K2VItemPartition,
+	pub prefix: Option<String>,
+	pub start: Option<String>,
+	pub end: Option<String>,
+}
+
+#[derive(Default)]
+pub struct SubscriptionManager {
+	item_subscriptions: Mutex<HashMap<PollKey, broadcast::Sender<K2VItem>>>,
+	range_subscriptions: Mutex<HashMap<PollRange, broadcast::Sender<K2VHistoryEntry>>>,
+}
+
+impl SubscriptionManager {
+	pub fn new() -> Self {
+		Self::default()
+	}
+
+	// ---- simple item polling ----
+
+	pub fn subscribe_item(&self, key: &PollKey) -> broadcast::Receiver<K2VItem> {
+		let mut subs = self.item_subscriptions.lock().unwrap();
+		if let Some(s) = subs.get(key) {
+			s.subscribe()
+		} else {
+			let (tx, rx) = broadcast::channel(8);
+			subs.insert(key.clone(), tx);
+			rx
+		}
+	}
+
+	pub fn notify_item(&self, item: &K2VItem) {
+		let key = PollKey {
+			partition: item.partition.clone(),
+			sort_key: item.sort_key.clone(),
+		};
+		let mut subs = self.item_subscriptions.lock().unwrap();
+		if let Some(s) = subs.get(&key) {
+			if s.send(item.clone()).is_err() {
+				// no more subscribers, remove channel from here
+				// (we will re-create it later if we need to subscribe again)
+				subs.remove(&key);
+			}
+		}
+	}
+
+	// ---- range polling ----
+
+	pub fn subscribe_range(&self, key: &PollRange) -> broadcast::Receiver<K2VHistoryEntry> {
+		let mut subs = self.range_subscriptions.lock().unwrap();
+		if let Some(s) = subs.get(key) {
+			s.subscribe()
+		} else {
+			let (tx, rx) = broadcast::channel(8);
+			subs.insert(key.clone(), tx);
+			rx
+		}
+	}
+
+	pub fn notify_range(&self, entry: &K2VHistoryEntry) {
+		let mut subs = self.range_subscriptions.lock().unwrap();
+		let mut dead_subs = vec![];
+
+		for (sub, chan) in subs.iter() {
+			if sub.matches(&entry) {
+				if chan.send(entry.clone()).is_err() {
+					dead_subs.push(sub.clone());
+				}
+			} else if chan.receiver_count() == 0 {
+				dead_subs.push(sub.clone());
+			}
+		}
+
+		for sub in dead_subs.iter() {
+			subs.remove(sub);
+		}
+	}
+}
+
+impl PollRange {
+	fn matches(&self, entry: &K2VHistoryEntry) -> bool {
+		entry.ins_item.partition == self.partition
+			&& self
+				.prefix
+				.as_ref()
+				.map(|x| entry.ins_item.sort_key.starts_with(x))
+				.unwrap_or(true)
+			&& self
+				.start
+				.as_ref()
+				.map(|x| entry.ins_item.sort_key >= *x)
+				.unwrap_or(true)
+			&& self
+				.end
+				.as_ref()
+				.map(|x| entry.ins_item.sort_key < *x)
+				.unwrap_or(true)
+	}
+}

src/model/key_table.rs

@@ -1,45 +1,121 @@
 use serde::{Deserialize, Serialize};
 
-use garage_table::crdt::*;
+use garage_util::crdt::{self, Crdt};
-use garage_table::*;
 use garage_util::data::*;
 
+use garage_table::{DeletedFilter, EmptyKey, Entry, TableSchema};
+
 use crate::permission::BucketKeyPerm;
 
-use crate::prev::v051::key_table as old;
+pub(crate) mod v05 {
+	use garage_util::crdt;
+	use serde::{Deserialize, Serialize};
 
-/// An api key
+	/// An api key
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct Key {
+	pub struct Key {
-	/// The id of the key (immutable), used as partition key
+		/// The id of the key (immutable), used as partition key
-	pub key_id: String,
+		pub key_id: String,
 
-	/// Internal state of the key
+		/// The secret_key associated
-	pub state: crdt::Deletable<KeyParams>,
+		pub secret_key: String,
+
+		/// Name for the key
+		pub name: crdt::Lww<String>,
+
+		/// Is the key deleted
+		pub deleted: crdt::Bool,
+
+		/// Buckets in which the key is authorized. Empty if `Key` is deleted
+		// CRDT interaction: deleted implies authorized_buckets is empty
+		pub authorized_buckets: crdt::LwwMap<String, PermissionSet>,
+	}
+
+	/// Permission given to a key in a bucket
+	#[derive(PartialOrd, Ord, PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct PermissionSet {
+		/// The key can be used to read the bucket
+		pub allow_read: bool,
+		/// The key can be used to write in the bucket
+		pub allow_write: bool,
+	}
+
+	impl crdt::AutoCrdt for PermissionSet {
+		const WARN_IF_DIFFERENT: bool = true;
+	}
+
+	impl garage_util::migrate::InitialFormat for Key {}
 }
 
-/// Configuration for a key
+mod v08 {
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	use super::v05;
-pub struct KeyParams {
+	use crate::permission::BucketKeyPerm;
-	/// The secret_key associated (immutable)
+	use garage_util::crdt;
-	pub secret_key: String,
+	use garage_util::data::Uuid;
+	use serde::{Deserialize, Serialize};
 
-	/// Name for the key
+	/// An api key
-	pub name: crdt::Lww<String>,
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct Key {
+		/// The id of the key (immutable), used as partition key
+		pub key_id: String,
 
-	/// Flag to allow users having this key to create buckets
+		/// Internal state of the key
-	pub allow_create_bucket: crdt::Lww<bool>,
+		pub state: crdt::Deletable<KeyParams>,
+	}
 
-	/// If the key is present: it gives some permissions,
+	/// Configuration for a key
-	/// a map of bucket IDs (uuids) to permissions.
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-	/// Otherwise no permissions are granted to key
+	pub struct KeyParams {
-	pub authorized_buckets: crdt::Map<Uuid, BucketKeyPerm>,
+		/// The secret_key associated (immutable)
+		pub secret_key: String,
 
-	/// A key can have a local view of buckets names it is
+		/// Name for the key
-	/// the only one to see, this is the namespace for these aliases
+		pub name: crdt::Lww<String>,
-	pub local_aliases: crdt::LwwMap<String, Option<Uuid>>,
+
+		/// Flag to allow users having this key to create buckets
+		pub allow_create_bucket: crdt::Lww<bool>,
+
+		/// If the key is present: it gives some permissions,
+		/// a map of bucket IDs (uuids) to permissions.
+		/// Otherwise no permissions are granted to key
+		pub authorized_buckets: crdt::Map<Uuid, BucketKeyPerm>,
+
+		/// A key can have a local view of buckets names it is
+		/// the only one to see, this is the namespace for these aliases
+		pub local_aliases: crdt::LwwMap<String, Option<Uuid>>,
+	}
+
+	impl garage_util::migrate::Migrate for Key {
+		type Previous = v05::Key;
+
+		fn migrate(old_k: v05::Key) -> Key {
+			let name = crdt::Lww::raw(old_k.name.timestamp(), old_k.name.get().clone());
+
+			let state = if old_k.deleted.get() {
+				crdt::Deletable::Deleted
+			} else {
+				// Authorized buckets is ignored here,
+				// migration is performed in specific migration code in
+				// garage/migrate.rs
+				crdt::Deletable::Present(KeyParams {
+					secret_key: old_k.secret_key,
+					name,
+					allow_create_bucket: crdt::Lww::new(false),
+					authorized_buckets: crdt::Map::new(),
+					local_aliases: crdt::LwwMap::new(),
+				})
+			};
+			Key {
+				key_id: old_k.key_id,
+				state,
+			}
+		}
+	}
 }
 
+pub use v08::*;
+
 impl KeyParams {
 	fn new(secret_key: &str, name: &str) -> Self {
 		KeyParams {
@@ -173,28 +249,4 @@ impl TableSchema for KeyTable {
 			}
 		}
 	}
-
-	fn try_migrate(bytes: &[u8]) -> Option<Self::E> {
-		let old_k = rmp_serde::decode::from_read_ref::<_, old::Key>(bytes).ok()?;
-		let name = crdt::Lww::raw(old_k.name.timestamp(), old_k.name.get().clone());
-
-		let state = if old_k.deleted.get() {
-			crdt::Deletable::Deleted
-		} else {
-			// Authorized buckets is ignored here,
-			// migration is performed in specific migration code in
-			// garage/migrate.rs
-			crdt::Deletable::Present(KeyParams {
-				secret_key: old_k.secret_key,
-				name,
-				allow_create_bucket: crdt::Lww::new(false),
-				authorized_buckets: crdt::Map::new(),
-				local_aliases: crdt::LwwMap::new(),
-			})
-		};
-		Some(Key {
-			key_id: old_k.key_id,
-			state,
-		})
-	}
 }

src/model/migrate.rs

@@ -2,6 +2,7 @@ use std::sync::Arc;
 
 use garage_util::crdt::*;
 use garage_util::data::*;
+use garage_util::encode::nonversioned_decode;
 use garage_util::error::Error as GarageError;
 use garage_util::time::*;
 
@@ -28,8 +29,8 @@ impl Migrate {
 		let mut old_buckets = vec![];
 		for res in tree.iter().map_err(GarageError::from)? {
 			let (_k, v) = res.map_err(GarageError::from)?;
-			let bucket = rmp_serde::decode::from_read_ref::<_, old_bucket::Bucket>(&v[..])
+			let bucket =
-				.map_err(GarageError::from)?;
+				nonversioned_decode::<old_bucket::Bucket>(&v[..]).map_err(GarageError::from)?;
 			old_buckets.push(bucket);
 		}
 

src/model/prev/v051/bucket_table.rs

@@ -3,7 +3,7 @@ use serde::{Deserialize, Serialize};
 use garage_table::crdt::Crdt;
 use garage_table::*;
 
-use super::key_table::PermissionSet;
+use crate::key_table::v05::PermissionSet;
 
 /// A bucket is a collection of objects
 ///

src/model/prev/v051/key_table.rs

@@ -1,50 +0,0 @@
-use serde::{Deserialize, Serialize};
-
-use garage_table::crdt::*;
-use garage_table::*;
-
-/// An api key
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct Key {
-	/// The id of the key (immutable), used as partition key
-	pub key_id: String,
-
-	/// The secret_key associated
-	pub secret_key: String,
-
-	/// Name for the key
-	pub name: crdt::Lww<String>,
-
-	/// Is the key deleted
-	pub deleted: crdt::Bool,
-
-	/// Buckets in which the key is authorized. Empty if `Key` is deleted
-	// CRDT interaction: deleted implies authorized_buckets is empty
-	pub authorized_buckets: crdt::LwwMap<String, PermissionSet>,
-}
-
-/// Permission given to a key in a bucket
-#[derive(PartialOrd, Ord, PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct PermissionSet {
-	/// The key can be used to read the bucket
-	pub allow_read: bool,
-	/// The key can be used to write in the bucket
-	pub allow_write: bool,
-}
-
-impl AutoCrdt for PermissionSet {
-	const WARN_IF_DIFFERENT: bool = true;
-}
-
-impl Crdt for Key {
-	fn merge(&mut self, other: &Self) {
-		self.name.merge(&other.name);
-		self.deleted.merge(&other.deleted);
-
-		if self.deleted.get() {
-			self.authorized_buckets.clear();
-		} else {
-			self.authorized_buckets.merge(&other.authorized_buckets);
-		}
-	}
-}

src/model/prev/v051/mod.rs

@@ -1,4 +1 @@
 pub(crate) mod bucket_table;
-pub(crate) mod key_table;
-pub(crate) mod object_table;
-pub(crate) mod version_table;

src/model/prev/v051/object_table.rs

@@ -1,149 +0,0 @@
-use serde::{Deserialize, Serialize};
-use std::collections::BTreeMap;
-
-use garage_util::data::*;
-
-use garage_table::crdt::*;
-
-/// An object
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct Object {
-	/// The bucket in which the object is stored, used as partition key
-	pub bucket: String,
-
-	/// The key at which the object is stored in its bucket, used as sorting key
-	pub key: String,
-
-	/// The list of currenty stored versions of the object
-	versions: Vec<ObjectVersion>,
-}
-
-impl Object {
-	/// Get a list of currently stored versions of `Object`
-	pub fn versions(&self) -> &[ObjectVersion] {
-		&self.versions[..]
-	}
-}
-
-/// Informations about a version of an object
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct ObjectVersion {
-	/// Id of the version
-	pub uuid: Uuid,
-	/// Timestamp of when the object was created
-	pub timestamp: u64,
-	/// State of the version
-	pub state: ObjectVersionState,
-}
-
-/// State of an object version
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub enum ObjectVersionState {
-	/// The version is being received
-	Uploading(ObjectVersionHeaders),
-	/// The version is fully received
-	Complete(ObjectVersionData),
-	/// The version uploaded containded errors or the upload was explicitly aborted
-	Aborted,
-}
-
-impl Crdt for ObjectVersionState {
-	fn merge(&mut self, other: &Self) {
-		use ObjectVersionState::*;
-		match other {
-			Aborted => {
-				*self = Aborted;
-			}
-			Complete(b) => match self {
-				Aborted => {}
-				Complete(a) => {
-					a.merge(b);
-				}
-				Uploading(_) => {
-					*self = Complete(b.clone());
-				}
-			},
-			Uploading(_) => {}
-		}
-	}
-}
-
-/// Data stored in object version
-#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub enum ObjectVersionData {
-	/// The object was deleted, this Version is a tombstone to mark it as such
-	DeleteMarker,
-	/// The object is short, it's stored inlined
-	Inline(ObjectVersionMeta, #[serde(with = "serde_bytes")] Vec<u8>),
-	/// The object is not short, Hash of first block is stored here, next segments hashes are
-	/// stored in the version table
-	FirstBlock(ObjectVersionMeta, Hash),
-}
-
-impl AutoCrdt for ObjectVersionData {
-	const WARN_IF_DIFFERENT: bool = true;
-}
-
-/// Metadata about the object version
-#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub struct ObjectVersionMeta {
-	/// Headers to send to the client
-	pub headers: ObjectVersionHeaders,
-	/// Size of the object
-	pub size: u64,
-	/// etag of the object
-	pub etag: String,
-}
-
-/// Additional headers for an object
-#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub struct ObjectVersionHeaders {
-	/// Content type of the object
-	pub content_type: String,
-	/// Any other http headers to send
-	pub other: BTreeMap<String, String>,
-}
-
-impl ObjectVersion {
-	fn cmp_key(&self) -> (u64, Uuid) {
-		(self.timestamp, self.uuid)
-	}
-
-	/// Is the object version completely received
-	pub fn is_complete(&self) -> bool {
-		matches!(self.state, ObjectVersionState::Complete(_))
-	}
-}
-
-impl Crdt for Object {
-	fn merge(&mut self, other: &Self) {
-		// Merge versions from other into here
-		for other_v in other.versions.iter() {
-			match self
-				.versions
-				.binary_search_by(|v| v.cmp_key().cmp(&other_v.cmp_key()))
-			{
-				Ok(i) => {
-					self.versions[i].state.merge(&other_v.state);
-				}
-				Err(i) => {
-					self.versions.insert(i, other_v.clone());
-				}
-			}
-		}
-
-		// Remove versions which are obsolete, i.e. those that come
-		// before the last version which .is_complete().
-		let last_complete = self
-			.versions
-			.iter()
-			.enumerate()
-			.rev()
-			.find(|(_, v)| v.is_complete())
-			.map(|(vi, _)| vi);
-
-		if let Some(last_vi) = last_complete {
-			self.versions = self.versions.drain(last_vi..).collect::<Vec<_>>();
-		}
-	}
-}

src/model/prev/v051/version_table.rs

@@ -1,79 +0,0 @@
-use serde::{Deserialize, Serialize};
-
-use garage_util::data::*;
-
-use garage_table::crdt::*;
-use garage_table::*;
-
-/// A version of an object
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct Version {
-	/// UUID of the version, used as partition key
-	pub uuid: Uuid,
-
-	// Actual data: the blocks for this version
-	// In the case of a multipart upload, also store the etags
-	// of individual parts and check them when doing CompleteMultipartUpload
-	/// Is this version deleted
-	pub deleted: crdt::Bool,
-	/// list of blocks of data composing the version
-	pub blocks: crdt::Map<VersionBlockKey, VersionBlock>,
-	/// Etag of each part in case of a multipart upload, empty otherwise
-	pub parts_etags: crdt::Map<u64, String>,
-
-	// Back link to bucket+key so that we can figure if
-	// this was deleted later on
-	/// Bucket in which the related object is stored
-	pub bucket: String,
-	/// Key in which the related object is stored
-	pub key: String,
-}
-
-#[derive(PartialEq, Eq, Clone, Copy, Debug, Serialize, Deserialize)]
-pub struct VersionBlockKey {
-	/// Number of the part
-	pub part_number: u64,
-	/// Offset of this sub-segment in its part
-	pub offset: u64,
-}
-
-impl Ord for VersionBlockKey {
-	fn cmp(&self, other: &Self) -> std::cmp::Ordering {
-		self.part_number
-			.cmp(&other.part_number)
-			.then(self.offset.cmp(&other.offset))
-	}
-}
-
-impl PartialOrd for VersionBlockKey {
-	fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
-		Some(self.cmp(other))
-	}
-}
-
-/// Informations about a single block
-#[derive(PartialEq, Eq, Ord, PartialOrd, Clone, Copy, Debug, Serialize, Deserialize)]
-pub struct VersionBlock {
-	/// Blake2 sum of the block
-	pub hash: Hash,
-	/// Size of the block
-	pub size: u64,
-}
-
-impl AutoCrdt for VersionBlock {
-	const WARN_IF_DIFFERENT: bool = true;
-}
-
-impl Crdt for Version {
-	fn merge(&mut self, other: &Self) {
-		self.deleted.merge(&other.deleted);
-
-		if self.deleted.get() {
-			self.blocks.clear();
-			self.parts_etags.clear();
-		} else {
-			self.blocks.merge(&other.blocks);
-			self.parts_etags.merge(&other.parts_etags);
-		}
-	}
-}

src/model/s3/block_ref_table.rs

@@ -1,4 +1,3 @@
-use serde::{Deserialize, Serialize};
 use std::sync::Arc;
 
 use garage_db as db;
@@ -10,19 +9,29 @@ use garage_table::*;
 
 use garage_block::manager::*;
 
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+mod v08 {
-pub struct BlockRef {
+	use garage_util::crdt;
-	/// Hash (blake2 sum) of the block, used as partition key
+	use garage_util::data::{Hash, Uuid};
-	pub block: Hash,
+	use serde::{Deserialize, Serialize};
 
-	/// Id of the Version for the object containing this block, used as sorting key
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-	pub version: Uuid,
+	pub struct BlockRef {
+		/// Hash (blake2 sum) of the block, used as partition key
+		pub block: Hash,
 
-	// Keep track of deleted status
+		/// Id of the Version for the object containing this block, used as sorting key
-	/// Is the Version that contains this block deleted
+		pub version: Uuid,
-	pub deleted: crdt::Bool,
+
+		// Keep track of deleted status
+		/// Is the Version that contains this block deleted
+		pub deleted: crdt::Bool,
+	}
+
+	impl garage_util::migrate::InitialFormat for BlockRef {}
 }
 
+pub use v08::*;
+
 impl Entry<Hash, Uuid> for BlockRef {
 	fn partition_key(&self) -> &Hash {
 		&self.block

src/model/s3/object_table.rs

@@ -1,10 +1,8 @@
 use serde::{Deserialize, Serialize};
-use std::collections::BTreeMap;
 use std::sync::Arc;
 
 use garage_db as db;
 
-use garage_util::background::BackgroundRunner;
 use garage_util::data::*;
 
 use garage_table::crdt::*;
@@ -14,25 +12,126 @@ use garage_table::*;
 use crate::index_counter::*;
 use crate::s3::version_table::*;
 
-use crate::prev::v051::object_table as old;
-
 pub const OBJECTS: &str = "objects";
 pub const UNFINISHED_UPLOADS: &str = "unfinished_uploads";
 pub const BYTES: &str = "bytes";
 
-/// An object
+mod v05 {
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	use garage_util::data::{Hash, Uuid};
-pub struct Object {
+	use serde::{Deserialize, Serialize};
-	/// The bucket in which the object is stored, used as partition key
+	use std::collections::BTreeMap;
-	pub bucket_id: Uuid,
 
-	/// The key at which the object is stored in its bucket, used as sorting key
+	/// An object
-	pub key: String,
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct Object {
+		/// The bucket in which the object is stored, used as partition key
+		pub bucket: String,
 
-	/// The list of currenty stored versions of the object
+		/// The key at which the object is stored in its bucket, used as sorting key
-	versions: Vec<ObjectVersion>,
+		pub key: String,
+
+		/// The list of currenty stored versions of the object
+		pub(super) versions: Vec<ObjectVersion>,
+	}
+
+	/// Informations about a version of an object
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct ObjectVersion {
+		/// Id of the version
+		pub uuid: Uuid,
+		/// Timestamp of when the object was created
+		pub timestamp: u64,
+		/// State of the version
+		pub state: ObjectVersionState,
+	}
+
+	/// State of an object version
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub enum ObjectVersionState {
+		/// The version is being received
+		Uploading(ObjectVersionHeaders),
+		/// The version is fully received
+		Complete(ObjectVersionData),
+		/// The version uploaded containded errors or the upload was explicitly aborted
+		Aborted,
+	}
+
+	/// Data stored in object version
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
+	pub enum ObjectVersionData {
+		/// The object was deleted, this Version is a tombstone to mark it as such
+		DeleteMarker,
+		/// The object is short, it's stored inlined
+		Inline(ObjectVersionMeta, #[serde(with = "serde_bytes")] Vec<u8>),
+		/// The object is not short, Hash of first block is stored here, next segments hashes are
+		/// stored in the version table
+		FirstBlock(ObjectVersionMeta, Hash),
+	}
+
+	/// Metadata about the object version
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
+	pub struct ObjectVersionMeta {
+		/// Headers to send to the client
+		pub headers: ObjectVersionHeaders,
+		/// Size of the object
+		pub size: u64,
+		/// etag of the object
+		pub etag: String,
+	}
+
+	/// Additional headers for an object
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
+	pub struct ObjectVersionHeaders {
+		/// Content type of the object
+		pub content_type: String,
+		/// Any other http headers to send
+		pub other: BTreeMap<String, String>,
+	}
+
+	impl garage_util::migrate::InitialFormat for Object {}
 }
 
+mod v08 {
+	use garage_util::data::Uuid;
+	use serde::{Deserialize, Serialize};
+
+	use super::v05;
+
+	pub use v05::{
+		ObjectVersion, ObjectVersionData, ObjectVersionHeaders, ObjectVersionMeta,
+		ObjectVersionState,
+	};
+
+	/// An object
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct Object {
+		/// The bucket in which the object is stored, used as partition key
+		pub bucket_id: Uuid,
+
+		/// The key at which the object is stored in its bucket, used as sorting key
+		pub key: String,
+
+		/// The list of currenty stored versions of the object
+		pub(super) versions: Vec<ObjectVersion>,
+	}
+
+	impl garage_util::migrate::Migrate for Object {
+		type Previous = v05::Object;
+
+		fn migrate(old: v05::Object) -> Object {
+			use garage_util::data::blake2sum;
+
+			Object {
+				bucket_id: blake2sum(old.bucket.as_bytes()),
+				key: old.key,
+				versions: old.versions,
+			}
+		}
+	}
+}
+
+pub use v08::*;
+
 impl Object {
 	/// Initialize an Object struct from parts
 	pub fn new(bucket_id: Uuid, key: String, versions: Vec<ObjectVersion>) -> Self {
@@ -69,28 +168,6 @@ impl Object {
 	}
 }
 
-/// Informations about a version of an object
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct ObjectVersion {
-	/// Id of the version
-	pub uuid: Uuid,
-	/// Timestamp of when the object was created
-	pub timestamp: u64,
-	/// State of the version
-	pub state: ObjectVersionState,
-}
-
-/// State of an object version
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub enum ObjectVersionState {
-	/// The version is being received
-	Uploading(ObjectVersionHeaders),
-	/// The version is fully received
-	Complete(ObjectVersionData),
-	/// The version uploaded containded errors or the upload was explicitly aborted
-	Aborted,
-}
-
 impl Crdt for ObjectVersionState {
 	fn merge(&mut self, other: &Self) {
 		use ObjectVersionState::*;
@@ -112,42 +189,10 @@ impl Crdt for ObjectVersionState {
 	}
 }
 
-/// Data stored in object version
-#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub enum ObjectVersionData {
-	/// The object was deleted, this Version is a tombstone to mark it as such
-	DeleteMarker,
-	/// The object is short, it's stored inlined
-	Inline(ObjectVersionMeta, #[serde(with = "serde_bytes")] Vec<u8>),
-	/// The object is not short, Hash of first block is stored here, next segments hashes are
-	/// stored in the version table
-	FirstBlock(ObjectVersionMeta, Hash),
-}
-
 impl AutoCrdt for ObjectVersionData {
 	const WARN_IF_DIFFERENT: bool = true;
 }
 
-/// Metadata about the object version
-#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub struct ObjectVersionMeta {
-	/// Headers to send to the client
-	pub headers: ObjectVersionHeaders,
-	/// Size of the object
-	pub size: u64,
-	/// etag of the object
-	pub etag: String,
-}
-
-/// Additional headers for an object
-#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
-pub struct ObjectVersionHeaders {
-	/// Content type of the object
-	pub content_type: String,
-	/// Any other http headers to send
-	pub other: BTreeMap<String, String>,
-}
-
 impl ObjectVersion {
 	fn cmp_key(&self) -> (u64, Uuid) {
 		(self.timestamp, self.uuid)
@@ -221,7 +266,6 @@ impl Crdt for Object {
 }
 
 pub struct ObjectTable {
-	pub background: Arc<BackgroundRunner>,
 	pub version_table: Arc<Table<VersionTable, TableShardedReplication>>,
 	pub object_counter_table: Arc<IndexCounter<Object>>,
 }
@@ -255,34 +299,34 @@ impl TableSchema for ObjectTable {
 			);
 		}
 
-		// 2. Spawn threads that propagates deletions to version table
+		// 2. Enqueue propagation deletions to version table
-		let version_table = self.version_table.clone();
+		if let (Some(old_v), Some(new_v)) = (old, new) {
-		let old = old.cloned();
+			// Propagate deletion of old versions
-		let new = new.cloned();
+			for v in old_v.versions.iter() {
-
+				let newly_deleted = match new_v
-		self.background.spawn(async move {
+					.versions
-			if let (Some(old_v), Some(new_v)) = (old, new) {
+					.binary_search_by(|nv| nv.cmp_key().cmp(&v.cmp_key()))
-				// Propagate deletion of old versions
+				{
-				for v in old_v.versions.iter() {
+					Err(_) => true,
-					let newly_deleted = match new_v
+					Ok(i) => {
-						.versions
+						new_v.versions[i].state == ObjectVersionState::Aborted
-						.binary_search_by(|nv| nv.cmp_key().cmp(&v.cmp_key()))
+							&& v.state != ObjectVersionState::Aborted
-					{
+					}
-						Err(_) => true,
+				};
-						Ok(i) => {
+				if newly_deleted {
-							new_v.versions[i].state == ObjectVersionState::Aborted
+					let deleted_version =
-								&& v.state != ObjectVersionState::Aborted
+						Version::new(v.uuid, old_v.bucket_id, old_v.key.clone(), true);
-						}
+					let res = self.version_table.queue_insert(tx, &deleted_version);
-					};
+					if let Err(e) = db::unabort(res)? {
-					if newly_deleted {
+						error!(
-						let deleted_version =
+							"Unable to enqueue version deletion propagation: {}. A repair will be needed.",
-							Version::new(v.uuid, old_v.bucket_id, old_v.key.clone(), true);
+							e
-						version_table.insert(&deleted_version).await?;
+						);
 					}
 				}
 			}
-			Ok(())
+		}
-		});
+
 		Ok(())
 	}
 
@@ -292,11 +336,6 @@ impl TableSchema for ObjectTable {
 			ObjectFilter::IsUploading => entry.versions.iter().any(|v| v.is_uploading()),
 		}
 	}
-
-	fn try_migrate(bytes: &[u8]) -> Option<Self::E> {
-		let old_obj = rmp_serde::decode::from_read_ref::<_, old::Object>(bytes).ok()?;
-		Some(migrate_object(old_obj))
-	}
 }
 
 impl CountedItem for Object {
@@ -341,64 +380,3 @@ impl CountedItem for Object {
 		]
 	}
 }
-
-// vvvvvvvv migration code, stupid stuff vvvvvvvvvvvv
-// (we just want to change bucket into bucket_id by hashing it)
-
-fn migrate_object(o: old::Object) -> Object {
-	let versions = o
-		.versions()
-		.iter()
-		.cloned()
-		.map(migrate_object_version)
-		.collect();
-	Object {
-		bucket_id: blake2sum(o.bucket.as_bytes()),
-		key: o.key,
-		versions,
-	}
-}
-
-fn migrate_object_version(v: old::ObjectVersion) -> ObjectVersion {
-	ObjectVersion {
-		uuid: Uuid::try_from(v.uuid.as_slice()).unwrap(),
-		timestamp: v.timestamp,
-		state: match v.state {
-			old::ObjectVersionState::Uploading(h) => {
-				ObjectVersionState::Uploading(migrate_object_version_headers(h))
-			}
-			old::ObjectVersionState::Complete(d) => {
-				ObjectVersionState::Complete(migrate_object_version_data(d))
-			}
-			old::ObjectVersionState::Aborted => ObjectVersionState::Aborted,
-		},
-	}
-}
-
-fn migrate_object_version_headers(h: old::ObjectVersionHeaders) -> ObjectVersionHeaders {
-	ObjectVersionHeaders {
-		content_type: h.content_type,
-		other: h.other,
-	}
-}
-
-fn migrate_object_version_data(d: old::ObjectVersionData) -> ObjectVersionData {
-	match d {
-		old::ObjectVersionData::DeleteMarker => ObjectVersionData::DeleteMarker,
-		old::ObjectVersionData::Inline(m, b) => {
-			ObjectVersionData::Inline(migrate_object_version_meta(m), b)
-		}
-		old::ObjectVersionData::FirstBlock(m, h) => ObjectVersionData::FirstBlock(
-			migrate_object_version_meta(m),
-			Hash::try_from(h.as_slice()).unwrap(),
-		),
-	}
-}
-
-fn migrate_object_version_meta(m: old::ObjectVersionMeta) -> ObjectVersionMeta {
-	ObjectVersionMeta {
-		headers: migrate_object_version_headers(m.headers),
-		size: m.size,
-		etag: m.etag,
-	}
-}

src/model/s3/version_table.rs

@@ -1,9 +1,7 @@
-use serde::{Deserialize, Serialize};
 use std::sync::Arc;
 
 use garage_db as db;
 
-use garage_util::background::BackgroundRunner;
 use garage_util::data::*;
 
 use garage_table::crdt::*;
@@ -12,32 +10,108 @@ use garage_table::*;
 
 use crate::s3::block_ref_table::*;
 
-use crate::prev::v051::version_table as old;
+mod v05 {
+	use garage_util::crdt;
+	use garage_util::data::{Hash, Uuid};
+	use serde::{Deserialize, Serialize};
 
-/// A version of an object
+	/// A version of an object
-#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
-pub struct Version {
+	pub struct Version {
-	/// UUID of the version, used as partition key
+		/// UUID of the version, used as partition key
-	pub uuid: Uuid,
+		pub uuid: Uuid,
 
-	// Actual data: the blocks for this version
+		// Actual data: the blocks for this version
-	// In the case of a multipart upload, also store the etags
+		// In the case of a multipart upload, also store the etags
-	// of individual parts and check them when doing CompleteMultipartUpload
+		// of individual parts and check them when doing CompleteMultipartUpload
-	/// Is this version deleted
+		/// Is this version deleted
-	pub deleted: crdt::Bool,
+		pub deleted: crdt::Bool,
-	/// list of blocks of data composing the version
+		/// list of blocks of data composing the version
-	pub blocks: crdt::Map<VersionBlockKey, VersionBlock>,
+		pub blocks: crdt::Map<VersionBlockKey, VersionBlock>,
-	/// Etag of each part in case of a multipart upload, empty otherwise
+		/// Etag of each part in case of a multipart upload, empty otherwise
-	pub parts_etags: crdt::Map<u64, String>,
+		pub parts_etags: crdt::Map<u64, String>,
 
-	// Back link to bucket+key so that we can figure if
+		// Back link to bucket+key so that we can figure if
-	// this was deleted later on
+		// this was deleted later on
-	/// Bucket in which the related object is stored
+		/// Bucket in which the related object is stored
-	pub bucket_id: Uuid,
+		pub bucket: String,
-	/// Key in which the related object is stored
+		/// Key in which the related object is stored
-	pub key: String,
+		pub key: String,
+	}
+
+	#[derive(PartialEq, Eq, Clone, Copy, Debug, Serialize, Deserialize)]
+	pub struct VersionBlockKey {
+		/// Number of the part
+		pub part_number: u64,
+		/// Offset of this sub-segment in its part
+		pub offset: u64,
+	}
+
+	/// Informations about a single block
+	#[derive(PartialEq, Eq, Ord, PartialOrd, Clone, Copy, Debug, Serialize, Deserialize)]
+	pub struct VersionBlock {
+		/// Blake2 sum of the block
+		pub hash: Hash,
+		/// Size of the block
+		pub size: u64,
+	}
+
+	impl garage_util::migrate::InitialFormat for Version {}
 }
 
+mod v08 {
+	use garage_util::crdt;
+	use garage_util::data::Uuid;
+	use serde::{Deserialize, Serialize};
+
+	use super::v05;
+
+	/// A version of an object
+	#[derive(PartialEq, Eq, Clone, Debug, Serialize, Deserialize)]
+	pub struct Version {
+		/// UUID of the version, used as partition key
+		pub uuid: Uuid,
+
+		// Actual data: the blocks for this version
+		// In the case of a multipart upload, also store the etags
+		// of individual parts and check them when doing CompleteMultipartUpload
+		/// Is this version deleted
+		pub deleted: crdt::Bool,
+		/// list of blocks of data composing the version
+		pub blocks: crdt::Map<VersionBlockKey, VersionBlock>,
+		/// Etag of each part in case of a multipart upload, empty otherwise
+		pub parts_etags: crdt::Map<u64, String>,
+
+		// Back link to bucket+key so that we can figure if
+		// this was deleted later on
+		/// Bucket in which the related object is stored
+		pub bucket_id: Uuid,
+		/// Key in which the related object is stored
+		pub key: String,
+	}
+
+	pub use v05::{VersionBlock, VersionBlockKey};
+
+	impl garage_util::migrate::Migrate for Version {
+		type Previous = v05::Version;
+
+		fn migrate(old: v05::Version) -> Version {
+			use garage_util::data::blake2sum;
+
+			Version {
+				uuid: old.uuid,
+				deleted: old.deleted,
+				blocks: old.blocks,
+				parts_etags: old.parts_etags,
+				bucket_id: blake2sum(old.bucket.as_bytes()),
+				key: old.key,
+			}
+		}
+	}
+}
+
+pub use v08::*;
+
 impl Version {
 	pub fn new(uuid: Uuid, bucket_id: Uuid, key: String, deleted: bool) -> Self {
 		Self {
@@ -65,14 +139,6 @@ impl Version {
 	}
 }
 
-#[derive(PartialEq, Eq, Clone, Copy, Debug, Serialize, Deserialize)]
-pub struct VersionBlockKey {
-	/// Number of the part
-	pub part_number: u64,
-	/// Offset of this sub-segment in its part
-	pub offset: u64,
-}
-
 impl Ord for VersionBlockKey {
 	fn cmp(&self, other: &Self) -> std::cmp::Ordering {
 		self.part_number
@@ -87,15 +153,6 @@ impl PartialOrd for VersionBlockKey {
 	}
 }
 
-/// Informations about a single block
-#[derive(PartialEq, Eq, Ord, PartialOrd, Clone, Copy, Debug, Serialize, Deserialize)]
-pub struct VersionBlock {
-	/// Blake2 sum of the block
-	pub hash: Hash,
-	/// Size of the block
-	pub size: u64,
-}
-
 impl AutoCrdt for VersionBlock {
 	const WARN_IF_DIFFERENT: bool = true;
 }
@@ -127,7 +184,6 @@ impl Crdt for Version {
 }
 
 pub struct VersionTable {
-	pub background: Arc<BackgroundRunner>,
 	pub block_ref_table: Arc<Table<BlockRefTable, TableShardedReplication>>,
 }
 
@@ -141,33 +197,26 @@ impl TableSchema for VersionTable {
 
 	fn updated(
 		&self,
-		_tx: &mut db::Transaction,
+		tx: &mut db::Transaction,
 		old: Option<&Self::E>,
 		new: Option<&Self::E>,
 	) -> db::TxOpResult<()> {
-		let block_ref_table = self.block_ref_table.clone();
+		if let (Some(old_v), Some(new_v)) = (old, new) {
-		let old = old.cloned();
+			// Propagate deletion of version blocks
-		let new = new.cloned();
+			if new_v.deleted.get() && !old_v.deleted.get() {
-
+				let deleted_block_refs = old_v.blocks.items().iter().map(|(_k, vb)| BlockRef {
-		self.background.spawn(async move {
+					block: vb.hash,
-			if let (Some(old_v), Some(new_v)) = (old, new) {
+					version: old_v.uuid,
-				// Propagate deletion of version blocks
+					deleted: true.into(),
-				if new_v.deleted.get() && !old_v.deleted.get() {
+				});
-					let deleted_block_refs = old_v
+				for block_ref in deleted_block_refs {
-						.blocks
+					let res = self.block_ref_table.queue_insert(tx, &block_ref);
-						.items()
+					if let Err(e) = db::unabort(res)? {
-						.iter()
+						error!("Unable to enqueue block ref deletion propagation: {}. A repair will be needed.", e);
-						.map(|(_k, vb)| BlockRef {
+					}
-							block: vb.hash,
-							version: old_v.uuid,
-							deleted: true.into(),
-						})
-						.collect::<Vec<_>>();
-					block_ref_table.insert_many(&deleted_block_refs[..]).await?;
 				}
 			}
-			Ok(())
+		}
-		});
 
 		Ok(())
 	}
@@ -175,42 +224,4 @@ impl TableSchema for VersionTable {
 	fn matches_filter(entry: &Self::E, filter: &Self::Filter) -> bool {
 		filter.apply(entry.deleted.get())
 	}
-
-	fn try_migrate(bytes: &[u8]) -> Option<Self::E> {
-		let old = rmp_serde::decode::from_read_ref::<_, old::Version>(bytes).ok()?;
-
-		let blocks = old
-			.blocks
-			.items()
-			.iter()
-			.map(|(k, v)| {
-				(
-					VersionBlockKey {
-						part_number: k.part_number,
-						offset: k.offset,
-					},
-					VersionBlock {
-						hash: Hash::try_from(v.hash.as_slice()).unwrap(),
-						size: v.size,
-					},
-				)
-			})
-			.collect::<crdt::Map<_, _>>();
-
-		let parts_etags = old
-			.parts_etags
-			.items()
-			.iter()
-			.map(|(k, v)| (*k, v.clone()))
-			.collect::<crdt::Map<_, _>>();
-
-		Some(Version {
-			uuid: Hash::try_from(old.uuid.as_slice()).unwrap(),
-			deleted: crdt::Bool::new(old.deleted.get()),
-			blocks,
-			parts_etags,
-			bucket_id: blake2sum(old.bucket.as_bytes()),
-			key: old.key,
-		})
-	}
 }

src/rpc/Cargo.toml

@@ -25,7 +25,6 @@ rand = "0.8"
 sodiumoxide = { version = "0.2.5-0", package = "kuska-sodiumoxide" }
 
 async-trait = "0.1.7"
-rmp-serde = "0.15"
 serde = { version = "1.0", default-features = false, features = ["derive", "rc"] }
 serde_bytes = "0.11"
 serde_json = "1.0"

src/rpc/layout.rs

@@ -5,6 +5,7 @@ use serde::{Deserialize, Serialize};
 
 use garage_util::crdt::{AutoCrdt, Crdt, LwwMap};
 use garage_util::data::*;
+use garage_util::encode::nonversioned_encode;
 use garage_util::error::*;
 
 use crate::ring::*;
@@ -35,6 +36,8 @@ pub struct ClusterLayout {
 	pub staging_hash: Hash,
 }
 
+impl garage_util::migrate::InitialFormat for ClusterLayout {}
+
 #[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
 pub struct NodeRoleV(pub Option<NodeRole>);
 
@@ -68,7 +71,7 @@ impl NodeRole {
 impl ClusterLayout {
 	pub fn new(replication_factor: usize) -> Self {
 		let empty_lwwmap = LwwMap::new();
-		let empty_lwwmap_hash = blake2sum(&rmp_to_vec_all_named(&empty_lwwmap).unwrap()[..]);
+		let empty_lwwmap_hash = blake2sum(&nonversioned_encode(&empty_lwwmap).unwrap()[..]);
 
 		ClusterLayout {
 			version: 0,
@@ -90,7 +93,7 @@ impl ClusterLayout {
 			Ordering::Equal => {
 				self.staging.merge(&other.staging);
 
-				let new_staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
+				let new_staging_hash = blake2sum(&nonversioned_encode(&self.staging).unwrap()[..]);
 				let changed = new_staging_hash != self.staging_hash;
 
 				self.staging_hash = new_staging_hash;
@@ -125,7 +128,7 @@ To know the correct value of the new layout version, invoke `garage layout show`
 		}
 
 		self.staging.clear();
-		self.staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
+		self.staging_hash = blake2sum(&nonversioned_encode(&self.staging).unwrap()[..]);
 
 		self.version += 1;
 
@@ -149,7 +152,7 @@ To know the correct value of the new layout version, invoke `garage layout show`
 		}
 
 		self.staging.clear();
-		self.staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
+		self.staging_hash = blake2sum(&nonversioned_encode(&self.staging).unwrap()[..]);
 
 		self.version += 1;
 
@@ -178,7 +181,7 @@ To know the correct value of the new layout version, invoke `garage layout show`
 	/// returns true if consistent, false if error
 	pub fn check(&self) -> bool {
 		// Check that the hash of the staging data is correct
-		let staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
+		let staging_hash = blake2sum(&nonversioned_encode(&self.staging).unwrap()[..]);
 		if staging_hash != self.staging_hash {
 			return false;
 		}

src/rpc/rpc_helper.rs

@@ -5,7 +5,6 @@ use std::time::Duration;
 use futures::future::join_all;
 use futures::stream::futures_unordered::FuturesUnordered;
 use futures::stream::StreamExt;
-use futures_util::future::FutureExt;
 use tokio::select;
 use tokio::sync::watch;
 
@@ -24,7 +23,6 @@ pub use netapp::message::{
 use netapp::peering::fullmesh::FullMeshPeeringStrategy;
 pub use netapp::{self, NetApp, NodeID};
 
-use garage_util::background::BackgroundRunner;
 use garage_util::data::*;
 use garage_util::error::Error;
 use garage_util::metrics::RecordDuration;
@@ -94,7 +92,6 @@ pub struct RpcHelper(Arc<RpcHelperInner>);
 struct RpcHelperInner {
 	our_node_id: Uuid,
 	fullmesh: Arc<FullMeshPeeringStrategy>,
-	background: Arc<BackgroundRunner>,
 	ring: watch::Receiver<Arc<Ring>>,
 	metrics: RpcMetrics,
 	rpc_timeout: Duration,
@@ -104,7 +101,6 @@ impl RpcHelper {
 	pub(crate) fn new(
 		our_node_id: Uuid,
 		fullmesh: Arc<FullMeshPeeringStrategy>,
-		background: Arc<BackgroundRunner>,
 		ring: watch::Receiver<Arc<Ring>>,
 		rpc_timeout: Option<Duration>,
 	) -> Self {
@@ -113,7 +109,6 @@ impl RpcHelper {
 		Self(Arc::new(RpcHelperInner {
 			our_node_id,
 			fullmesh,
-			background,
 			ring,
 			metrics,
 			rpc_timeout: rpc_timeout.unwrap_or(DEFAULT_TIMEOUT),
@@ -377,16 +372,13 @@ impl RpcHelper {
 
 			if !resp_stream.is_empty() {
 				// Continue remaining requests in background.
-				// Continue the remaining requests immediately using tokio::spawn
+				// Note: these requests can get interrupted on process shutdown,
-				// but enqueue a task in the background runner
+				// we must not count on them being executed for certain.
-				// to ensure that the process won't exit until the requests are done
+				// For all background things that have to happen with certainty,
-				// (if we had just enqueued the resp_stream.collect directly in the background runner,
+				// they have to be put in a proper queue that is persisted to disk.
-				// the requests might have been put on hold in the background runner's queue,
+				tokio::spawn(async move {
-				// in which case they might timeout or otherwise fail)
-				let wait_finished_fut = tokio::spawn(async move {
 					resp_stream.collect::<Vec<Result<_, _>>>().await;
 				});
-				self.0.background.spawn(wait_finished_fut.map(|_| Ok(())));
 			}
 		}
 

src/rpc/system.rs

@@ -21,7 +21,6 @@ use netapp::peering::fullmesh::FullMeshPeeringStrategy;
 use netapp::util::parse_and_resolve_peer_addr_async;
 use netapp::{NetApp, NetworkKey, NodeID, NodeKey};
 
-use garage_util::background::BackgroundRunner;
 use garage_util::config::Config;
 #[cfg(feature = "kubernetes-discovery")]
 use garage_util::config::KubernetesDiscoveryConfig;
@@ -50,8 +49,6 @@ pub const GARAGE_VERSION_TAG: u64 = 0x6761726167650008; // garage 0x0008
 /// RPC endpoint used for calls related to membership
 pub const SYSTEM_RPC_PATH: &str = "garage_rpc/membership.rs/SystemRpc";
 
-pub const CONNECT_ERROR_MESSAGE: &str = "Error establishing RPC connection to remote node. This can happen if the remote node is not reachable on the network, but also if the two nodes are not configured with the same rpc_secret";
-
 /// RPC messages related to membership
 #[derive(Debug, Serialize, Deserialize, Clone)]
 pub enum SystemRpc {
@@ -76,13 +73,17 @@ impl Rpc for SystemRpc {
 	type Response = Result<SystemRpc, Error>;
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct PeerList(Vec<(Uuid, SocketAddr)>);
+impl garage_util::migrate::InitialFormat for PeerList {}
+
 /// This node's membership manager
 pub struct System {
 	/// The id of this node
 	pub id: Uuid,
 
 	persist_cluster_layout: Persister<ClusterLayout>,
-	persist_peer_list: Persister<Vec<(Uuid, SocketAddr)>>,
+	persist_peer_list: Persister<PeerList>,
 
 	local_status: ArcSwap<NodeStatus>,
 	node_status: RwLock<HashMap<Uuid, (u64, NodeStatus)>>,
@@ -110,9 +111,6 @@ pub struct System {
 	pub ring: watch::Receiver<Arc<Ring>>,
 	update_ring: Mutex<watch::Sender<Arc<Ring>>>,
 
-	/// The job runner of this node
-	pub background: Arc<BackgroundRunner>,
-
 	/// Path to metadata directory
 	pub metadata_dir: PathBuf,
 }
@@ -232,7 +230,6 @@ impl System {
 	/// Create this node's membership manager
 	pub fn new(
 		network_key: NetworkKey,
-		background: Arc<BackgroundRunner>,
 		replication_mode: ReplicationMode,
 		config: &Config,
 	) -> Result<Arc<Self>, Error> {
@@ -354,7 +351,6 @@ impl System {
 			rpc: RpcHelper::new(
 				netapp.id.into(),
 				fullmesh,
-				background.clone(),
 				ring.clone(),
 				config.rpc_timeout_msec.map(Duration::from_millis),
 			),
@@ -372,7 +368,6 @@ impl System {
 
 			ring,
 			update_ring: Mutex::new(update_ring),
-			background,
 			metadata_dir: config.metadata_dir.clone(),
 		});
 		sys.system_endpoint.set_handler(sys.clone());
@@ -444,17 +439,14 @@ impl System {
 				))
 			})?;
 		let mut errors = vec![];
-		for ip in addrs.iter() {
+		for addr in addrs.iter() {
-			match self
+			match self.netapp.clone().try_connect(*addr, pubkey).await {
-				.netapp
-				.clone()
-				.try_connect(*ip, pubkey)
-				.await
-				.err_context(CONNECT_ERROR_MESSAGE)
-			{
 				Ok(()) => return Ok(()),
 				Err(e) => {
-					errors.push((*ip, e));
+					errors.push((
+						*addr,
+						Error::Message(connect_error_message(*addr, pubkey, e)),
+					));
 				}
 			}
 		}
@@ -529,56 +521,61 @@ impl System {
 	// ---- INTERNALS ----
 
 	#[cfg(feature = "consul-discovery")]
-	async fn advertise_to_consul(self: Arc<Self>) -> Result<(), Error> {
+	async fn advertise_to_consul(self: Arc<Self>) {
 		let c = match &self.consul_discovery {
 			Some(c) => c,
-			_ => return Ok(()),
+			_ => return,
 		};
 
 		let rpc_public_addr = match self.rpc_public_addr {
 			Some(addr) => addr,
 			None => {
 				warn!("Not advertising to Consul because rpc_public_addr is not defined in config file and could not be autodetected.");
-				return Ok(());
+				return;
 			}
 		};
 
-		c.publish_consul_service(
+		if let Err(e) = c
-			self.netapp.id,
+			.publish_consul_service(
-			&self.local_status.load_full().hostname,
+				self.netapp.id,
-			rpc_public_addr,
+				&self.local_status.load_full().hostname,
-		)
+				rpc_public_addr,
-		.await
+			)
-		.err_context("Error while publishing Consul service")
+			.await
+		{
+			error!("Error while publishing Consul service: {}", e);
+		}
 	}
 
 	#[cfg(feature = "kubernetes-discovery")]
-	async fn advertise_to_kubernetes(self: Arc<Self>) -> Result<(), Error> {
+	async fn advertise_to_kubernetes(self: Arc<Self>) {
 		let k = match &self.kubernetes_discovery {
 			Some(k) => k,
-			_ => return Ok(()),
+			_ => return,
 		};
 
 		let rpc_public_addr = match self.rpc_public_addr {
 			Some(addr) => addr,
 			None => {
 				warn!("Not advertising to Kubernetes because rpc_public_addr is not defined in config file and could not be autodetected.");
-				return Ok(());
+				return;
 			}
 		};
 
-		publish_kubernetes_node(
+		if let Err(e) = publish_kubernetes_node(
 			k,
 			self.netapp.id,
 			&self.local_status.load_full().hostname,
 			rpc_public_addr,
 		)
 		.await
-		.err_context("Error while publishing node to kubernetes")
+		{
+			error!("Error while publishing node to Kubernetes: {}", e);
+		}
 	}
 
 	/// Save network configuration to disc
-	async fn save_cluster_layout(self: Arc<Self>) -> Result<(), Error> {
+	async fn save_cluster_layout(&self) -> Result<(), Error> {
 		let ring: Arc<Ring> = self.ring.borrow().clone();
 		self.persist_cluster_layout
 			.save_async(&ring.layout)
@@ -630,11 +627,7 @@ impl System {
 		if info.cluster_layout_version > local_info.cluster_layout_version
 			|| info.cluster_layout_staging_hash != local_info.cluster_layout_staging_hash
 		{
-			let self2 = self.clone();
+			tokio::spawn(self.clone().pull_cluster_layout(from));
-			self.background.spawn_cancellable(async move {
-				self2.pull_cluster_layout(from).await;
-				Ok(())
-			});
 		}
 
 		self.node_status
@@ -676,18 +669,21 @@ impl System {
 			drop(update_ring);
 
 			let self2 = self.clone();
-			self.background.spawn_cancellable(async move {
+			tokio::spawn(async move {
-				self2
+				if let Err(e) = self2
 					.rpc
 					.broadcast(
 						&self2.system_endpoint,
 						SystemRpc::AdvertiseClusterLayout(layout),
 						RequestStrategy::with_priority(PRIO_HIGH),
 					)
-					.await?;
+					.await
-				Ok(())
+				{
+					warn!("Error while broadcasting new cluster layout: {}", e);
+				}
 			});
-			self.background.spawn(self.clone().save_cluster_layout());
+
+			self.save_cluster_layout().await?;
 		}
 
 		Ok(SystemRpc::Ok)
@@ -734,7 +730,7 @@ impl System {
 
 				// Add peer list from list stored on disk
 				if let Ok(peers) = self.persist_peer_list.load_async().await {
-					ping_list.extend(peers.iter().map(|(id, addr)| ((*id).into(), *addr)))
+					ping_list.extend(peers.0.iter().map(|(id, addr)| ((*id).into(), *addr)))
 				}
 
 				// Fetch peer list from Consul
@@ -773,12 +769,12 @@ impl System {
 				}
 
 				for (node_id, node_addr) in ping_list {
-					tokio::spawn(
+					let self2 = self.clone();
-						self.netapp
+					tokio::spawn(async move {
-							.clone()
+						if let Err(e) = self2.netapp.clone().try_connect(node_addr, node_id).await {
-							.try_connect(node_addr, node_id)
+							error!("{}", connect_error_message(node_addr, node_id, e));
-							.map(|r| r.err_context(CONNECT_ERROR_MESSAGE)),
+						}
-					);
+					});
 				}
 			}
 
@@ -787,11 +783,10 @@ impl System {
 			}
 
 			#[cfg(feature = "consul-discovery")]
-			self.background.spawn(self.clone().advertise_to_consul());
+			tokio::spawn(self.clone().advertise_to_consul());
 
 			#[cfg(feature = "kubernetes-discovery")]
-			self.background
+			tokio::spawn(self.clone().advertise_to_kubernetes());
-				.spawn(self.clone().advertise_to_kubernetes());
 
 			let restart_at = tokio::time::sleep(DISCOVERY_INTERVAL);
 			select! {
@@ -815,12 +810,16 @@ impl System {
 		// and append it to the list we are about to save,
 		// so that no peer ID gets lost in the process.
 		if let Ok(mut prev_peer_list) = self.persist_peer_list.load_async().await {
-			prev_peer_list.retain(|(id, _ip)| peer_list.iter().all(|(id2, _ip2)| id2 != id));
+			prev_peer_list
-			peer_list.extend(prev_peer_list);
+				.0
+				.retain(|(id, _ip)| peer_list.iter().all(|(id2, _ip2)| id2 != id));
+			peer_list.extend(prev_peer_list.0);
 		}
 
 		// Save new peer list to file
-		self.persist_peer_list.save_async(&peer_list).await
+		self.persist_peer_list
+			.save_async(&PeerList(peer_list))
+			.await
 	}
 
 	async fn pull_cluster_layout(self: Arc<Self>, peer: Uuid) {
@@ -881,3 +880,11 @@ async fn resolve_peers(peers: &[String]) -> Vec<(NodeID, SocketAddr)> {
 
 	ret
 }
+
+fn connect_error_message(
+	addr: SocketAddr,
+	pubkey: ed25519::PublicKey,
+	e: netapp::error::Error,
+) -> String {
+	format!("Error establishing RPC connection to remote node: {}@{}.\nThis can happen if the remote node is not reachable on the network, but also if the two nodes are not configured with the same rpc_secret.\n{}", hex::encode(pubkey), addr, e)
+}

src/table/Cargo.toml

@@ -21,13 +21,13 @@ garage_util = { version = "0.8.1", path = "../util" }
 opentelemetry = "0.17"
 
 async-trait = "0.1.7"
+arc-swap = "1.0"
 bytes = "1.0"
 hex = "0.4"
 hexdump = "0.1"
 tracing = "0.1.30"
 rand = "0.8"
 
-rmp-serde = "0.15"
 serde = { version = "1.0", default-features = false, features = ["derive", "rc"] }
 serde_bytes = "0.11"
 

src/table/data.rs

@@ -10,6 +10,7 @@ use garage_db::counted_tree_hack::CountedTree;
 
 use garage_util::data::*;
 use garage_util::error::*;
+use garage_util::migrate::Migrate;
 
 use garage_rpc::system::System;
 
@@ -31,16 +32,16 @@ pub struct TableData<F: TableSchema, R: TableReplication> {
 	pub(crate) merkle_tree: db::Tree,
 	pub(crate) merkle_todo: db::Tree,
 	pub(crate) merkle_todo_notify: Notify,
+
+	pub(crate) insert_queue: db::Tree,
+	pub(crate) insert_queue_notify: Notify,
+
 	pub(crate) gc_todo: CountedTree,
 
 	pub(crate) metrics: TableMetrics,
 }
 
-impl<F, R> TableData<F, R>
+impl<F: TableSchema, R: TableReplication> TableData<F, R> {
-where
-	F: TableSchema,
-	R: TableReplication,
-{
 	pub fn new(system: Arc<System>, instance: F, replication: R, db: &db::Db) -> Arc<Self> {
 		let store = db
 			.open_tree(&format!("{}:table", F::TABLE_NAME))
@@ -53,9 +54,13 @@ where
 			.open_tree(&format!("{}:merkle_todo", F::TABLE_NAME))
 			.expect("Unable to open DB Merkle TODO tree");
 
+		let insert_queue = db
+			.open_tree(&format!("{}:insert_queue", F::TABLE_NAME))
+			.expect("Unable to open insert queue DB tree");
+
 		let gc_todo = db
 			.open_tree(&format!("{}:gc_todo_v2", F::TABLE_NAME))
-			.expect("Unable to open DB tree");
+			.expect("Unable to open GC DB tree");
 		let gc_todo = CountedTree::new(gc_todo).expect("Cannot count gc_todo_v2");
 
 		let metrics = TableMetrics::new(
@@ -74,6 +79,8 @@ where
 			merkle_tree,
 			merkle_todo,
 			merkle_todo_notify: Notify::new(),
+			insert_queue,
+			insert_queue_notify: Notify::new(),
 			gc_todo,
 			metrics,
 		})
@@ -173,50 +180,57 @@ where
 
 	pub(crate) fn update_entry(&self, update_bytes: &[u8]) -> Result<(), Error> {
 		let update = self.decode_entry(update_bytes)?;
-		let tree_key = self.tree_key(update.partition_key(), update.sort_key());
 
-		self.update_entry_with(&tree_key[..], |ent| match ent {
+		self.update_entry_with(
-			Some(mut ent) => {
+			update.partition_key(),
-				ent.merge(&update);
+			update.sort_key(),
-				ent
+			|_tx, ent| match ent {
-			}
+				Some(mut ent) => {
-			None => update.clone(),
+					ent.merge(&update);
-		})?;
+					Ok(ent)
+				}
+				None => Ok(update.clone()),
+			},
+		)?;
 		Ok(())
 	}
 
 	pub fn update_entry_with(
 		&self,
-		tree_key: &[u8],
+		partition_key: &F::P,
-		f: impl Fn(Option<F::E>) -> F::E,
+		sort_key: &F::S,
+		update_fn: impl Fn(&mut db::Transaction, Option<F::E>) -> db::TxResult<F::E, Error>,
 	) -> Result<Option<F::E>, Error> {
+		let tree_key = self.tree_key(partition_key, sort_key);
+
 		let changed = self.store.db().transaction(|mut tx| {
-			let (old_entry, old_bytes, new_entry) = match tx.get(&self.store, tree_key)? {
+			let (old_entry, old_bytes, new_entry) = match tx.get(&self.store, &tree_key)? {
 				Some(old_bytes) => {
 					let old_entry = self.decode_entry(&old_bytes).map_err(db::TxError::Abort)?;
-					let new_entry = f(Some(old_entry.clone()));
+					let new_entry = update_fn(&mut tx, Some(old_entry.clone()))?;
 					(Some(old_entry), Some(old_bytes), new_entry)
 				}
-				None => (None, None, f(None)),
+				None => (None, None, update_fn(&mut tx, None)?),
 			};
 
-			// Scenario 1: the value changed, so of course there is a change
+			// Changed can be true in two scenarios
-			let value_changed = Some(&new_entry) != old_entry.as_ref();
+			// Scenario 1: the actual represented value changed,
-
+			//   so of course the messagepack encoding changed as well
 			// Scenario 2: the value didn't change but due to a migration in the
-			// data format, the messagepack encoding changed. In this case
+			//   data format, the messagepack encoding changed. In this case,
-			// we have to write the migrated value in the table and update
+			//   we also have to write the migrated value in the table and update
-			// the associated Merkle tree entry.
+			//   the associated Merkle tree entry.
-			let new_bytes = rmp_to_vec_all_named(&new_entry)
+			let new_bytes = new_entry
+				.encode()
 				.map_err(Error::RmpEncode)
 				.map_err(db::TxError::Abort)?;
-			let encoding_changed = Some(&new_bytes[..]) != old_bytes.as_ref().map(|x| &x[..]);
+			let changed = Some(&new_bytes[..]) != old_bytes.as_deref();
 			drop(old_bytes);
 
-			if value_changed || encoding_changed {
+			if changed {
-				let new_bytes_hash = blake2sum(&new_bytes[..]);
+				let new_bytes_hash = blake2sum(&new_bytes);
-				tx.insert(&self.merkle_todo, tree_key, new_bytes_hash.as_slice())?;
+				tx.insert(&self.merkle_todo, &tree_key, new_bytes_hash.as_slice())?;
-				tx.insert(&self.store, tree_key, new_bytes)?;
+				tx.insert(&self.store, &tree_key, new_bytes)?;
 
 				self.instance
 					.updated(&mut tx, old_entry.as_ref(), Some(&new_entry))?;
@@ -242,7 +256,7 @@ where
 				let pk_hash = Hash::try_from(&tree_key[..32]).unwrap();
 				let nodes = self.replication.write_nodes(&pk_hash);
 				if nodes.first() == Some(&self.system.id) {
-					GcTodoEntry::new(tree_key.to_vec(), new_bytes_hash).save(&self.gc_todo)?;
+					GcTodoEntry::new(tree_key, new_bytes_hash).save(&self.gc_todo)?;
 				}
 			}
 
@@ -258,10 +272,11 @@ where
 			.db()
 			.transaction(|mut tx| match tx.get(&self.store, k)? {
 				Some(cur_v) if cur_v == v => {
+					let old_entry = self.decode_entry(v).map_err(db::TxError::Abort)?;
+
 					tx.remove(&self.store, k)?;
 					tx.insert(&self.merkle_todo, k, vec![])?;
 
-					let old_entry = self.decode_entry(v).map_err(db::TxError::Abort)?;
 					self.instance.updated(&mut tx, Some(&old_entry), None)?;
 					Ok(true)
 				}
@@ -285,10 +300,11 @@ where
 			.db()
 			.transaction(|mut tx| match tx.get(&self.store, k)? {
 				Some(cur_v) if blake2sum(&cur_v[..]) == vhash => {
+					let old_entry = self.decode_entry(&cur_v[..]).map_err(db::TxError::Abort)?;
+
 					tx.remove(&self.store, k)?;
 					tx.insert(&self.merkle_todo, k, vec![])?;
 
-					let old_entry = self.decode_entry(&cur_v[..]).map_err(db::TxError::Abort)?;
 					self.instance.updated(&mut tx, Some(&old_entry), None)?;
 					Ok(true)
 				}
@@ -302,27 +318,54 @@ where
 		Ok(removed)
 	}
 
+	// ---- Insert queue functions ----
+
+	pub(crate) fn queue_insert(
+		&self,
+		tx: &mut db::Transaction,
+		ins: &F::E,
+	) -> db::TxResult<(), Error> {
+		let tree_key = self.tree_key(ins.partition_key(), ins.sort_key());
+
+		let new_entry = match tx.get(&self.insert_queue, &tree_key)? {
+			Some(old_v) => {
+				let mut entry = self.decode_entry(&old_v).map_err(db::TxError::Abort)?;
+				entry.merge(ins);
+				entry.encode()
+			}
+			None => ins.encode(),
+		};
+		let new_entry = new_entry
+			.map_err(Error::RmpEncode)
+			.map_err(db::TxError::Abort)?;
+		tx.insert(&self.insert_queue, &tree_key, new_entry)?;
+
+		self.insert_queue_notify.notify_one();
+
+		Ok(())
+	}
+
 	// ---- Utility functions ----
 
 	pub fn tree_key(&self, p: &F::P, s: &F::S) -> Vec<u8> {
 		let mut ret = p.hash().to_vec();
-		ret.extend(s.sort_key());
+		ret.extend(s.sort_key().borrow());
 		ret
 	}
 
 	pub fn decode_entry(&self, bytes: &[u8]) -> Result<F::E, Error> {
-		match rmp_serde::decode::from_read_ref::<_, F::E>(bytes) {
+		match F::E::decode(bytes) {
-			Ok(x) => Ok(x),
+			Some(x) => Ok(x),
-			Err(e) => match F::try_migrate(bytes) {
+			None => {
-				Some(x) => Ok(x),
+				error!("Unable to decode entry of {}", F::TABLE_NAME);
-				None => {
+				for line in hexdump::hexdump_iter(bytes) {
-					warn!("Unable to decode entry of {}: {}", F::TABLE_NAME, e);
+					debug!("{}", line);
-					for line in hexdump::hexdump_iter(bytes) {
-						debug!("{}", line);
-					}
-					Err(e.into())
 				}
-			},
+				Err(Error::Message(format!(
+					"Unable to decode entry of {}",
+					F::TABLE_NAME
+				)))
+			}
 		}
 	}
 

src/table/gc.rs

@@ -31,7 +31,7 @@ const TABLE_GC_BATCH_SIZE: usize = 1024;
 // and the moment the garbage collection actually happens)
 const TABLE_GC_DELAY: Duration = Duration::from_secs(24 * 3600);
 
-pub(crate) struct TableGc<F: TableSchema + 'static, R: TableReplication + 'static> {
+pub(crate) struct TableGc<F: TableSchema, R: TableReplication> {
 	system: Arc<System>,
 	data: Arc<TableData<F, R>>,
 
@@ -49,29 +49,26 @@ impl Rpc for GcRpc {
 	type Response = Result<GcRpc, Error>;
 }
 
-impl<F, R> TableGc<F, R>
+impl<F: TableSchema, R: TableReplication> TableGc<F, R> {
-where
+	pub(crate) fn new(system: Arc<System>, data: Arc<TableData<F, R>>) -> Arc<Self> {
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
-	pub(crate) fn launch(system: Arc<System>, data: Arc<TableData<F, R>>) -> Arc<Self> {
 		let endpoint = system
 			.netapp
 			.endpoint(format!("garage_table/gc.rs/Rpc:{}", F::TABLE_NAME));
 
 		let gc = Arc::new(Self {
-			system: system.clone(),
+			system,
 			data,
 			endpoint,
 		});
-
 		gc.endpoint.set_handler(gc.clone());
 
-		system.background.spawn_worker(GcWorker::new(gc.clone()));
-
 		gc
 	}
 
+	pub(crate) fn spawn_workers(self: &Arc<Self>, bg: &BackgroundRunner) {
+		bg.spawn_worker(GcWorker::new(self.clone()));
+	}
+
 	async fn gc_loop_iter(&self) -> Result<Option<Duration>, Error> {
 		let now = now_msec();
 
@@ -276,11 +273,7 @@ where
 }
 
 #[async_trait]
-impl<F, R> EndpointHandler<GcRpc> for TableGc<F, R>
+impl<F: TableSchema, R: TableReplication> EndpointHandler<GcRpc> for TableGc<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	async fn handle(self: &Arc<Self>, message: &GcRpc, _from: NodeID) -> Result<GcRpc, Error> {
 		match message {
 			GcRpc::Update(items) => {
@@ -298,20 +291,12 @@ where
 	}
 }
 
-struct GcWorker<F, R>
+struct GcWorker<F: TableSchema, R: TableReplication> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	gc: Arc<TableGc<F, R>>,
 	wait_delay: Duration,
 }
 
-impl<F, R> GcWorker<F, R>
+impl<F: TableSchema, R: TableReplication> GcWorker<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	fn new(gc: Arc<TableGc<F, R>>) -> Self {
 		Self {
 			gc,
@@ -321,11 +306,7 @@ where
 }
 
 #[async_trait]
-impl<F, R> Worker for GcWorker<F, R>
+impl<F: TableSchema, R: TableReplication> Worker for GcWorker<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	fn name(&self) -> String {
 		format!("{} GC", F::TABLE_NAME)
 	}
@@ -347,10 +328,7 @@ where
 		}
 	}
 
-	async fn wait_for_work(&mut self, must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
-		if *must_exit.borrow() {
-			return WorkerState::Done;
-		}
 		tokio::time::sleep(self.wait_delay).await;
 		WorkerState::Busy
 	}

src/table/lib.rs

@@ -4,16 +4,18 @@
 #[macro_use]
 extern crate tracing;
 
-mod metrics;
 pub mod schema;
 pub mod util;
 
 pub mod data;
+pub mod replication;
+pub mod table;
+
 mod gc;
 mod merkle;
-pub mod replication;
+mod metrics;
+mod queue;
 mod sync;
-pub mod table;
 
 pub use schema::*;
 pub use table::*;

src/table/merkle.rs

@@ -3,12 +3,14 @@ use std::time::Duration;
 
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
+use tokio::select;
 use tokio::sync::watch;
 
 use garage_db as db;
 
 use garage_util::background::*;
 use garage_util::data::*;
+use garage_util::encode::{nonversioned_decode, nonversioned_encode};
 use garage_util::error::Error;
 
 use garage_rpc::ring::*;
@@ -64,22 +66,18 @@ pub enum MerkleNode {
 	Leaf(Vec<u8>, Hash),
 }
 
-impl<F, R> MerkleUpdater<F, R>
+impl<F: TableSchema, R: TableReplication> MerkleUpdater<F, R> {
-where
+	pub(crate) fn new(data: Arc<TableData<F, R>>) -> Arc<Self> {
-	F: TableSchema + 'static,
+		let empty_node_hash = blake2sum(&nonversioned_encode(&MerkleNode::Empty).unwrap()[..]);
-	R: TableReplication + 'static,
-{
-	pub(crate) fn launch(background: &BackgroundRunner, data: Arc<TableData<F, R>>) -> Arc<Self> {
-		let empty_node_hash = blake2sum(&rmp_to_vec_all_named(&MerkleNode::Empty).unwrap()[..]);
 
-		let ret = Arc::new(Self {
+		Arc::new(Self {
 			data,
 			empty_node_hash,
-		});
+		})
+	}
 
-		background.spawn_worker(MerkleWorker(ret.clone()));
+	pub(crate) fn spawn_workers(self: &Arc<Self>, background: &BackgroundRunner) {
-
+		background.spawn_worker(MerkleWorker(self.clone()));
-		ret
 	}
 
 	fn updater_loop_iter(&self) -> Result<WorkerState, Error> {
@@ -276,7 +274,7 @@ where
 			tx.remove(&self.data.merkle_tree, k.encode())?;
 			Ok(self.empty_node_hash)
 		} else {
-			let vby = rmp_to_vec_all_named(v).map_err(|e| db::TxError::Abort(e.into()))?;
+			let vby = nonversioned_encode(v).map_err(|e| db::TxError::Abort(e.into()))?;
 			let rethash = blake2sum(&vby[..]);
 			tx.insert(&self.data.merkle_tree, k.encode(), vby)?;
 			Ok(rethash)
@@ -302,17 +300,10 @@ where
 	}
 }
 
-struct MerkleWorker<F, R>(Arc<MerkleUpdater<F, R>>)
+struct MerkleWorker<F: TableSchema, R: TableReplication>(Arc<MerkleUpdater<F, R>>);
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static;
 
 #[async_trait]
-impl<F, R> Worker for MerkleWorker<F, R>
+impl<F: TableSchema, R: TableReplication> Worker for MerkleWorker<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	fn name(&self) -> String {
 		format!("{} Merkle", F::TABLE_NAME)
 	}
@@ -339,11 +330,11 @@ where
 		.unwrap()
 	}
 
-	async fn wait_for_work(&mut self, must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
-		if *must_exit.borrow() {
+		select! {
-			return WorkerState::Done;
+			_ = tokio::time::sleep(Duration::from_secs(60)) => (),
+			_ = self.0.data.merkle_todo_notify.notified() => (),
 		}
-		tokio::time::sleep(Duration::from_secs(10)).await;
 		WorkerState::Busy
 	}
 }
@@ -374,7 +365,7 @@ impl MerkleNode {
 	fn decode_opt(ent: &Option<db::Value>) -> Result<Self, Error> {
 		match ent {
 			None => Ok(MerkleNode::Empty),
-			Some(v) => Ok(rmp_serde::decode::from_read_ref::<_, MerkleNode>(&v[..])?),
+			Some(v) => Ok(nonversioned_decode::<MerkleNode>(&v[..])?),
 		}
 	}
 

src/table/queue.rs

@@ -0,0 +1,77 @@
+use std::sync::Arc;
+use std::time::Duration;
+
+use async_trait::async_trait;
+use tokio::select;
+use tokio::sync::watch;
+
+use garage_util::background::*;
+use garage_util::error::Error;
+
+use crate::replication::*;
+use crate::schema::*;
+use crate::table::*;
+
+const BATCH_SIZE: usize = 100;
+
+pub(crate) struct InsertQueueWorker<F, R>(pub(crate) Arc<Table<F, R>>)
+where
+	F: TableSchema,
+	R: TableReplication;
+
+#[async_trait]
+impl<F: TableSchema, R: TableReplication> Worker for InsertQueueWorker<F, R> {
+	fn name(&self) -> String {
+		format!("{} queue", F::TABLE_NAME)
+	}
+
+	fn status(&self) -> WorkerStatus {
+		WorkerStatus {
+			queue_length: Some(self.0.data.insert_queue.len().unwrap_or(0) as u64),
+			..Default::default()
+		}
+	}
+
+	async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
+		let mut kv_pairs = vec![];
+		let mut values = vec![];
+
+		for entry_kv in self.0.data.insert_queue.iter()? {
+			let (k, v) = entry_kv?;
+
+			values.push(self.0.data.decode_entry(&v)?);
+			kv_pairs.push((k, v));
+
+			if kv_pairs.len() > BATCH_SIZE {
+				break;
+			}
+		}
+
+		if kv_pairs.is_empty() {
+			return Ok(WorkerState::Idle);
+		}
+
+		self.0.insert_many(values).await?;
+
+		self.0.data.insert_queue.db().transaction(|mut tx| {
+			for (k, v) in kv_pairs.iter() {
+				if let Some(v2) = tx.get(&self.0.data.insert_queue, k)? {
+					if &v2 == v {
+						tx.remove(&self.0.data.insert_queue, k)?;
+					}
+				}
+			}
+			Ok(())
+		})?;
+
+		Ok(WorkerState::Busy)
+	}
+
+	async fn wait_for_work(&mut self) -> WorkerState {
+		select! {
+			_ = tokio::time::sleep(Duration::from_secs(600)) => (),
+			_ = self.0.data.insert_queue_notify.notified() => (),
+		}
+		WorkerState::Busy
+	}
+}

src/table/replication/parameters.rs

@@ -2,7 +2,7 @@ use garage_rpc::ring::*;
 use garage_util::data::*;
 
 /// Trait to describe how a table shall be replicated
-pub trait TableReplication: Send + Sync {
+pub trait TableReplication: Send + Sync + 'static {
 	// See examples in table_sharded.rs and table_fullcopy.rs
 	// To understand various replication methods
 

src/table/schema.rs

@@ -2,11 +2,14 @@ use serde::{Deserialize, Serialize};
 
 use garage_db as db;
 use garage_util::data::*;
+use garage_util::migrate::Migrate;
 
 use crate::crdt::Crdt;
 
 /// Trait for field used to partition data
-pub trait PartitionKey {
+pub trait PartitionKey:
+	Clone + PartialEq + Serialize + for<'de> Deserialize<'de> + Send + Sync + 'static
+{
 	/// Get the key used to partition
 	fn hash(&self) -> Hash;
 }
@@ -27,18 +30,24 @@ impl PartitionKey for FixedBytes32 {
 }
 
 /// Trait for field used to sort data
-pub trait SortKey {
+pub trait SortKey: Clone + Serialize + for<'de> Deserialize<'de> + Send + Sync + 'static {
+	type B<'a>: std::borrow::Borrow<[u8]>;
+
 	/// Get the key used to sort
-	fn sort_key(&self) -> &[u8];
+	fn sort_key(&self) -> Self::B<'_>;
 }
 
 impl SortKey for String {
+	type B<'a> = &'a [u8];
+
 	fn sort_key(&self) -> &[u8] {
 		self.as_bytes()
 	}
 }
 
 impl SortKey for FixedBytes32 {
+	type B<'a> = &'a [u8];
+
 	fn sort_key(&self) -> &[u8] {
 		self.as_slice()
 	}
@@ -46,7 +55,7 @@ impl SortKey for FixedBytes32 {
 
 /// Trait for an entry in a table. It must be sortable and partitionnable.
 pub trait Entry<P: PartitionKey, S: SortKey>:
-	Crdt + PartialEq + Clone + Serialize + for<'de> Deserialize<'de> + Send + Sync
+	Crdt + PartialEq + Clone + Migrate + Send + Sync + 'static
 {
 	/// Get the key used to partition
 	fn partition_key(&self) -> &P;
@@ -65,23 +74,16 @@ pub trait TableSchema: Send + Sync + 'static {
 	const TABLE_NAME: &'static str;
 
 	/// The partition key used in that table
-	type P: PartitionKey + Clone + PartialEq + Serialize + for<'de> Deserialize<'de> + Send + Sync;
+	type P: PartitionKey;
 	/// The sort key used int that table
-	type S: SortKey + Clone + Serialize + for<'de> Deserialize<'de> + Send + Sync;
+	type S: SortKey;
 
 	/// They type for an entry in that table
 	type E: Entry<Self::P, Self::S>;
 
 	/// The type for a filter that can be applied to select entries
 	/// (e.g. filter out deleted entries)
-	type Filter: Clone + Serialize + for<'de> Deserialize<'de> + Send + Sync;
+	type Filter: Clone + Serialize + for<'de> Deserialize<'de> + Send + Sync + 'static;
-
-	// Action to take if not able to decode current version:
-	// try loading from an older version
-	/// Try migrating an entry from an older version
-	fn try_migrate(_bytes: &[u8]) -> Option<Self::E> {
-		None
-	}
 
 	/// Actions triggered by data changing in a table. If such actions
 	/// include updates to the local database that should be applied

src/table/sync.rs

@@ -2,6 +2,7 @@ use std::collections::VecDeque;
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
+use arc_swap::ArcSwapOption;
 use async_trait::async_trait;
 use futures_util::stream::*;
 use opentelemetry::KeyValue;
@@ -13,7 +14,8 @@ use tokio::sync::{mpsc, watch};
 
 use garage_util::background::*;
 use garage_util::data::*;
-use garage_util::error::Error;
+use garage_util::encode::{debug_serialize, nonversioned_encode};
+use garage_util::error::{Error, OkOrMessage};
 
 use garage_rpc::ring::*;
 use garage_rpc::system::System;
@@ -27,12 +29,12 @@ use crate::*;
 // Do anti-entropy every 10 minutes
 const ANTI_ENTROPY_INTERVAL: Duration = Duration::from_secs(10 * 60);
 
-pub struct TableSyncer<F: TableSchema + 'static, R: TableReplication + 'static> {
+pub struct TableSyncer<F: TableSchema, R: TableReplication> {
 	system: Arc<System>,
 	data: Arc<TableData<F, R>>,
 	merkle: Arc<MerkleUpdater<F, R>>,
 
-	add_full_sync_tx: mpsc::UnboundedSender<()>,
+	add_full_sync_tx: ArcSwapOption<mpsc::UnboundedSender<()>>,
 	endpoint: Arc<Endpoint<SyncRpc, Self>>,
 }
 
@@ -60,12 +62,8 @@ struct TodoPartition {
 	retain: bool,
 }
 
-impl<F, R> TableSyncer<F, R>
+impl<F: TableSchema, R: TableReplication> TableSyncer<F, R> {
-where
+	pub(crate) fn new(
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
-	pub(crate) fn launch(
 		system: Arc<System>,
 		data: Arc<TableData<F, R>>,
 		merkle: Arc<MerkleUpdater<F, R>>,
@@ -74,34 +72,40 @@ where
 			.netapp
 			.endpoint(format!("garage_table/sync.rs/Rpc:{}", F::TABLE_NAME));
 
-		let (add_full_sync_tx, add_full_sync_rx) = mpsc::unbounded_channel();
-
 		let syncer = Arc::new(Self {
-			system: system.clone(),
+			system,
 			data,
 			merkle,
-			add_full_sync_tx,
+			add_full_sync_tx: ArcSwapOption::new(None),
 			endpoint,
 		});
-
 		syncer.endpoint.set_handler(syncer.clone());
 
-		system.background.spawn_worker(SyncWorker {
-			syncer: syncer.clone(),
-			ring_recv: system.ring.clone(),
-			ring: system.ring.borrow().clone(),
-			add_full_sync_rx,
-			todo: vec![],
-			next_full_sync: Instant::now() + Duration::from_secs(20),
-		});
-
 		syncer
 	}
 
-	pub fn add_full_sync(&self) {
+	pub(crate) fn spawn_workers(self: &Arc<Self>, bg: &BackgroundRunner) {
-		if self.add_full_sync_tx.send(()).is_err() {
+		let (add_full_sync_tx, add_full_sync_rx) = mpsc::unbounded_channel();
-			error!("({}) Could not add full sync", F::TABLE_NAME);
+		self.add_full_sync_tx
-		}
+			.store(Some(Arc::new(add_full_sync_tx)));
+
+		bg.spawn_worker(SyncWorker {
+			syncer: self.clone(),
+			ring_recv: self.system.ring.clone(),
+			ring: self.system.ring.borrow().clone(),
+			add_full_sync_rx,
+			todo: vec![],
+			next_full_sync: Instant::now() + Duration::from_secs(20),
+		});
+	}
+
+	pub fn add_full_sync(&self) -> Result<(), Error> {
+		let tx = self.add_full_sync_tx.load();
+		let tx = tx
+			.as_ref()
+			.ok_or_message("table sync worker is not running")?;
+		tx.send(()).ok_or_message("send error")?;
+		Ok(())
 	}
 
 	// ----
@@ -295,7 +299,7 @@ where
 			);
 			return Ok(());
 		}
-		let root_ck_hash = hash_of::<MerkleNode>(&root_ck)?;
+		let root_ck_hash = hash_of_merkle_node(&root_ck)?;
 
 		// Check if they have the same root checksum
 		// If so, do nothing.
@@ -452,16 +456,12 @@ where
 // ======= SYNCHRONIZATION PROCEDURE -- RECEIVER SIDE ======
 
 #[async_trait]
-impl<F, R> EndpointHandler<SyncRpc> for TableSyncer<F, R>
+impl<F: TableSchema, R: TableReplication> EndpointHandler<SyncRpc> for TableSyncer<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	async fn handle(self: &Arc<Self>, message: &SyncRpc, from: NodeID) -> Result<SyncRpc, Error> {
 		match message {
 			SyncRpc::RootCkHash(range, h) => {
 				let (_root_ck_key, root_ck) = self.get_root_ck(*range)?;
-				let hash = hash_of::<MerkleNode>(&root_ck)?;
+				let hash = hash_of_merkle_node(&root_ck)?;
 				Ok(SyncRpc::RootCkDifferent(hash != *h))
 			}
 			SyncRpc::GetNode(k) => {
@@ -490,7 +490,7 @@ where
 
 // -------- Sync Worker ---------
 
-struct SyncWorker<F: TableSchema + 'static, R: TableReplication + 'static> {
+struct SyncWorker<F: TableSchema, R: TableReplication> {
 	syncer: Arc<TableSyncer<F, R>>,
 	ring_recv: watch::Receiver<Arc<Ring>>,
 	ring: Arc<Ring>,
@@ -499,7 +499,7 @@ struct SyncWorker<F: TableSchema + 'static, R: TableReplication + 'static> {
 	next_full_sync: Instant,
 }
 
-impl<F: TableSchema + 'static, R: TableReplication + 'static> SyncWorker<F, R> {
+impl<F: TableSchema, R: TableReplication> SyncWorker<F, R> {
 	fn add_full_sync(&mut self) {
 		let system = &self.syncer.system;
 		let data = &self.syncer.data;
@@ -565,7 +565,7 @@ impl<F: TableSchema + 'static, R: TableReplication + 'static> SyncWorker<F, R> {
 }
 
 #[async_trait]
-impl<F: TableSchema + 'static, R: TableReplication + 'static> Worker for SyncWorker<F, R> {
+impl<F: TableSchema, R: TableReplication> Worker for SyncWorker<F, R> {
 	fn name(&self) -> String {
 		format!("{} sync", F::TABLE_NAME)
 	}
@@ -586,10 +586,7 @@ impl<F: TableSchema + 'static, R: TableReplication + 'static> Worker for SyncWor
 		}
 	}
 
-	async fn wait_for_work(&mut self, must_exit: &watch::Receiver<bool>) -> WorkerState {
+	async fn wait_for_work(&mut self) -> WorkerState {
-		if *must_exit.borrow() {
-			return WorkerState::Done;
-		}
 		select! {
 			s = self.add_full_sync_rx.recv() => {
 				if let Some(()) = s {
@@ -618,8 +615,8 @@ impl<F: TableSchema + 'static, R: TableReplication + 'static> Worker for SyncWor
 
 // ---- UTIL ----
 
-fn hash_of<T: Serialize>(x: &T) -> Result<Hash, Error> {
+fn hash_of_merkle_node(x: &MerkleNode) -> Result<Hash, Error> {
-	Ok(blake2sum(&rmp_to_vec_all_named(x)?[..]))
+	Ok(blake2sum(&nonversioned_encode(x)?[..]))
 }
 
 fn join_ordered<'a, K: Ord + Eq, V1, V2>(

src/table/table.rs

@@ -14,9 +14,11 @@ use opentelemetry::{
 
 use garage_db as db;
 
+use garage_util::background::BackgroundRunner;
 use garage_util::data::*;
 use garage_util::error::Error;
 use garage_util::metrics::RecordDuration;
+use garage_util::migrate::Migrate;
 
 use garage_rpc::system::System;
 use garage_rpc::*;
@@ -25,16 +27,18 @@ use crate::crdt::Crdt;
 use crate::data::*;
 use crate::gc::*;
 use crate::merkle::*;
+use crate::queue::InsertQueueWorker;
 use crate::replication::*;
 use crate::schema::*;
 use crate::sync::*;
 use crate::util::*;
 
-pub struct Table<F: TableSchema + 'static, R: TableReplication + 'static> {
+pub struct Table<F: TableSchema, R: TableReplication> {
 	pub system: Arc<System>,
 	pub data: Arc<TableData<F, R>>,
 	pub merkle_updater: Arc<MerkleUpdater<F, R>>,
 	pub syncer: Arc<TableSyncer<F, R>>,
+	gc: Arc<TableGc<F, R>>,
 	endpoint: Arc<Endpoint<TableRpc<F>, Self>>,
 }
 
@@ -61,11 +65,7 @@ impl<F: TableSchema> Rpc for TableRpc<F> {
 	type Response = Result<TableRpc<F>, Error>;
 }
 
-impl<F, R> Table<F, R>
+impl<F: TableSchema, R: TableReplication> Table<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	// =============== PUBLIC INTERFACE FUNCTIONS (new, insert, get, etc) ===============
 
 	pub fn new(instance: F, replication: R, system: Arc<System>, db: &db::Db) -> Arc<Self> {
@@ -75,15 +75,16 @@ where
 
 		let data = TableData::new(system.clone(), instance, replication, db);
 
-		let merkle_updater = MerkleUpdater::launch(&system.background, data.clone());
+		let merkle_updater = MerkleUpdater::new(data.clone());
 
-		let syncer = TableSyncer::launch(system.clone(), data.clone(), merkle_updater.clone());
+		let syncer = TableSyncer::new(system.clone(), data.clone(), merkle_updater.clone());
-		TableGc::launch(system.clone(), data.clone());
+		let gc = TableGc::new(system.clone(), data.clone());
 
 		let table = Arc::new(Self {
 			system,
 			data,
 			merkle_updater,
+			gc,
 			syncer,
 			endpoint,
 		});
@@ -93,6 +94,13 @@ where
 		table
 	}
 
+	pub fn spawn_workers(self: &Arc<Self>, bg: &BackgroundRunner) {
+		self.merkle_updater.spawn_workers(bg);
+		self.syncer.spawn_workers(bg);
+		self.gc.spawn_workers(bg);
+		bg.spawn_worker(InsertQueueWorker(self.clone()));
+	}
+
 	pub async fn insert(&self, e: &F::E) -> Result<(), Error> {
 		let tracer = opentelemetry::global::tracer("garage_table");
 		let span = tracer.start(format!("{} insert", F::TABLE_NAME));
@@ -111,7 +119,7 @@ where
 		let hash = e.partition_key().hash();
 		let who = self.data.replication.write_nodes(&hash);
 
-		let e_enc = Arc::new(ByteBuf::from(rmp_to_vec_all_named(e)?));
+		let e_enc = Arc::new(ByteBuf::from(e.encode()?));
 		let rpc = TableRpc::<F>::Update(vec![e_enc]);
 
 		self.system
@@ -128,6 +136,11 @@ where
 		Ok(())
 	}
 
+	/// Insert item locally
+	pub fn queue_insert(&self, tx: &mut db::Transaction, e: &F::E) -> db::TxResult<(), Error> {
+		self.data.queue_insert(tx, e)
+	}
+
 	pub async fn insert_many<I, IE>(&self, entries: I) -> Result<(), Error>
 	where
 		I: IntoIterator<Item = IE> + Send + Sync,
@@ -157,7 +170,7 @@ where
 			let entry = entry.borrow();
 			let hash = entry.partition_key().hash();
 			let who = self.data.replication.write_nodes(&hash);
-			let e_enc = Arc::new(ByteBuf::from(rmp_to_vec_all_named(entry)?));
+			let e_enc = Arc::new(ByteBuf::from(entry.encode()?));
 			for node in who {
 				call_list.entry(node).or_default().push(e_enc.clone());
 			}
@@ -259,9 +272,11 @@ where
 			if not_all_same {
 				let self2 = self.clone();
 				let ent2 = ret_entry.clone();
-				self.system
+				tokio::spawn(async move {
-					.background
+					if let Err(e) = self2.repair_on_read(&who[..], ent2).await {
-					.spawn_cancellable(async move { self2.repair_on_read(&who[..], ent2).await });
+						warn!("Error doing repair on read: {}", e);
+					}
+				});
 			}
 		}
 
@@ -358,11 +373,12 @@ where
 				.into_iter()
 				.map(|k| ret.get(&k).unwrap().clone())
 				.collect::<Vec<_>>();
-			self.system.background.spawn_cancellable(async move {
+			tokio::spawn(async move {
 				for v in to_repair {
-					self2.repair_on_read(&who[..], v).await?;
+					if let Err(e) = self2.repair_on_read(&who[..], v).await {
+						warn!("Error doing repair on read: {}", e);
+					}
 				}
-				Ok(())
 			});
 		}
 
@@ -393,7 +409,7 @@ where
 	// =============== UTILITY FUNCTION FOR CLIENT OPERATIONS ===============
 
 	async fn repair_on_read(&self, who: &[Uuid], what: F::E) -> Result<(), Error> {
-		let what_enc = Arc::new(ByteBuf::from(rmp_to_vec_all_named(&what)?));
+		let what_enc = Arc::new(ByteBuf::from(what.encode()?));
 		self.system
 			.rpc
 			.try_call_many(
@@ -408,11 +424,7 @@ where
 }
 
 #[async_trait]
-impl<F, R> EndpointHandler<TableRpc<F>> for Table<F, R>
+impl<F: TableSchema, R: TableReplication> EndpointHandler<TableRpc<F>> for Table<F, R> {
-where
-	F: TableSchema + 'static,
-	R: TableReplication + 'static,
-{
 	async fn handle(
 		self: &Arc<Self>,
 		msg: &TableRpc<F>,

src/table/util.rs

@@ -7,6 +7,8 @@ use crate::schema::*;
 #[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct EmptyKey;
 impl SortKey for EmptyKey {
+	type B<'a> = &'a [u8];
+
 	fn sort_key(&self) -> &[u8] {
 		&[]
 	}

src/util/Cargo.toml

@@ -23,6 +23,7 @@ bytes = "1.0"
 digest = "0.10"
 err-derive = "0.3"
 git-version = "0.3.4"
+hexdump = "0.1"
 xxhash-rust = { version = "0.8", default-features = false, features = ["xxh3"] }
 hex = "0.4"
 lazy_static = "1.4"

src/util/background/job_worker.rs

@@ -1,48 +0,0 @@
-//! Job worker: a generic worker that just processes incoming
-//! jobs one by one
-
-use std::sync::Arc;
-
-use async_trait::async_trait;
-use tokio::sync::{mpsc, Mutex};
-
-use crate::background::worker::*;
-use crate::background::*;
-
-pub(crate) struct JobWorker {
-	pub(crate) index: usize,
-	pub(crate) job_chan: Arc<Mutex<mpsc::UnboundedReceiver<(Job, bool)>>>,
-	pub(crate) next_job: Option<Job>,
-}
-
-#[async_trait]
-impl Worker for JobWorker {
-	fn name(&self) -> String {
-		format!("Job worker #{}", self.index)
-	}
-
-	async fn work(&mut self, _must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error> {
-		match self.next_job.take() {
-			None => return Ok(WorkerState::Idle),
-			Some(job) => {
-				job.await?;
-				Ok(WorkerState::Busy)
-			}
-		}
-	}
-
-	async fn wait_for_work(&mut self, must_exit: &watch::Receiver<bool>) -> WorkerState {
-		loop {
-			match self.job_chan.lock().await.recv().await {
-				Some((job, cancellable)) => {
-					if cancellable && *must_exit.borrow() {
-						continue;
-					}
-					self.next_job = Some(job);
-					return WorkerState::Busy;
-				}
-				None => return WorkerState::Done,
-			}
-		}
-	}
-}

src/util/background/mod.rs

@@ -1,27 +1,19 @@
 //! Job runner for futures and async functions
 
-pub mod job_worker;
+pub mod vars;
 pub mod worker;
 
-use core::future::Future;
-
 use std::collections::HashMap;
-use std::pin::Pin;
 use std::sync::Arc;
 
 use serde::{Deserialize, Serialize};
-use tokio::sync::{mpsc, watch, Mutex};
+use tokio::sync::{mpsc, watch};
 
-use crate::error::Error;
 use worker::WorkerProcessor;
 pub use worker::{Worker, WorkerState};
 
-pub(crate) type JobOutput = Result<(), Error>;
-pub(crate) type Job = Pin<Box<dyn Future<Output = JobOutput> + Send>>;
-
 /// Job runner for futures and async functions
 pub struct BackgroundRunner {
-	send_job: mpsc::UnboundedSender<(Job, bool)>,
 	send_worker: mpsc::UnboundedSender<Box<dyn Worker>>,
 	worker_info: Arc<std::sync::Mutex<HashMap<usize, WorkerInfo>>>,
 }
@@ -49,10 +41,7 @@ pub struct WorkerStatus {
 
 impl BackgroundRunner {
 	/// Create a new BackgroundRunner
-	pub fn new(
+	pub fn new(stop_signal: watch::Receiver<bool>) -> (Arc<Self>, tokio::task::JoinHandle<()>) {
-		n_runners: usize,
-		stop_signal: watch::Receiver<bool>,
-	) -> (Arc<Self>, tokio::task::JoinHandle<()>) {
 		let (send_worker, worker_out) = mpsc::unbounded_channel::<Box<dyn Worker>>();
 
 		let worker_info = Arc::new(std::sync::Mutex::new(HashMap::new()));
@@ -63,24 +52,7 @@ impl BackgroundRunner {
 			worker_processor.run().await;
 		});
 
-		let (send_job, queue_out) = mpsc::unbounded_channel();
-		let queue_out = Arc::new(Mutex::new(queue_out));
-
-		for i in 0..n_runners {
-			let queue_out = queue_out.clone();
-
-			send_worker
-				.send(Box::new(job_worker::JobWorker {
-					index: i,
-					job_chan: queue_out.clone(),
-					next_job: None,
-				}))
-				.ok()
-				.unwrap();
-		}
-
 		let bgrunner = Arc::new(Self {
-			send_job,
 			send_worker,
 			worker_info,
 		});
@@ -91,31 +63,6 @@ impl BackgroundRunner {
 		self.worker_info.lock().unwrap().clone()
 	}
 
-	/// Spawn a task to be run in background
-	pub fn spawn<T>(&self, job: T)
-	where
-		T: Future<Output = JobOutput> + Send + 'static,
-	{
-		let boxed: Job = Box::pin(job);
-		self.send_job
-			.send((boxed, false))
-			.ok()
-			.expect("Could not put job in queue");
-	}
-
-	/// Spawn a task to be run in background. It may get discarded before running if spawned while
-	/// the runner is stopping
-	pub fn spawn_cancellable<T>(&self, job: T)
-	where
-		T: Future<Output = JobOutput> + Send + 'static,
-	{
-		let boxed: Job = Box::pin(job);
-		self.send_job
-			.send((boxed, true))
-			.ok()
-			.expect("Could not put job in queue");
-	}
-
 	pub fn spawn_worker<W>(&self, worker: W)
 	where
 		W: Worker + 'static,

src/util/background/vars.rs

@@ -0,0 +1,113 @@
+use std::collections::HashMap;
+use std::str::FromStr;
+
+use crate::error::{Error, OkOrMessage};
+use crate::migrate::Migrate;
+use crate::persister::PersisterShared;
+
+pub struct BgVars {
+	vars: HashMap<&'static str, Box<dyn BgVarTrait>>,
+}
+
+impl BgVars {
+	pub fn new() -> Self {
+		Self {
+			vars: HashMap::new(),
+		}
+	}
+
+	pub fn register_rw<V, T, GF, SF>(
+		&mut self,
+		p: &PersisterShared<V>,
+		name: &'static str,
+		get_fn: GF,
+		set_fn: SF,
+	) where
+		V: Migrate + Default + Send + Sync,
+		T: FromStr + ToString + Send + Sync + 'static,
+		GF: Fn(&PersisterShared<V>) -> T + Send + Sync + 'static,
+		SF: Fn(&PersisterShared<V>, T) -> Result<(), Error> + Send + Sync + 'static,
+	{
+		let p1 = p.clone();
+		let get_fn = move || get_fn(&p1);
+
+		let p2 = p.clone();
+		let set_fn = move |v| set_fn(&p2, v);
+
+		self.vars.insert(name, Box::new(BgVar { get_fn, set_fn }));
+	}
+
+	pub fn register_ro<V, T, GF>(&mut self, p: &PersisterShared<V>, name: &'static str, get_fn: GF)
+	where
+		V: Migrate + Default + Send + Sync,
+		T: FromStr + ToString + Send + Sync + 'static,
+		GF: Fn(&PersisterShared<V>) -> T + Send + Sync + 'static,
+	{
+		let p1 = p.clone();
+		let get_fn = move || get_fn(&p1);
+
+		let set_fn = move |_| Err(Error::Message(format!("Cannot set value of {}", name)));
+
+		self.vars.insert(name, Box::new(BgVar { get_fn, set_fn }));
+	}
+
+	pub fn get(&self, var: &str) -> Result<String, Error> {
+		Ok(self
+			.vars
+			.get(var)
+			.ok_or_message("variable does not exist")?
+			.get())
+	}
+
+	pub fn get_all(&self) -> Vec<(&'static str, String)> {
+		self.vars.iter().map(|(k, v)| (*k, v.get())).collect()
+	}
+
+	pub fn set(&self, var: &str, val: &str) -> Result<(), Error> {
+		self.vars
+			.get(var)
+			.ok_or_message("variable does not exist")?
+			.set(val)
+	}
+}
+
+impl Default for BgVars {
+	fn default() -> Self {
+		Self::new()
+	}
+}
+
+// ----
+
+trait BgVarTrait: Send + Sync + 'static {
+	fn get(&self) -> String;
+	fn set(&self, v: &str) -> Result<(), Error>;
+}
+
+struct BgVar<T, GF, SF>
+where
+	T: FromStr + ToString + Send + Sync + 'static,
+	GF: Fn() -> T + Send + Sync + 'static,
+	SF: Fn(T) -> Result<(), Error> + Sync + Send + 'static,
+{
+	get_fn: GF,
+	set_fn: SF,
+}
+
+impl<T, GF, SF> BgVarTrait for BgVar<T, GF, SF>
+where
+	T: FromStr + ToString + Sync + Send + 'static,
+	GF: Fn() -> T + Sync + Send + 'static,
+	SF: Fn(T) -> Result<(), Error> + Sync + Send + 'static,
+{
+	fn get(&self) -> String {
+		(self.get_fn)().to_string()
+	}
+
+	fn set(&self, vstr: &str) -> Result<(), Error> {
+		let value = vstr
+			.parse()
+			.map_err(|_| Error::Message(format!("invalid value: {}", vstr)))?;
+		(self.set_fn)(value)
+	}
+}

src/util/background/worker.rs

@@ -1,6 +1,6 @@
 use std::collections::HashMap;
 use std::sync::Arc;
-use std::time::{Duration, Instant};
+use std::time::Duration;
 
 use async_trait::async_trait;
 use futures::future::*;
@@ -14,6 +14,10 @@ use crate::background::{WorkerInfo, WorkerStatus};
 use crate::error::Error;
 use crate::time::now_msec;
 
+// All workers that haven't exited for this time after an exit signal was recieved
+// will be interrupted in the middle of whatever they are doing.
+const EXIT_DEADLINE: Duration = Duration::from_secs(8);
+
 #[derive(PartialEq, Copy, Clone, Serialize, Deserialize, Debug)]
 pub enum WorkerState {
 	Busy,
@@ -50,10 +54,8 @@ pub trait Worker: Send {
 	async fn work(&mut self, must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error>;
 
 	/// Wait for work: await for some task to become available.  This future can be interrupted in
-	/// the middle for any reason.  This future doesn't have to await on must_exit.changed(), we
+	/// the middle for any reason, for example if an interrupt signal was recieved.
-	/// are doing it for you.  Therefore it only receives a read refernce to must_exit which allows
+	async fn wait_for_work(&mut self) -> WorkerState;
-	/// it to check if we are exiting.
-	async fn wait_for_work(&mut self, must_exit: &watch::Receiver<bool>) -> WorkerState;
 }
 
 pub(crate) struct WorkerProcessor {
@@ -93,11 +95,9 @@ impl WorkerProcessor {
 						let task_id = next_task_id;
 						next_task_id += 1;
 						let stop_signal = self.stop_signal.clone();
-						let stop_signal_worker = self.stop_signal.clone();
 						let mut worker = WorkerHandler {
 								task_id,
 								stop_signal,
-								stop_signal_worker,
 								worker: new_worker,
 								state: WorkerState::Busy,
 								errors: 0,
@@ -153,26 +153,14 @@ impl WorkerProcessor {
 		}
 
 		// We are exiting, drain everything
-		let drain_half_time = Instant::now() + Duration::from_secs(5);
 		let drain_everything = async move {
-			while let Some(mut worker) = workers.next().await {
+			while let Some(worker) = workers.next().await {
-				if worker.state == WorkerState::Done {
+				info!(
-					info!(
+					"Worker {} (TID {}) exited (last state: {:?})",
-						"Worker {} (TID {}) exited",
+					worker.worker.name(),
-						worker.worker.name(),
+					worker.task_id,
-						worker.task_id
+					worker.state
-					);
+				);
-				} else if Instant::now() > drain_half_time {
-					warn!("Worker {} (TID {}) interrupted between two iterations in state {:?} (this should be fine)", worker.worker.name(), worker.task_id, worker.state);
-				} else {
-					workers.push(
-						async move {
-							worker.step().await;
-							worker
-						}
-						.boxed(),
-					);
-				}
 			}
 		};
 
@@ -180,7 +168,7 @@ impl WorkerProcessor {
 			_ = drain_everything => {
 				info!("All workers exited peacefully \\o/");
 			}
-			_ = tokio::time::sleep(Duration::from_secs(9)) => {
+			_ = tokio::time::sleep(EXIT_DEADLINE) => {
 				error!("Some workers could not exit in time, we are cancelling some things in the middle");
 			}
 		}
@@ -190,7 +178,6 @@ impl WorkerProcessor {
 struct WorkerHandler {
 	task_id: usize,
 	stop_signal: watch::Receiver<bool>,
-	stop_signal_worker: watch::Receiver<bool>,
 	worker: Box<dyn Worker>,
 	state: WorkerState,
 	errors: usize,
@@ -225,33 +212,19 @@ impl WorkerHandler {
 			},
 			WorkerState::Throttled(delay) => {
 				// Sleep for given delay and go back to busy state
-				if !*self.stop_signal.borrow() {
+				select! {
-					select! {
+					_ = tokio::time::sleep(Duration::from_secs_f32(delay)) => {
-						_ = tokio::time::sleep(Duration::from_secs_f32(delay)) => (),
+						self.state = WorkerState::Busy;
-						_ = self.stop_signal.changed() => (),
 					}
+					_ = self.stop_signal.changed() => (),
 				}
-				self.state = WorkerState::Busy;
 			}
 			WorkerState::Idle => {
-				if *self.stop_signal.borrow() {
+				select! {
-					select! {
+					new_st = self.worker.wait_for_work() => {
-						new_st = self.worker.wait_for_work(&self.stop_signal_worker) => {
+						self.state = new_st;
-							self.state = new_st;
-						}
-						_ = tokio::time::sleep(Duration::from_secs(1)) => {
-							// stay in Idle state
-						}
-					}
-				} else {
-					select! {
-						new_st = self.worker.wait_for_work(&self.stop_signal_worker) => {
-							self.state = new_st;
-						}
-						_ = self.stop_signal.changed() => {
-							// stay in Idle state
-						}
 					}
+					_ = self.stop_signal.changed() => (),
 				}
 			}
 			WorkerState::Done => unreachable!(),

src/util/data.rs

@@ -140,34 +140,3 @@ pub fn fasthash(data: &[u8]) -> FastHash {
 pub fn gen_uuid() -> Uuid {
 	rand::thread_rng().gen::<[u8; 32]>().into()
 }
-
-// RMP serialization with names of fields and variants
-
-/// Serialize to MessagePack
-pub fn rmp_to_vec_all_named<T>(val: &T) -> Result<Vec<u8>, rmp_serde::encode::Error>
-where
-	T: Serialize + ?Sized,
-{
-	let mut wr = Vec::with_capacity(128);
-	let mut se = rmp_serde::Serializer::new(&mut wr)
-		.with_struct_map()
-		.with_string_variants();
-	val.serialize(&mut se)?;
-	Ok(wr)
-}
-
-/// Serialize to JSON, truncating long result
-pub fn debug_serialize<T: Serialize>(x: T) -> String {
-	match serde_json::to_string(&x) {
-		Ok(ss) => {
-			if ss.len() > 100 {
-				// TODO this can panic if 100 is not a codepoint boundary, but inside a 2 Bytes
-				// (or more) codepoint
-				ss[..100].to_string()
-			} else {
-				ss
-			}
-		}
-		Err(e) => format!("<JSON serialization error: {}>", e),
-	}
-}

src/util/encode.rs

@@ -0,0 +1,42 @@
+use serde::{Deserialize, Serialize};
+
+/// Serialize to MessagePacki, without versionning
+/// (see garage_util::migrate for functions that manage versionned
+/// data formats)
+pub fn nonversioned_encode<T>(val: &T) -> Result<Vec<u8>, rmp_serde::encode::Error>
+where
+	T: Serialize + ?Sized,
+{
+	let mut wr = Vec::with_capacity(128);
+	let mut se = rmp_serde::Serializer::new(&mut wr)
+		.with_struct_map()
+		.with_string_variants();
+	val.serialize(&mut se)?;
+	Ok(wr)
+}
+
+/// Deserialize from MessagePacki, without versionning
+/// (see garage_util::migrate for functions that manage versionned
+/// data formats)
+pub fn nonversioned_decode<T>(bytes: &[u8]) -> Result<T, rmp_serde::decode::Error>
+where
+	T: for<'de> Deserialize<'de> + ?Sized,
+{
+	rmp_serde::decode::from_read_ref::<_, T>(bytes)
+}
+
+/// Serialize to JSON, truncating long result
+pub fn debug_serialize<T: Serialize>(x: T) -> String {
+	match serde_json::to_string(&x) {
+		Ok(ss) => {
+			if ss.len() > 100 {
+				// TODO this can panic if 100 is not a codepoint boundary, but inside a 2 Bytes
+				// (or more) codepoint
+				ss[..100].to_string()
+			} else {
+				ss
+			}
+		}
+		Err(e) => format!("<JSON serialization error: {}>", e),
+	}
+}

src/util/error.rs

@@ -7,6 +7,7 @@ use err_derive::Error;
 use serde::{de::Visitor, Deserialize, Deserializer, Serialize, Serializer};
 
 use crate::data::*;
+use crate::encode::debug_serialize;
 
 /// Regroup all Garage errors
 #[derive(Debug, Error)]

src/util/lib.rs

@@ -8,11 +8,12 @@ pub mod background;
 pub mod config;
 pub mod crdt;
 pub mod data;
+pub mod encode;
 pub mod error;
 pub mod formater;
 pub mod metrics;
+pub mod migrate;
 pub mod persister;
 pub mod time;
-pub mod token_bucket;
 pub mod tranquilizer;
 pub mod version;

src/util/migrate.rs

@@ -0,0 +1,159 @@
+use serde::{Deserialize, Serialize};
+
+/// Indicates that this type has an encoding that can be migrated from
+/// a previous version upon upgrades of Garage.
+pub trait Migrate: Serialize + for<'de> Deserialize<'de> + 'static {
+	/// A sequence of bytes to add at the beginning of the serialized
+	/// string, to identify that the data is of this version.
+	const VERSION_MARKER: &'static [u8] = b"";
+
+	/// The previous version of this data type, from which items of this version
+	/// can be migrated.
+	type Previous: Migrate;
+
+	/// The migration function that transforms a value decoded in the old format
+	/// to an up-to-date value.
+	fn migrate(previous: Self::Previous) -> Self;
+
+	/// Decode an encoded version of this type, going through a migration if necessary.
+	fn decode(bytes: &[u8]) -> Option<Self> {
+		let marker_len = Self::VERSION_MARKER.len();
+		if bytes.get(..marker_len) == Some(Self::VERSION_MARKER) {
+			if let Ok(value) = rmp_serde::decode::from_read_ref::<_, Self>(&bytes[marker_len..]) {
+				return Some(value);
+			}
+		}
+
+		Self::Previous::decode(bytes).map(Self::migrate)
+	}
+
+	/// Encode this type with optionnal version marker
+	fn encode(&self) -> Result<Vec<u8>, rmp_serde::encode::Error> {
+		let mut wr = Vec::with_capacity(128);
+		wr.extend_from_slice(Self::VERSION_MARKER);
+		let mut se = rmp_serde::Serializer::new(&mut wr)
+			.with_struct_map()
+			.with_string_variants();
+		self.serialize(&mut se)?;
+		Ok(wr)
+	}
+}
+
+/// Indicates that this type has no previous encoding version to be migrated from.
+pub trait InitialFormat: Serialize + for<'de> Deserialize<'de> + 'static {
+	/// A sequence of bytes to add at the beginning of the serialized
+	/// string, to identify that the data is of this version.
+	const VERSION_MARKER: &'static [u8] = b"";
+}
+
+impl<T: InitialFormat> Migrate for T {
+	const VERSION_MARKER: &'static [u8] = <T as InitialFormat>::VERSION_MARKER;
+
+	type Previous = NoPrevious;
+
+	fn migrate(_previous: Self::Previous) -> Self {
+		unreachable!();
+	}
+}
+
+/// Internal type used by InitialFormat, not meant for general use.
+#[derive(Serialize, Deserialize)]
+pub enum NoPrevious {}
+
+impl Migrate for NoPrevious {
+	type Previous = NoPrevious;
+
+	fn migrate(_previous: Self::Previous) -> Self {
+		unreachable!();
+	}
+
+	fn decode(_bytes: &[u8]) -> Option<Self> {
+		None
+	}
+
+	fn encode(&self) -> Result<Vec<u8>, rmp_serde::encode::Error> {
+		unreachable!()
+	}
+}
+
+#[cfg(test)]
+mod test {
+	use super::*;
+
+	#[derive(Serialize, Deserialize, PartialEq, Eq, Debug)]
+	struct V1 {
+		a: usize,
+		b: String,
+	}
+	impl InitialFormat for V1 {}
+
+	#[derive(Serialize, Deserialize, PartialEq, Eq, Debug)]
+	struct V2 {
+		a: usize,
+		b: Vec<String>,
+		c: String,
+	}
+	impl Migrate for V2 {
+		const VERSION_MARKER: &'static [u8] = b"GtestV2";
+		type Previous = V1;
+		fn migrate(prev: V1) -> V2 {
+			V2 {
+				a: prev.a,
+				b: vec![prev.b],
+				c: String::new(),
+			}
+		}
+	}
+
+	#[test]
+	fn test_v1() {
+		let x = V1 {
+			a: 12,
+			b: "hello".into(),
+		};
+		let x_enc = x.encode().unwrap();
+		let y = V1::decode(&x_enc).unwrap();
+		assert_eq!(x, y);
+	}
+
+	#[test]
+	fn test_v2() {
+		let x = V2 {
+			a: 12,
+			b: vec!["hello".into(), "world".into()],
+			c: "plop".into(),
+		};
+		let x_enc = x.encode().unwrap();
+		assert_eq!(&x_enc[..V2::VERSION_MARKER.len()], V2::VERSION_MARKER);
+		let y = V2::decode(&x_enc).unwrap();
+		assert_eq!(x, y);
+	}
+
+	#[test]
+	fn test_migrate() {
+		let x = V1 {
+			a: 12,
+			b: "hello".into(),
+		};
+		let x_enc = x.encode().unwrap();
+
+		let xx = V1::decode(&x_enc).unwrap();
+		assert_eq!(x, xx);
+
+		let y = V2::decode(&x_enc).unwrap();
+		assert_eq!(
+			y,
+			V2 {
+				a: 12,
+				b: vec!["hello".into()],
+				c: "".into(),
+			}
+		);
+
+		let y_enc = y.encode().unwrap();
+		assert_eq!(&y_enc[..V2::VERSION_MARKER.len()], V2::VERSION_MARKER);
+
+		let z = V2::decode(&y_enc).unwrap();
+		assert_eq!(y, z);
+	}
+}

src/util/persister.rs

@@ -1,23 +1,19 @@
 use std::io::{Read, Write};
 use std::path::{Path, PathBuf};
+use std::sync::{Arc, RwLock};
 
 use tokio::io::{AsyncReadExt, AsyncWriteExt};
 
-use serde::{Deserialize, Serialize};
-
-use crate::data::*;
 use crate::error::Error;
+use crate::migrate::Migrate;
 
-pub struct Persister<T: Serialize + for<'de> Deserialize<'de>> {
+pub struct Persister<T: Migrate> {
 	path: PathBuf,
 
 	_marker: std::marker::PhantomData<T>,
 }
 
-impl<T> Persister<T>
+impl<T: Migrate> Persister<T> {
-where
-	T: Serialize + for<'de> Deserialize<'de>,
-{
 	pub fn new(base_dir: &Path, file_name: &str) -> Self {
 		let mut path = base_dir.to_path_buf();
 		path.push(file_name);
@@ -27,18 +23,37 @@ where
 		}
 	}
 
+	fn decode(&self, bytes: &[u8]) -> Result<T, Error> {
+		match T::decode(bytes) {
+			Some(v) => Ok(v),
+			None => {
+				error!(
+					"Unable to decode persisted data file {}",
+					self.path.display()
+				);
+				for line in hexdump::hexdump_iter(bytes) {
+					debug!("{}", line);
+				}
+				Err(Error::Message(format!(
+					"Unable to decode persisted data file {}",
+					self.path.display()
+				)))
+			}
+		}
+	}
+
 	pub fn load(&self) -> Result<T, Error> {
 		let mut file = std::fs::OpenOptions::new().read(true).open(&self.path)?;
 
 		let mut bytes = vec![];
 		file.read_to_end(&mut bytes)?;
 
-		let value = rmp_serde::decode::from_read_ref(&bytes[..])?;
+		let value = self.decode(&bytes[..])?;
 		Ok(value)
 	}
 
 	pub fn save(&self, t: &T) -> Result<(), Error> {
-		let bytes = rmp_to_vec_all_named(t)?;
+		let bytes = t.encode()?;
 
 		let mut file = std::fs::OpenOptions::new()
 			.write(true)
@@ -57,12 +72,12 @@ where
 		let mut bytes = vec![];
 		file.read_to_end(&mut bytes).await?;
 
-		let value = rmp_serde::decode::from_read_ref(&bytes[..])?;
+		let value = self.decode(&bytes[..])?;
 		Ok(value)
 	}
 
 	pub async fn save_async(&self, t: &T) -> Result<(), Error> {
-		let bytes = rmp_to_vec_all_named(t)?;
+		let bytes = t.encode()?;
 
 		let mut file = tokio::fs::File::create(&self.path).await?;
 		file.write_all(&bytes[..]).await?;
@@ -70,3 +85,36 @@ where
 		Ok(())
 	}
 }
+
+pub struct PersisterShared<V: Migrate + Default>(Arc<(Persister<V>, RwLock<V>)>);
+
+impl<V: Migrate + Default> Clone for PersisterShared<V> {
+	fn clone(&self) -> PersisterShared<V> {
+		PersisterShared(self.0.clone())
+	}
+}
+
+impl<V: Migrate + Default> PersisterShared<V> {
+	pub fn new(base_dir: &Path, file_name: &str) -> Self {
+		let persister = Persister::new(base_dir, file_name);
+		let value = persister.load().unwrap_or_default();
+		Self(Arc::new((persister, RwLock::new(value))))
+	}
+
+	pub fn get_with<F, R>(&self, f: F) -> R
+	where
+		F: FnOnce(&V) -> R,
+	{
+		let value = self.0 .1.read().unwrap();
+		f(&value)
+	}
+
+	pub fn set_with<F>(&self, f: F) -> Result<(), Error>
+	where
+		F: FnOnce(&mut V),
+	{
+		let mut value = self.0 .1.write().unwrap();
+		f(&mut value);
+		self.0 .0.save(&value)
+	}
+}

src/util/token_bucket.rs

@@ -1,40 +0,0 @@
-use std::time::{Duration, Instant};
-
-use tokio::time::sleep;
-
-pub struct TokenBucket {
-	// Replenish rate: number of tokens per second
-	replenish_rate: u64,
-	// Current number of tokens
-	tokens: u64,
-	// Last replenish time
-	last_replenish: Instant,
-}
-
-impl TokenBucket {
-	pub fn new(replenish_rate: u64) -> Self {
-		Self {
-			replenish_rate,
-			tokens: 0,
-			last_replenish: Instant::now(),
-		}
-	}
-
-	pub async fn take(&mut self, tokens: u64) {
-		while self.tokens < tokens {
-			let needed = tokens - self.tokens;
-			let delay = (needed as f64) / (self.replenish_rate as f64);
-			sleep(Duration::from_secs_f64(delay)).await;
-			self.replenish();
-		}
-		self.tokens -= tokens;
-	}
-
-	pub fn replenish(&mut self) {
-		let now = Instant::now();
-		let new_tokens =
-			((now - self.last_replenish).as_secs_f64() * (self.replenish_rate as f64)) as u64;
-		self.tokens += new_tokens;
-		self.last_replenish = now;
-	}
-}