WIP: garage_rpc: store layout version history

src/rpc/layout/history.rs

@@ -0,0 +1,170 @@
+use std::cmp::Ordering;
+use std::sync::Arc;
+
+use garage_util::crdt::{Crdt, Lww, LwwMap};
+use garage_util::data::*;
+use garage_util::encode::nonversioned_encode;
+use garage_util::error::*;
+
+use super::schema::*;
+use super::*;
+
+impl LayoutHistory {
+	pub fn new(replication_factor: usize) -> Self {
+		let version = LayoutVersion::new(replication_factor);
+
+		let staging_parameters = Lww::<LayoutParameters>::new(version.parameters);
+		let empty_lwwmap = LwwMap::new();
+
+		let mut ret = LayoutHistory {
+			versions: vec![version].into_boxed_slice().into(),
+			update_trackers: Default::default(),
+			staging_parameters,
+			staging_roles: empty_lwwmap,
+			staging_hash: [0u8; 32].into(),
+		};
+		ret.staging_hash = ret.calculate_staging_hash();
+		ret
+	}
+
+	pub fn current(&self) -> &LayoutVersion {
+		self.versions.last().as_ref().unwrap()
+	}
+
+	pub(crate) fn calculate_staging_hash(&self) -> Hash {
+		let hashed_tuple = (&self.staging_roles, &self.staging_parameters);
+		blake2sum(&nonversioned_encode(&hashed_tuple).unwrap()[..])
+	}
+
+	// ================== updates to layout, public interface ===================
+
+	pub fn merge(&mut self, other: &LayoutHistory) -> bool {
+		let mut changed = false;
+
+		// Merge staged layout changes
+		match other.current().version.cmp(&self.current().version) {
+			Ordering::Greater => {
+				self.staging_parameters = other.staging_parameters.clone();
+				self.staging_roles = other.staging_roles.clone();
+				self.staging_hash = other.staging_hash;
+				changed = true;
+			}
+			Ordering::Equal => {
+				self.staging_parameters.merge(&other.staging_parameters);
+				self.staging_roles.merge(&other.staging_roles);
+
+				let new_staging_hash = self.calculate_staging_hash();
+				if new_staging_hash != self.staging_hash {
+					changed = true;
+				}
+
+				self.staging_hash = new_staging_hash;
+			}
+			Ordering::Less => (),
+		}
+
+		// Add any new versions to history
+		let mut versions = self.versions.to_vec();
+		for v2 in other.versions.iter() {
+			if let Some(v1) = versions.iter().find(|v| v.version == v2.version) {
+				if v1 != v2 {
+					error!("Inconsistent layout histories: different layout compositions for version {}. Your cluster will be broken as long as this layout version is not replaced.", v2.version);
+				}
+			} else if versions.iter().all(|v| v.version != v2.version - 1) {
+				error!(
+					"Cannot receive new layout version {}, version {} is missing",
+					v2.version,
+					v2.version - 1
+				);
+			} else {
+				versions.push(v2.clone());
+				changed = true;
+			}
+		}
+		self.versions = Arc::from(versions.into_boxed_slice());
+
+		// Merge trackers
+		self.update_trackers.merge(&other.update_trackers);
+
+		changed
+	}
+
+	pub fn apply_staged_changes(mut self, version: Option<u64>) -> Result<(Self, Message), Error> {
+		match version {
+			None => {
+				let error = r#"
+Please pass the new layout version number to ensure that you are writing the correct version of the cluster layout.
+To know the correct value of the new layout version, invoke `garage layout show` and review the proposed changes.
+				"#;
+				return Err(Error::Message(error.into()));
+			}
+			Some(v) => {
+				if v != self.current().version + 1 {
+					return Err(Error::Message("Invalid new layout version".into()));
+				}
+			}
+		}
+
+		let mut new_version = self.current().clone();
+		new_version.version += 1;
+
+		new_version.roles.merge(&self.staging_roles);
+		new_version.roles.retain(|(_, _, v)| v.0.is_some());
+		new_version.parameters = *self.staging_parameters.get();
+
+		self.staging_roles.clear();
+		self.staging_hash = self.calculate_staging_hash();
+
+		let msg = new_version.calculate_partition_assignment()?;
+
+		let mut versions = self.versions.to_vec();
+		versions.push(new_version);
+		self.versions = Arc::from(versions.into_boxed_slice());
+
+		Ok((self, msg))
+	}
+
+	pub fn revert_staged_changes(mut self, version: Option<u64>) -> Result<Self, Error> {
+		match version {
+			None => {
+				let error = r#"
+Please pass the new layout version number to ensure that you are writing the correct version of the cluster layout.
+To know the correct value of the new layout version, invoke `garage layout show` and review the proposed changes.
+				"#;
+				return Err(Error::Message(error.into()));
+			}
+			Some(v) => {
+				if v != self.current().version + 1 {
+					return Err(Error::Message("Invalid new layout version".into()));
+				}
+			}
+		}
+
+		self.staging_roles.clear();
+		self.staging_parameters.update(self.current().parameters);
+		self.staging_hash = self.calculate_staging_hash();
+
+		// TODO this is stupid, we should have a separate version counter/LWW
+		// for the staging params
+		let mut new_version = self.current().clone();
+		new_version.version += 1;
+
+		let mut versions = self.versions.to_vec();
+		versions.push(new_version);
+		self.versions = Arc::from(versions.into_boxed_slice());
+
+		Ok(self)
+	}
+
+	pub fn check(&self) -> Result<(), String> {
+		// Check that the hash of the staging data is correct
+		let staging_hash = self.calculate_staging_hash();
+		if staging_hash != self.staging_hash {
+			return Err("staging_hash is incorrect".into());
+		}
+
+		// TODO: anythign more ?
+
+		self.current().check()
+	}
+}

src/rpc/layout/mod.rs

@@ -0,0 +1,32 @@
+mod history;
+mod schema;
+mod tracker;
+mod version;
+
+pub use history::*;
+pub use schema::*;
+pub use version::*;
+
+// ---- defines: partitions ----
+
+/// A partition id, which is stored on 16 bits
+/// i.e. we have up to 2**16 partitions.
+/// (in practice we have exactly 2**PARTITION_BITS partitions)
+pub type Partition = u16;
+
+// TODO: make this constant parametrizable in the config file
+// For deployments with many nodes it might make sense to bump
+// it up to 10.
+// Maximum value : 16
+/// How many bits from the hash are used to make partitions. Higher numbers means more fairness in
+/// presence of numerous nodes, but exponentially bigger ring. Max 16
+pub const PARTITION_BITS: usize = 8;
+
+const NB_PARTITIONS: usize = 1usize << PARTITION_BITS;
+
+// ---- defines: nodes ----
+
+// Type to store compactly the id of a node in the system
+// Change this to u16 the day we want to have more than 256 nodes in a cluster
+pub type CompactNodeType = u8;
+pub const MAX_NODE_NUMBER: usize = 256;

src/rpc/layout/schema.rs

@@ -0,0 +1,286 @@
+mod v08 {
+	use crate::layout::CompactNodeType;
+	use garage_util::crdt::LwwMap;
+	use garage_util::data::{Hash, Uuid};
+	use serde::{Deserialize, Serialize};
+
+	/// The layout of the cluster, i.e. the list of roles
+	/// which are assigned to each cluster node
+	#[derive(Clone, Debug, Serialize, Deserialize)]
+	pub struct ClusterLayout {
+		pub version: u64,
+
+		pub replication_factor: usize,
+		pub roles: LwwMap<Uuid, NodeRoleV>,
+
+		/// node_id_vec: a vector of node IDs with a role assigned
+		/// in the system (this includes gateway nodes).
+		/// The order here is different than the vec stored by `roles`, because:
+		/// 1. non-gateway nodes are first so that they have lower numbers
+		/// 2. nodes that don't have a role are excluded (but they need to
+		///    stay in the CRDT as tombstones)
+		pub node_id_vec: Vec<Uuid>,
+		/// the assignation of data partitions to node, the values
+		/// are indices in node_id_vec
+		#[serde(with = "serde_bytes")]
+		pub ring_assignation_data: Vec<CompactNodeType>,
+
+		/// Role changes which are staged for the next version of the layout
+		pub staging: LwwMap<Uuid, NodeRoleV>,
+		pub staging_hash: Hash,
+	}
+
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
+	pub struct NodeRoleV(pub Option<NodeRole>);
+
+	/// The user-assigned roles of cluster nodes
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
+	pub struct NodeRole {
+		/// Datacenter at which this entry belong. This information is used to
+		/// perform a better geodistribution
+		pub zone: String,
+		/// The capacity of the node
+		/// If this is set to None, the node does not participate in storing data for the system
+		/// and is only active as an API gateway to other nodes
+		pub capacity: Option<u64>,
+		/// A set of tags to recognize the node
+		pub tags: Vec<String>,
+	}
+
+	impl garage_util::migrate::InitialFormat for ClusterLayout {}
+}
+
+mod v09 {
+	use super::v08;
+	use crate::layout::CompactNodeType;
+	use garage_util::crdt::{Lww, LwwMap};
+	use garage_util::data::{Hash, Uuid};
+	use serde::{Deserialize, Serialize};
+	pub use v08::{NodeRole, NodeRoleV};
+
+	/// The layout of the cluster, i.e. the list of roles
+	/// which are assigned to each cluster node
+	#[derive(Clone, Debug, Serialize, Deserialize)]
+	pub struct ClusterLayout {
+		pub version: u64,
+
+		pub replication_factor: usize,
+
+		/// This attribute is only used to retain the previously computed partition size,
+		/// to know to what extent does it change with the layout update.
+		pub partition_size: u64,
+		/// Parameters used to compute the assignment currently given by
+		/// ring_assignment_data
+		pub parameters: LayoutParameters,
+
+		pub roles: LwwMap<Uuid, NodeRoleV>,
+
+		/// see comment in v08::ClusterLayout
+		pub node_id_vec: Vec<Uuid>,
+		/// see comment in v08::ClusterLayout
+		#[serde(with = "serde_bytes")]
+		pub ring_assignment_data: Vec<CompactNodeType>,
+
+		/// Parameters to be used in the next partition assignment computation.
+		pub staging_parameters: Lww<LayoutParameters>,
+		/// Role changes which are staged for the next version of the layout
+		pub staging_roles: LwwMap<Uuid, NodeRoleV>,
+		pub staging_hash: Hash,
+	}
+
+	/// This struct is used to set the parameters to be used in the assignment computation
+	/// algorithm. It is stored as a Crdt.
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Copy, Debug, Serialize, Deserialize)]
+	pub struct LayoutParameters {
+		pub zone_redundancy: ZoneRedundancy,
+	}
+
+	/// Zone redundancy: if set to AtLeast(x), the layout calculation will aim to store copies
+	/// of each partition on at least that number of different zones.
+	/// Otherwise, copies will be stored on the maximum possible number of zones.
+	#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Copy, Debug, Serialize, Deserialize)]
+	pub enum ZoneRedundancy {
+		AtLeast(usize),
+		Maximum,
+	}
+
+	impl garage_util::migrate::Migrate for ClusterLayout {
+		const VERSION_MARKER: &'static [u8] = b"G09layout";
+
+		type Previous = v08::ClusterLayout;
+
+		fn migrate(previous: Self::Previous) -> Self {
+			use itertools::Itertools;
+
+			// In the old layout, capacities are in an arbitrary unit,
+			// but in the new layout they are in bytes.
+			// Here we arbitrarily multiply everything by 1G,
+			// such that 1 old capacity unit = 1GB in the new units.
+			// This is totally arbitrary and won't work for most users.
+			let cap_mul = 1024 * 1024 * 1024;
+			let roles = multiply_all_capacities(previous.roles, cap_mul);
+			let staging_roles = multiply_all_capacities(previous.staging, cap_mul);
+			let node_id_vec = previous.node_id_vec;
+
+			// Determine partition size
+			let mut tmp = previous.ring_assignation_data.clone();
+			tmp.sort();
+			let partition_size = tmp
+				.into_iter()
+				.dedup_with_count()
+				.map(|(npart, node)| {
+					roles
+						.get(&node_id_vec[node as usize])
+						.and_then(|p| p.0.as_ref().and_then(|r| r.capacity))
+						.unwrap_or(0) / npart as u64
+				})
+				.min()
+				.unwrap_or(0);
+
+			// By default, zone_redundancy is maximum possible value
+			let parameters = LayoutParameters {
+				zone_redundancy: ZoneRedundancy::Maximum,
+			};
+
+			Self {
+				version: previous.version,
+				replication_factor: previous.replication_factor,
+				partition_size,
+				parameters,
+				roles,
+				node_id_vec,
+				ring_assignment_data: previous.ring_assignation_data,
+				staging_parameters: Lww::new(parameters),
+				staging_roles,
+				staging_hash: [0u8; 32].into(), // will be set in the next migration
+			}
+		}
+	}
+
+	fn multiply_all_capacities(
+		old_roles: LwwMap<Uuid, NodeRoleV>,
+		mul: u64,
+	) -> LwwMap<Uuid, NodeRoleV> {
+		let mut new_roles = LwwMap::new();
+		for (node, ts, role) in old_roles.items() {
+			let mut role = role.clone();
+			if let NodeRoleV(Some(NodeRole {
+				capacity: Some(ref mut cap),
+				..
+			})) = role
+			{
+				*cap *= mul;
+			}
+			new_roles.merge_raw(node, *ts, &role);
+		}
+		new_roles
+	}
+}
+
+mod v010 {
+	use super::v09;
+	use crate::layout::CompactNodeType;
+	use garage_util::crdt::{Lww, LwwMap};
+	use garage_util::data::{Hash, Uuid};
+	use serde::{Deserialize, Serialize};
+	use std::collections::HashMap;
+	use std::sync::Arc;
+	pub use v09::{LayoutParameters, NodeRole, NodeRoleV, ZoneRedundancy};
+
+	/// The layout of the cluster, i.e. the list of roles
+	/// which are assigned to each cluster node
+	#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+	pub struct LayoutVersion {
+		pub version: u64,
+
+		pub replication_factor: usize,
+
+		/// This attribute is only used to retain the previously computed partition size,
+		/// to know to what extent does it change with the layout update.
+		pub partition_size: u64,
+		/// Parameters used to compute the assignment currently given by
+		/// ring_assignment_data
+		pub parameters: LayoutParameters,
+
+		pub roles: LwwMap<Uuid, NodeRoleV>,
+
+		/// see comment in v08::ClusterLayout
+		pub node_id_vec: Vec<Uuid>,
+		/// see comment in v08::ClusterLayout
+		#[serde(with = "serde_bytes")]
+		pub ring_assignment_data: Vec<CompactNodeType>,
+	}
+
+	/// The history of cluster layouts
+	#[derive(Clone, Debug, Serialize, Deserialize)]
+	pub struct LayoutHistory {
+		/// The versions currently in use in the cluster
+		pub versions: Arc<[LayoutVersion]>,
+
+		/// Update trackers
+		pub update_trackers: UpdateTrackers,
+
+		/// Parameters to be used in the next partition assignment computation.
+		pub staging_parameters: Lww<LayoutParameters>,
+		/// Role changes which are staged for the next version of the layout
+		pub staging_roles: LwwMap<Uuid, NodeRoleV>,
+		/// Hash of the serialized staging_parameters + staging_roles
+		pub staging_hash: Hash,
+	}
+
+	/// The tracker of acknowlegments and data syncs around the cluster
+	#[derive(Clone, Debug, Serialize, Deserialize, Default)]
+	pub struct UpdateTrackers {
+		/// The highest layout version number each node has ack'ed
+		pub ack_map: UpdateTracker,
+		/// The highest layout version number each node has synced data for
+		pub sync_map: UpdateTracker,
+		/// The highest layout version number each node has
+		/// ack'ed that all other nodes have synced data for
+		pub sync_ack_map: UpdateTracker,
+	}
+
+	/// The history of cluster layouts
+	#[derive(Clone, Debug, Serialize, Deserialize, Default)]
+	pub struct UpdateTracker(pub HashMap<Uuid, u64>);
+
+	impl garage_util::migrate::Migrate for LayoutHistory {
+		const VERSION_MARKER: &'static [u8] = b"G010lh";
+
+		type Previous = v09::ClusterLayout;
+
+		fn migrate(previous: Self::Previous) -> Self {
+			let version = LayoutVersion {
+				version: previous.version,
+				replication_factor: previous.replication_factor,
+				partition_size: previous.partition_size,
+				parameters: previous.parameters,
+				roles: previous.roles,
+				node_id_vec: previous.node_id_vec,
+				ring_assignment_data: previous.ring_assignment_data,
+			};
+			let update_tracker = UpdateTracker(
+				version
+					.nongateway_nodes()
+					.iter()
+					.map(|x| (*x, version.version))
+					.collect::<HashMap<Uuid, u64>>(),
+			);
+			let mut ret = Self {
+				versions: Arc::from(vec![version].into_boxed_slice()),
+				update_trackers: UpdateTrackers {
+					ack_map: update_tracker.clone(),
+					sync_map: update_tracker.clone(),
+					sync_ack_map: update_tracker.clone(),
+				},
+				staging_parameters: previous.staging_parameters,
+				staging_roles: previous.staging_roles,
+				staging_hash: [0u8; 32].into(),
+			};
+			ret.staging_hash = ret.calculate_staging_hash();
+			ret
+		}
+	}
+}
+
+pub use v010::*;

src/rpc/layout/tracker.rs

@@ -0,0 +1,21 @@
+use super::*;
+
+impl UpdateTracker {
+	fn merge(&mut self, other: &UpdateTracker) {
+		for (k, v) in other.0.iter() {
+			if let Some(v_mut) = self.0.get_mut(k) {
+				*v_mut = std::cmp::max(*v_mut, *v);
+			} else {
+				self.0.insert(*k, *v);
+			}
+		}
+	}
+}
+
+impl UpdateTrackers {
+	pub(crate) fn merge(&mut self, other: &UpdateTrackers) {
+		self.ack_map.merge(&other.ack_map);
+		self.sync_map.merge(&other.sync_map);
+		self.sync_ack_map.merge(&other.sync_ack_map);
+	}
+}