Files
rustfs/ARCHITECTURE.md
T
Zhengchao An 2ebf8bc138 refactor(ecstore): drop the client shim, import rustfs-s3-client directly (#6668)
* refactor(ecstore): drop the client shim, import rustfs-s3-client directly

Completes the migration window opened by the rustfs-s3-client extraction (rustfs/backlog#1842 PR3): every consumer now imports the client crate directly and the crate::client shim is deleted.

- All in-crate crate::client:: paths (tier warm backends, tier core, lifecycle tier_sweeper, replication storage boundary, set_disk) now import rustfs_s3_client::* directly; crates/ecstore/src/client/mod.rs and the lib.rs mod client declaration are gone.
- The two server-side modules historically misfiled under client/ move to their real homes: object_api_utils.rs to crates/ecstore/src/object_api/ (it builds engine-side object readers/writers), and object_handlers_common.rs to crates/ecstore/src/bucket/lifecycle/ (it is the lifecycle noncurrent-version cleanup helper). The latter now routes its replication calls through the lifecycle replication_sink boundary (schedule_delete wrapper and the sink's ReplicationObjectBridge re-export), as the lifecycle guard requires.
- The ecstore public facade drops api::client: object_api_utils is exposed as api::object_api_utils, and the rustfs crate takes admin_handler_utils (AdminError) from rustfs-s3-client directly (new dependency).
- Guard updates: the migration guard no longer pins mod client in ecstore's lib.rs or the admin_handler_utils facade module (it pins the new api::object_api_utils facade instead), and the module-lint register follows object_api_utils.rs to its new path.

Verification: cargo check -p rustfs-ecstore --all-targets and -p rustfs; cargo fmt --all; tier/transition/lifecycle-focused nextest (626 passed) and the decommission/rebalance/heal families in a filtered run (603 passed; the full-suite parallel run only fails on this machine's known decommission/rebalance baseline flakes, which pass in filtered reruns and fail identically on pristine origin/main); layer/migration/s3s/logging/error-format/doc-path guard scripts all pass.

* docs(architecture): record the S3 client extraction and reword invariant 4 (#6669)

Closes the documentation step of rustfs/backlog#1842. ARCHITECTURE.md invariant 4 now states the serving-vs-consuming distinction the adversarial ruling asked for: ecstore must not serve HTTP/S3 wire types, while consuming remote S3 endpoints is a legitimate engine capability that lives in the extracted rustfs-s3-client crate. The violation note is updated from the pre-extraction snapshot (58 files, embedded client) to the current ratcheted state (shrink-only S3S_ECSTORE_FILES_BASELINE in scripts/check_s3s_footprint.sh, object_lock converted first), and the crate map gains s3-client. ecstore-module-split-plan.md gets the client-directory entry the plan was missing: a Current Shape row and a completed-extraction section describing the pure-move + shim + direct-import sequence and the re-homing of the two misfiled server-side modules.
2026-08-26 22:02:36 +08:00

340 lines
17 KiB
Markdown

# ARCHITECTURE.md
> Last updated: 2026-08-12 · Revision: 3
>
> This document describes the high-level architecture of RustFS.
> If you want to familiarize yourself with the code base, you are in the right place!
>
> See also [CONTRIBUTING.md](CONTRIBUTING.md) for development workflow.
> See also [docs/architecture](docs/architecture/overview.md) for active
> architecture migration guardrails.
## Bird's Eye View
RustFS is a high-performance, S3-compatible distributed object storage system written
in Rust. It uses erasure coding for data durability, supports multi-tenancy through
IAM/STS, and provides a web-based admin console.
A running RustFS node exposes:
- **S3 API** (port 9000) — the primary data path for object CRUD
- **Admin API** (port 9000, `/minio/` prefix) — cluster management, IAM, metrics
- **Console** (port 9001) — web UI backed by the Admin API
- **Inter-node RPC** (gRPC/tonic) — cluster communication for distributed mode
The core data flow for a PUT request looks like:
```
HTTP request
→ server (TLS, auth, routing, compression)
→ app/object_usecase (validation, policy, lifecycle)
→ storage/ecfs (erasure coding, encryption, checksums)
→ ecstore (disk pool selection, data distribution)
→ rio (reader pipeline: encrypt → compress → hash → write)
→ io-core (buffer pool, storage profiling, admission control)
→ local disk / remote disk via RPC
```
## Code Map
The repository is a Cargo workspace with a flat `crates/` layout:
```
rustfs/ # Workspace root (virtual manifest)
├── rustfs/ # Main binary + library crate
│ └── src/
│ ├── main.rs # Entry point, startup sequence
│ ├── lib.rs # Module tree root
│ ├── server/ # HTTP server, TLS, routing, middleware
│ ├── admin/ # Admin API handlers and console
│ ├── app/ # Use-case layer (object, bucket, multipart)
│ ├── storage/ # Storage engine interface and implementation
│ ├── auth.rs # S3 request authentication
│ ├── config/ # CLI args, config parsing, workload profiles
│ └── ...
├── crates/ # library crates (authoritative list: Cargo.toml [workspace].members)
│ ├── ecstore/ # Erasure-coded storage engine
│ ├── rio/ # Reader I/O pipeline (encrypt, compress, hash)
│ ├── io-core/ # Buffer pool, storage profiling, backpressure/deadlock policy, lock optimizer, operation progress
│ ├── io-metrics/ # I/O metrics collection
│ ├── common/ # Shared runtime state, globals, data usage types
│ ├── config/ # Configuration types and parsing
│ ├── utils/ # Pure utility functions
│ ├── ... # (see "Crate Reference" below)
│ └── e2e_test/ # End-to-end integration tests
└── docs/ # Design documents and analysis
```
### Main Crate Layers (`rustfs/src/`)
The main crate is organized in layers, top to bottom:
| Layer | Directory | Responsibility |
|-------|-----------|----------------|
| **Server** | `server/` | HTTP listener, TLS, CORS, compression, middleware, graceful shutdown |
| **Admin** | `admin/` | Admin API routing, 30+ handler modules, web console |
| **App** | `app/` | Use-case orchestration: object (per-operation modules under `app/object/`, re-exported as `object_usecase`), bucket_usecase, multipart_usecase |
| **Storage** | `storage/` | S3 API translation, erasure-coded FS, SSE encryption, RPC, concurrency |
| **Auth** | `auth.rs` | S3 signature verification, credential validation |
| **Config** | `config/` | CLI parsing, config struct, workload profiles |
A request flows **downward** through the layers. No layer should reach upward
(e.g., storage must not import from admin).
### Crate Reference
`Cargo.toml` is the authoritative workspace membership and `cargo tree` is the
authoritative dependency graph. This overview deliberately avoids line-count
and dependency-depth snapshots because both quickly become stale during
refactors.
#### By Domain
| Domain | Current workspace crates | Responsibility |
|--------|--------------------------|----------------|
| Foundation | `checksums`, `common`, `config`, `data-usage`, `heal-contracts`, `scanner-contracts`, `utils` | Shared configuration, data-usage models, heal/scanner domain contracts, utilities, and checksums. |
| I/O and storage | `concurrency`, `ecstore`, `filemeta`, `heal`, `io-core`, `io-metrics`, `lifecycle`, `lock`, `object-capacity`, `object-data-cache`, `replication`, `rio`, `rio-v2`, `s3-client`, `scanner`, `storage-api` | Erasure-coded object storage, metadata, recovery, lifecycle, replication, locking, cache, I/O pipelines, and the engine-side S3 client for remote tier/transition targets. |
| Security and identity | `credentials`, `crypto`, `iam`, `keystone`, `kms`, `policy`, `security-governance`, `signer`, `tls-runtime`, `trusted-proxies` | Credentials, authentication, authorization, encryption, key management, TLS, and security contracts. |
| Protocols and contracts | `extension-schema`, `madmin`, `protos`, `protocols`, `s3-ops`, `s3-types`, `s3select-api`, `s3select-query` | Admin, inter-node, S3, S3 Select, and optional protocol contracts. |
| Operations and integration | `audit`, `notify`, `obs`, `targets`, `zip` | Auditing, observability, event delivery, notification targets, and archive support. |
| Test support | `e2e_test`, `test-utils` | End-to-end validation and shared test bootstrap utilities. |
The `rustfs` binary crate composes these libraries into the running server.
`ecstore` remains the storage engine at the architectural center; its internal
module split is tracked under `docs/architecture/`. `rio-v2` is the
feature-gated MinIO on-disk format compatibility I/O layer; it ships in no
default build (lifecycle:
[docs/architecture/minio-file-format-compat.md](docs/architecture/minio-file-format-compat.md)).
## Architecture Invariants
> These are rules that the codebase should follow. Some are currently violated
> (marked with ⚠️). Documenting them here makes the violations explicit and
> trackable.
1. **Layers flow downward.** Server → Admin/App → Storage → ecstore → rio/io-core.
No upward imports.
2. **Leaf crates have zero internal dependencies.** `config`, `credentials`, `crypto`,
`io-metrics`, and `madmin` should depend only on external crates.
- ✅ RESOLVED: the historical `utils → config` and `common → filemeta`/`madmin`
edges were removed; do not reintroduce them (see Known Structural Issues).
3. **Each type has exactly one definition.** Types shared across crates must be defined
in one crate and re-exported or imported by others.
- ⚠️ VIOLATED: `ReplicationStats` names three unrelated types
(`crates/data-usage/src/data_usage.rs`,
`crates/obs/src/metrics/collectors/replication.rs`,
`crates/ecstore/src/bucket/replication/replication_state.rs`) — a naming
collision, not copies; renaming is tracked in rustfs/backlog#1847.
- `LastMinuteLatency` has two deliberately different implementations: the
per-second bucketed accumulator in `crates/scanner-contracts/src/last_minute.rs` and
the in-memory endpoint-health sample tracker in
`crates/ecstore/src/bucket/bucket_target_sys.rs` (its doc comment explains
why it stays local).
- ✅ RESOLVED: `BackpressureConfig` and `DataUsageInfo` each have exactly one
definition (`crates/io-core/src/backpressure.rs`,
`crates/data-usage/src/data_usage.rs`). The zero-consumer
`BackpressureSettings` copy that lingered in io-metrics was removed
(rustfs/backlog#1833).
4. **ecstore does not *serve* HTTP or the S3 wire protocol.** It operates on
storage-level abstractions (objects, buckets, disks, pools) and holds no
wire or DTO types of the serving surface. *Consuming* remote S3-compatible
endpoints (ILM tier warm backends, transition targets) is a legitimate
engine capability, but it lives in the dedicated `rustfs-s3-client` crate
(`crates/s3-client`, extracted from the formerly embedded
`crates/ecstore/src/client/` by rustfs/backlog#1842), not inside ecstore.
- ⚠️ PARTIALLY VIOLATED: serving-side `s3s` references remain in ecstore
(bucket metadata/replication/lifecycle DTOs and error mapping). The
count is ratcheted shrink-only by `scripts/check_s3s_footprint.sh`
(`S3S_ECSTORE_FILES_BASELINE`; the `object_lock` module was converted to
storage-level types as the first ratchet step). Target state: the
baseline reaches zero and ecstore's `Cargo.toml` drops `s3s`.
5. **The `rustfs` binary crate is the only place that wires everything together.**
Individual crates should be testable in isolation.
6. **Error types use `thiserror` with descriptive names** (e.g., `StorageError`,
not bare `Error`).
- ✅ RESOLVED (strategy): `snafu` is gone from source
(`rg -l snafu crates/ rustfs/` is empty) and library code no longer uses
`anyhow` (remaining hits are test code and the `e2e_test` crate; `heal`
uses `thiserror`).
- ⚠️ VIOLATED (naming): 6 crates still export a bare `pub enum Error`:
`crypto`, `filemeta`, `heal`, `iam`, `policy`, and `replication`
(`src/resync.rs`) — all `thiserror`-derived.
## Known Structural Issues
> This section documents known problems in the current architecture.
> It exists so the team can track and address them deliberately.
### Critical
- **scanner/data-usage duplicate `.usage-cache.bin` serialization types.** The
original finding ("common/scanner code duplication, ~3K lines") is resolved:
`scanner` imports the shared data-usage types from `rustfs-data-usage` (see
the `pub use rustfs_data_usage::…` re-exports at the top of
`crates/scanner/src/data_usage_define.rs`). What remains: `scanner` and
`data-usage` each hold their own serialization types for the scanner cache
file (`DataUsageCacheInfo`/`DataUsageEntryInfo` in
`crates/scanner/src/data_usage_define.rs` vs
`DataUsageCacheInfo`/`DataUsageEntry` in
`crates/data-usage/src/data_usage.rs`); convergence is tracked in
rustfs/backlog#1828.
- **ecstore is a monolith (265 files, ~288K lines — roughly half is inline
`#[cfg(test)]` code).** Measured with
`find crates/ecstore/src -name '*.rs' | xargs wc -l`. It contains disk
management, bucket management, erasure coding, replication, lifecycle, RPC,
and configuration — all in one crate. It should be decomposed along its
existing subdirectories; the split plan lives in
[docs/architecture/ecstore-module-split-plan.md](docs/architecture/ecstore-module-split-plan.md).
### High
- **Dependency inversions.** Historical `utils → config` and
`common → filemeta/madmin` edges must stay removed so leaf/helper crates do
not regain upward dependencies.
- **Three-layer backpressure/deadlock policy bridging** across io-core,
concurrency, and `rustfs/src/storage`. The config types are no longer
duplicated (`BackpressureConfig` and `DeadlockDetectorConfig` are each
defined once, in io-core). Storage policies expose and consume explicit
projections into the concurrency/io-core policy shapes, and workload
admission snapshots are composed through provider registries; later work
should use those bridges before deleting compatibility wrappers.
### Medium
- **Bare `Error` naming.** Error-handling strategy has converged on `thiserror`
(no `snafu`, no `anyhow` in library code); the remaining inconsistency is the
bare `pub enum Error` naming in the 6 crates listed under Invariant 6.
- **`common` is mostly parked domain code, not shared utilities.** Of its
6,724 lines, ~83% is scanner/heal domain code stranded there to break
dependency cycles (`metrics.rs`, ~4,810 lines of scanner-domain metrics;
`heal_channel.rs`, ~776 lines of heal-domain channel types). The
"common vs utils" naming ambiguity is secondary to moving that code to its
domain owners.
## Cross-Cutting Concerns
### Error Handling
The project convention is `thiserror` for typed errors with descriptive names.
See `AGENTS.md`: "Prefer thiserror for library-facing error types."
```rust
// GOOD
#[derive(Debug, thiserror::Error)]
pub enum StorageError {
#[error("disk not found: {0}")]
DiskNotFound(String),
}
// AVOID
pub enum Error { ... } // too generic
anyhow::Result<T> // in library code (OK in tests/CLI)
```
### Logging & Tracing
- Use `tracing` crate (`info!`, `warn!`, `error!`, `debug!`, `trace!`)
- Structured fields: `tracing::info!(bucket = %name, "created bucket")`
- Spans for request-scoped context
### Metrics
- Prometheus-style metrics via `rustfs-obs` runtime and schema
- I/O-specific counters via `rustfs-io-metrics`
- Registration happens at crate level, collection/reporting in `rustfs-obs`
### Testing
- Unit tests: `#[cfg(test)] mod tests` in the same file
- Integration tests: inside respective crates (not top-level `tests/`)
- E2E tests: `crates/e2e_test/` — tests against a running server
- Run all: `make test` or `cargo nextest run`
## Startup Sequence
The binary (`main.rs`) boots in this order:
1. Environment variable compatibility (`MINIO_*``RUSTFS_*`)
2. Tokio runtime construction
3. CLI argument parsing
4. License, observability, TLS, trusted proxies initialization
5. Config parsing, server address resolution
6. Credentials, endpoints, local disks, lock client initialization
7. Capacity management initialization
8. HTTP server start (S3 API + optional console)
9. ECStore initialization (erasure coding storage engine)
10. Global config, background replication, KMS
11. Optional: FTP/FTPS/WebDAV servers
12. Event notifier, audit system, deadlock detector
13. Bucket metadata, IAM, Keystone, OIDC
14. Scanner and heal manager
15. Metrics system, mark `FullReady`
16. Wait for shutdown signal → graceful shutdown
## Dependency Diagram (Simplified)
```
┌─────────┐
│ rustfs │ (binary + lib)
│ main │
└────┬────┘
┌───────────────┼───────────────┐
│ │ │
┌────▼────┐ ┌────▼────┐ ┌──────▼─────┐
│ server │ │ admin │ │ app │
│ (HTTP) │ │(console)│ │(use-cases) │
└────┬────┘ └────┬────┘ └──────┬─────┘
│ │ │
└───────────────┼───────────────┘
┌──────▼──────┐
│ storage │
│ (ecfs, SSE, │
│ RPC, ACL) │
└──────┬──────┘
┌──────────────────┼──────────────────┐
│ │ │
┌─────▼──────┐ ┌──────▼──────┐ ┌──────▼──────┐
│ ecstore │ │ rio │ │ io-core │
│ (core) │ │ (readers) │ │ (buffers) │
└─────┬──────┘ └─────────────┘ └─────────────┘
┌─────┬──┼──┬─────┬──────┐
│ │ │ │ │ │
common utils config policy filemeta ...
```
## How to Navigate
- **"Where does S3 PutObject go?"**
`server/` routes → `app/object_usecase` validates → `storage/ecfs` encodes →
`ecstore` distributes → `rio` encrypts/compresses → `io-core` supplies buffers
- **"Where are bucket policies enforced?"**
`app/bucket_usecase` calls into `crates/policy/`
- **"Where is replication configured?"**
`admin/handlers/replication.rs` and `admin/handlers/site_replication.rs` for API,
`ecstore/src/bucket/replication/` for engine
- **"Where do I add a new admin endpoint?"**
Add handler in `admin/handlers/`, register in `admin/router.rs`
- **"Where do I add a new metric?"**
Define descriptor/collector in `crates/obs/src/metrics/`, expose via `/minio/v2/metrics`
---
*Inspired by [matklad's ARCHITECTURE.md](https://matklad.github.io/2021/02/06/ARCHITECTURE.md.html)
and [rust-analyzer's architecture.md](https://github.com/rust-analyzer/rust-analyzer/blob/master/docs/book/src/contributing/architecture.md).*