Files
rustfs/crates/io-metrics
houseme 0a00d8d500 perf(server): lighten internode data-plane stack (#3735)
* refactor(server): split internode dispatch scaffold

* test(server): cover internode dispatch prefix split

* refactor(server): name internode stack boundaries

* perf(server): skip internode request logging layer

* perf(server): skip internode trace layer

* perf(server): use lite internode request context

* feat(metrics): track internode rpc duration

* feat(ecstore): add put object stage summary logs

* test(metrics): update internode descriptor expectations

* fix(server): tighten internode path matching

* fix(pr): address review follow-up comments

* style(ecstore): simplify commit tail duration field

* refactor(ecstore): group put stage summary fields

* refactor(ecstore): inline put stage summary log

* fix(s3): return storage class for object attributes

* merge: sync latest main and resolve object attributes conflict

* fmt

* fix(server): remove duplicate rpc imports

* build(deps): bump memmap2 for RUSTSEC-2026-0186

* fix(s3select): align object_store with datafusion

* chore(deps): prune workspace dependencies

* perf(fuzz): optimize CI runtime with build/run split and matrix parallelization

  Separate fuzz harness compilation from execution to eliminate redundant
  builds across targets. Introduce matrix-based parallel execution for
  PR smoke and nightly fuzz jobs.

  Changes:

  - Split CI workflow into `fuzz-build` (compile once) and matrix run jobs
    (`pr-fuzz-smoke`, `nightly-fuzz-corpus`) that execute targets in parallel
  - Add `BUILD_ONLY` mode to run_ci_targets.sh / run_nightly_targets.sh
  - Add run_single_target.sh for matrix jobs (no build phase)
  - Optimize `local_metadata` fuzz target: reduce prefix iterations from
    8-10 (4 functions each) to 5 critical prefixes (parser-only), cutting
    per-iteration cost by ~3-5x
  - Move archive path validation (`validate_extract_relative_path`,
    `normalize_extract_entry_key`) from `rustfs` to `rustfs-utils::path`,
    eliminating `rustfs` binary crate dependency from fuzz harness
  - Remove `rustfs` from fuzz/Cargo.toml (drops significant transitive deps)
  - Add unit tests for archive path validation in rustfs-utils
  - Update fuzz/README.md with new workflow and script documentation

  Expected CI improvement: PR smoke wall-clock from ~120min (frequent
  timeout) to ~40min; nightly from ~180min to ~60min.

* refactor(fuzz): consolidate scripts and fix prefix test alignment

Replace three duplicated shell scripts (run_ci_targets.sh,
run_nightly_targets.sh, run_single_target.sh) with a single
parameterized run.sh that supports BUILD_ONLY, SKIP_BUILD, and
MAX_TOTAL_TIME environment variables.

Fix local_metadata fuzz target prefix testing: replace always-true
'len > 0' guard with lengths aligned to xl.meta binary layout
(4/5/8/12 bytes for magic+version+header fields). Remove redundant
empty-slice test.

Hoist RUSTFLAGS to workflow top-level env to eliminate per-job
duplication. Update README with unified script documentation.

Net: -118 lines, zero functionality loss.

* perf(fuzz): optimize CI runtime with build/run split and matrix parallelization

Restructure fuzz CI workflow to eliminate redundant compilation and
run targets in parallel via matrix strategy.

Workflow changes:
- Split into fuzz-build (compile once) and matrix run jobs
- PR smoke: 3 targets parallel, 60s each, timeout 30min (was 120min)
- Nightly: 3 targets parallel, 300s each, timeout 60min (was 180min)
- Pass compiled harness via actions/artifact between jobs
- Hoist RUSTFLAGS to workflow top-level env

Script consolidation:
- Replace 3 duplicated scripts with single parameterized run.sh
- Supports BUILD_ONLY, SKIP_BUILD, MAX_TOTAL_TIME env vars

Target optimizations:
- Remove rustfs binary crate from fuzz dependencies (was pulling
  979 transitive deps); move archive path validation to rustfs-utils
- Optimize local_metadata: reduce prefix iterations from 8-10x4
  calls to 5 prefixes with parser-only (no decompress), aligned
  with xl.meta binary layout (4/5/8/12 bytes)
- Add unit tests for archive path validation in rustfs-utils
- Update fuzz/README.md with unified script documentation

Expected: PR smoke wall-clock from ~120min (frequent timeout)
to ~40min; nightly from ~180min to ~60min.

* fix(rpc): resolve internode metrics via app context

* fmt

* ci: speed up fuzz smoke artifact restore

---------

Signed-off-by: houseme <housemecn@gmail.com>
2026-06-23 21:36:39 +08:00
..

rustfs-io-metrics

CI Status Crates.io

· Home · Docs · Issues · Discussions


Overview

rustfs-io-metrics is the metrics and configuration module for RustFS, a distributed object storage system. It provides:

  • Cache Configuration: L1/L2 tiered cache configuration management
  • Adaptive TTL: Dynamic TTL adjustment based on access frequency
  • Metrics Collection: Unified metrics recording and reporting
  • Bandwidth Monitoring: Real-time bandwidth observation and analysis
  • Performance Metrics: I/O performance metrics collection
  • Unified Configuration: Centralized configuration management
  • Exporter Boundary: Emit via metrics, export via rustfs-obs, no Prometheus HTTP endpoint

Features

Cache Configuration

Tiered cache configuration management:

use rustfs_io_metrics::{CacheConfig, CacheConfigError};

// Create configuration
let config = CacheConfig::new();

// Validate configuration
if let Err(e) = config.validate() {
    println!("Invalid configuration: {}", e);
}

// Custom configuration
let config = CacheConfig {
    max_capacity: 10_000,
    default_ttl_seconds: 300,
    max_memory_bytes: 100 * 1024 * 1024,  // 100 MB
    ..Default::default()
};

Adaptive TTL

Dynamic TTL adjustment based on access frequency:

use rustfs_io_metrics::{AdaptiveTTL, AdaptiveTTLStats};
use std::time::Duration;

let config = CacheConfig::new().with_ttl_range(60, 300, 3600);
let ttl = AdaptiveTTL::new(config);

// Cold object (few accesses)
let cold_ttl = ttl.calculate_ttl(Duration::from_secs(60), 1, 0.8);
println!("Cold object TTL: {:?}", cold_ttl);

// Hot object (many accesses)
let hot_ttl = ttl.calculate_ttl(Duration::from_secs(60), 100, 0.8);
println!("Hot object TTL: {:?}", hot_ttl);

Access Tracking

Track cache item access patterns:

use rustfs_io_metrics::{AccessTracker, AccessRecord};
use std::time::Duration;

let mut tracker = AccessTracker::new(1000, Duration::from_secs(300));

// Record accesses
tracker.record_access("object-key-1", 1024);
tracker.record_access("object-key-1", 1024);
tracker.record_access("object-key-2", 2048);

// Get access count
let count = tracker.get_access_count("object-key-1");
println!("Access count: {}", count);

// Detect hot/cold
if tracker.is_hot("object-key-1", 1) {
    println!("Hot object");
}

// Get top keys
let top_keys = tracker.top_keys(10);
for (key, count) in top_keys {
    println!("{}: {} accesses", key, count);
}

Metrics Recording

Unified metrics recording functions:

use rustfs_io_metrics::{
    // I/O scheduler metrics
    record_io_scheduler_decision,
    record_io_strategy_change,
    record_io_load_level,
    
    // Cache metrics
    record_cache_size,
    
    // Backpressure metrics
    record_backpressure_event,
    record_backpressure_state,
    
    // Timeout metrics
    record_timeout_event,
    record_operation_duration,
};

// Record I/O scheduler decision
record_io_scheduler_decision("sequential", "high_priority");

// Record cache size
record_cache_size("L1", 1024, 1);

// Record backpressure event
record_backpressure_event("warning", 0.85);

// Record operation timeout
record_timeout_event("GetObject", Duration::from_secs(30));

Internode Transport Metrics

Internode metrics are recorded by src/internode_metrics.rs. Aggregate metrics remain unlabeled for compatibility with existing dashboards:

Metric Meaning
rustfs_system_network_internode_sent_bytes_total Total internode bytes sent by this node.
rustfs_system_network_internode_recv_bytes_total Total internode bytes received by this node.
rustfs_system_network_internode_requests_outgoing_total Total outgoing internode requests.
rustfs_system_network_internode_requests_incoming_total Total incoming internode requests.
rustfs_system_network_internode_errors_total Total internode errors.
rustfs_system_network_internode_dial_errors_total Failed internode connection attempts.
rustfs_system_network_internode_dial_avg_time_nanos Average internode dial duration.

Operation-level metrics use the same low-cardinality label set:

Metric Labels Meaning
rustfs_system_network_internode_operation_sent_bytes_total operation, backend Bytes sent for an internode operation.
rustfs_system_network_internode_operation_recv_bytes_total operation, backend Bytes received for an internode operation.
rustfs_system_network_internode_operation_requests_outgoing_total operation, backend Outgoing request attempts for an internode operation.
rustfs_system_network_internode_operation_requests_incoming_total operation, backend Incoming request attempts for an internode operation.
rustfs_system_network_internode_operation_errors_total operation, backend Failed internode operation attempts.
rustfs_system_network_internode_operation_classified_errors_total operation, backend, classification Classified internode transport failures.
rustfs_system_network_internode_operation_retries_total operation, backend, classification Retry attempts for retryable internode transport failures.
rustfs_system_network_internode_operation_retry_successes_total operation, backend, classification Successful recoveries after retryable internode transport failures.
rustfs_system_storage_erasure_write_quorum_failures_total stage, dominant_error Erasure write quorum failures grouped by failure stage and dominant error class.

Current operation values are read_file_stream, put_file_stream, walk_dir, grpc_read_all, and grpc_write_all. Current backend values are tcp-http for the InternodeDataTransport TCP/HTTP path and grpc for the remaining gRPC byte paths. The compatibility wrapper uses unknown only for callers that have not been classified yet.

Success/failure is intentionally not a high-cardinality label today. Failures are represented by rustfs_system_network_internode_operation_errors_total; successful completions are not emitted as a dedicated result-labeled metric. Adding completion/result labels is a follow-up once stream completion semantics are defined consistently for request setup, body transfer, and shutdown.

Current low-cardinality classification values come from the TCP/HTTP internode path and include:

  • connect_timeout
  • connection_refused
  • dns_resolution_failed
  • connection_reset
  • body_stream_aborted
  • http_429
  • http_502
  • http_503
  • http_504
  • http_status_other
  • unknown

scripts/run_internode_transport_baseline.sh --metrics-url ... records metric deltas with operation and backend columns, so the TCP baseline can attribute bytes and request/error counts to tcp-http transport operations.

Unified Configuration

Centralized configuration management:

use rustfs_io_metrics::{
    IoConfig, CacheSettings, IoSchedulerSettings,
    BackpressureSettings, TimeoutSettings,
};

let config = IoConfig::new()
    .with_cache(CacheSettings::new()
        .with_max_capacity(10_000)
        .with_ttl(std::time::Duration::from_secs(300)))
    .with_scheduler(IoSchedulerSettings::new()
        .with_max_concurrent_reads(64))
    .with_backpressure(BackpressureSettings::new())
    .with_timeout(TimeoutSettings::new());

// Access configuration
println!("Cache capacity: {}", config.cache.max_capacity);
println!("Max concurrent reads: {}", config.scheduler.max_concurrent_reads);

Module Structure

rustfs-io-metrics/
├── src/
│   ├── lib.rs               # Module entry
│   ├── cache_config.rs      # Cache configuration
│   ├── adaptive_ttl.rs      # Adaptive TTL
│   ├── config.rs            # Unified configuration
│   ├── io_metrics.rs        # I/O metrics
│   ├── backpressure_metrics.rs # Backpressure metrics
│   ├── deadlock_metrics.rs  # Deadlock metrics
│   ├── lock_metrics.rs      # Lock metrics
│   ├── timeout_metrics.rs   # Timeout metrics
│   ├── internode_metrics.rs # Internode transport metrics
│   ├── bandwidth.rs         # Bandwidth monitoring
│   ├── global_metrics.rs    # Global metrics
│   └── performance.rs       # Performance metrics
└── Cargo.toml

Testing

# Run all tests
cargo test --package rustfs-io-metrics

# Run specific tests
cargo test --package rustfs-io-metrics --lib adaptive_ttl

# Run benchmarks
cargo bench --package rustfs-io-metrics --bench metrics_pipeline

Documentation

This crate records metrics through the Rust metrics crate and leaves exporting to rustfs-obs or the application-level observability pipeline. It does not expose Prometheus-compatible HTTP endpoints such as /rustfs/v2/metrics/cluster or /rustfs/v2/metrics/node.

API documentation can be generated locally:

cargo doc --package rustfs-io-metrics --no-deps --open

Useful source references:

  • rustfs-io-core: Core I/O scheduling
  • rustfs: Main storage service

License

Apache License 2.0