Releases

This file scopes barca by release so the scope does not quietly expand.

0.1.x (shipped)

The Rust rewrite. Replaced the pure-Python prototype with a hybrid Rust+Python architecture: Rust parses, plans, and dispatches; Python executes.

Shipped:

• @asset with dependency tracking (inputs=), caching (run_hash), partitions
• @sensor for external state observation
• @task for side-effect operations (always re-runs, never cached)
• @sink for writing outputs to disk
• partitions(), partitions_from(), collect() for partition workflows
• Static analysis via ruff (never imports user code)
• Turso/libSQL for persistence (.barca/metadata.db)
• CLI: barca get, barca run, barca plan, barca history, barca stats
• barca serve HTTP API (axum)
• Retries with backoff (retries=, retry_backoff=)
• Benchmarks: 13-97x faster than Dagster/Prefect across all workloads

0.2.0 (shipped)

The execution engine rewrite. Replaced the old per-thread dispatch system with a stateless worker pool coordinated via Unix domain sockets.

Shipped:

• UDS coordination protocol: length-prefixed JSON over Unix domain sockets, 290K msg/s at 128 workers
• Stateless workers: receive one task at a time from Rust via a global ready queue. No pre-assigned queues, no head-of-line blocking.
• parallel() and parallel_map(): dynamic fan-out at runtime via SIGSTOP/SIGCONT. Scales to 1000+ items, nested parallel works recursively.
• Type-safe coordinator: load_phase() consumes the planner’s Phase directly. StepId on every Item. Step accounting invariant (assert on count mismatch).
• orjson optional dependency for faster Python serialization
• -4000 lines net: removed executor.rs, scheduler.rs, work_plan.rs, old dispatch loop

Issues closed: #70 (UDS communication), #58 (asset vs task model)

0.3.0 (shipped)

Goal: scheduling, remote I/O, observability.

Delivered:

• Cron scheduling enforcement — barca serve fires Schedule("0 5 * * *") assets, sensors, and tasks on their cron tick (#54). Includes:
  • Durable catch-up — last-fire times persist; a job whose tick passed during downtime fires once on restart.
  • Configurable timezone — --timezone local|utc|<IANA>.
  • Parallel runs — independent runs execute concurrently (bounded pool), with DB writes serialized; a job never overlaps itself.
  • Observability — GET /schedule and barca list <files> (next fire times); live reload under --watch.
• Python server client — barca.Client (stdlib-only) to trigger runs, poll status, and inspect schedules over the HTTP API.

0.4.0 (shipped)

Goal: remote artifact storage and execution reliability.

Delivered:

• Remote artifact backends — the artifact store and @sink destinations accept object-store URIs via fsspec scheme dispatch: Azure ADLS Gen2 (abfs:///abfss://) first-class, S3 and GCS pluggable (#55). Includes:
  • Optional extras — barca[azure], barca[s3], barca[gcs], barca[remote]; the core install stays zero-dependency.
  • BARCA_ARTIFACT_URI — points the primary artifact store at a remote prefix; BARCA_STORAGE_OPTIONS passes per-protocol fsspec options. Credentials use each backend’s native default chain.
  • Staged writes — serializer → local temp file → atomic rename (local) or chunked upload (remote); large payloads are never buffered in memory and a crash never leaves a partial artifact.
  • v1 limitations (explicit errors): partitions_from and parallel() result values require a local artifact store.
• @sink execution — previously parsed but inert; sinks now write after the parent asset materializes, with serializer= override, format precedence (kwarg → extension → primary format), per-partition filename suffixing, and error isolation (a sink failure never fails the asset; outcomes are logged and persisted). Sinks and @asset(serializer=) now participate in the definition hash — a one-time global cache invalidation on upgrade.
• Error surfacing — worker failures carry the exception type and a barca-frame-filtered traceback all the way to BarcaError and the DB.
• Real retries — backoff is actually applied (retry_backoff * attempt, non-blocking), every attempt runs in a fresh worker process, and attempt counts are recorded accurately; timeouts are reported as TimeoutError instead of a worker disconnect.

Planned (carried forward):

• Reproducible Docker benchmarks — containerized cross-framework comparisons with proper timeouts and resource constraints (#65)
• Backend abstraction — pluggable DB + storage, enabling Docker and shared deployments (#56)
• APM integration — Datadog + Sentry for observability (#59)
• Alerting hooks — Slack webhooks, email notifications (#52)

0.6.0 (current)

Goal: object storage on equal footing across clouds, held to one contract.

• First-class S3, GCS, and Cloudflare R2 — alongside Azure, the object stores are now peers, all held to the same shared-state contract (conditional create, cross-machine cache hit, concurrent-writer conflict → replay). R2 rides on the S3 backend (s3:// + an R2 endpoint in storage_options); see Remote Storage.
• Backend conformance suite — the identical pull/push/conflict/stale-token assertions run against every backend on every PR, using local emulators (MinIO for S3/R2, fake-gcs-server for GCS, Azurite for Azure) — no cloud credentials. A stale-token guard makes a too-lenient emulator fail loud instead of false-passing.
• Fixes surfaced by that suite:
  • Azure concurrent first-push races (FileExistsError from adlfs create-only) are now classified as conflicts and replayed, not hard errors.
  • GCS conditional overwrite (previously broken — the etag was fed to int(if_generation_match)); GCS shared state now rides entirely on the google-cloud-storage SDK, conditioning on the numeric generation.
  • barca[gcs]/barca[remote] now declare google-cloud-storage, which the state push path imports directly (gcsfs does not pull it in).

0.5.0 (shipped)

Goal: shared state across machines.

• Shared remote materialization state — the metadata DB lives as a blob ({uri}/{env}/state/metadata.db), pulled at run start and pushed with an etag/generation-conditional upload at run end (conflict → pull + replay). A run on one machine hits artifacts materialized by another.
• Content-addressed artifacts — {artifacts}/{node}/{run_hash}{ext} in every mode (local included): immutable objects, cross-machine cache hits.
• barca.toml — first config file ([remote] section, default_env), env vars override; see Configuration.
• --env environments — dev/staging/prod fully separated state, local and remote.

Planned follow-ups: barca serve with shared state (currently gated to state = "off"), barca gc for content-addressed artifact garbage collection, partitioned cache checks.

Future (unscoped)

Not yet assigned to a release:

• Partition filtering on CLI (--partition flag) (#57)
• File versioning for artifacts (#61)
• Notebook integration
• TUI
• Distributed execution (multi-machine)
• Web UI