Changelog

May 29, 2026 · View on GitHub

All notable changes to Beava are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

[0.0.6] - 2026-05-29

Resolves a SEV-1 reported by a downstream consumer: HTTP serving (including POST /ping) stalled for 60–90 s during every snapshot write, tripping docker healthchecks into a restart loop.

Fixed

  • Snapshot writes no longer block the data plane. Snapshots now run in a forked child via copy-on-write (Valkey BGSAVE pattern, default on unix); the apply thread holds the state lock only for the fork() syscall, so /ping and all HTTP/TCP requests stay responsive while a multi-hundred-MB snapshot is written. Set BEAVA_SNAPSHOT_FORK=0 to fall back to the in-process path.
  • WAL truncation reclaims on-disk bytes. Compaction rewrites the segment to drop snapshot-covered records instead of logging "truncated" while the file kept growing.
  • POST /push rejects control characters in string fields (HTTP 400, control_character_in_string) so corrupt bytes can't poison the WAL and recur as a decode WARN on every restart.
  • Hardened WAL recovery: applied-watermark replay, forced re-register on schema gaps, bounded fork-child reaping, and EINTR-safe fork wait.

Added

  • Configurable snapshot cadence — BEAVA_SNAPSHOT_INTERVAL_MS, BEAVA_SNAPSHOT_MIN_EVENTS (Redis-style conditional snapshot), and BEAVA_SNAPSHOT_FORK.
  • Snapshot metrics on the admin /metrics endpoint: last duration, bytes, and fsync time.

Docs

  • Install docs feature pip / brew / docker as the three primary paths.
  • Corrected homepage, quickstart, and field-guide snippets to match the shipped SDK and HTTP API; fixed homepage mobile layout overflow.

[0.0.4] - 2026-05-13

First release with behavioural fixes since v0.0.0. Three operator bugs that mis-routed data through the engine + SDK are gone; the rest of the release is regression coverage so the bug class can't return.

Fixed

  • WindowedOp::update_at honours the dispatcher's pre_val. Windowed field-bearing ops (mean/avg/sum/top_k/quantile with window=) were re-extracting from extracted[field_idx] instead of using the pre_val the outer apply-loop had already resolved via the agg-local → union-index remap. End-to-end symptoms: mean("price", window="…") returned Null and top_k("category", window="…") returned a different field's distribution whenever a prior windowed op in the same agg referenced a different field. Fix routes pre_val through the windowed arm.
  • EventTypeMixState::update_at honours the dispatcher's pre_val. Same bug class as the windowed fix; the second update_at arm that discarded pre_val and re-extracted from the wrong index space.
  • SDK ~bv.col(x) emits (not x), not !(x). The server's where-parser rejects bare unary ! (unexpected character '!') but accepts the not keyword. Every register that used the idiomatic ~bv.col(...) inside where= had been failing with aggregation_invalid_where.

Tests added (33 regression tests across five files)

  • crates/beava-core/tests/windowed_op_uses_caller_pre_val.rs — 4 tests.
  • crates/beava-core/tests/field_ordering_invariant.rs — 4 tests covering multi-agg reversed field order, non-overlapping subsets, top_k-after-mean stacking, geo lat/lon position swaps.
  • crates/beava-core/tests/agg_combinations_matrix.rs — 12 tests covering (where × window × field) 232^{3} matrix and 3+ mixed windowed ops in one agg.
  • crates/beava-server/tests/wire_roundtrip_parity.rs — 3 tests asserting HTTP-JSON and TCP-msgpack produce identical state for a complex schema.
  • python/tests/test_op_roundtrips_extended.py + test_col_overload_server_acceptance.py + internal/test_col_invert.py — 18 Python tests covering var / std / percentile / top_k / decay / geo / event_type_mix end-to-end plus bv.col overload server-acceptance.

Examples

  • examples/python/agent_runtime.py reverts the bv.col("ok") == False workaround back to idiomatic ~bv.col("ok").
  • marketplace_rerank.py no longer returns None for avg_view_price_30m and surfaces 'watches' for top_category_30m (it had been a corrupted price distribution pre-fix).

[0.0.3] - 2026-05-12

Syncs the Rust workspace version with the Python package version so beava --version reports the same string as pip show beava. Otherwise source-identical to v0.0.2.

[0.0.2] - 2026-05-12

Homebrew install path. Wires per-platform tarballs into the release pipeline so the homebrew-bump.yml workflow can bump Formula/beava.rb cleanly. Also syncs the Rust crate version with the Python package version so beava --version matches the pip metadata. Otherwise source-identical to v0.0.1.

[0.0.1] - 2026-05-12

First release published to PyPI. Source-identical to v0.0.0; the version bump exists so the inaugural PyPI publish can claim the beava project name under the Trusted-Publisher OIDC flow enabled in this release cycle.

0.0.0 - 2026-05-12

Initial public release. Beava is a single-binary real-time feature server for fraud, ad-tech, and behavioral analytics — declare a feature, push events, query it, in under ten minutes, with curl alone.

Added

  • Single-binary real-time feature server with HTTP/1.1 + framed-TCP wire transports; in-memory state, WAL + periodic snapshot durability, sub-millisecond batch-get.
  • 54 aggregation operators across 7 families: core, sketch, point-ordinal, recency, decay, velocity, and buffer-geo.
  • Python SDK with declarative @bv.event decorator and pipeline DSL — events.with_columns(...).group_by(K).agg(...).
  • @bv.table derivation outputs for aggregation results (output-only; no upsert / delete / retract).
  • TypeScript SDK (@beava/sdk) — communicate-only: push, register, get, batch-get.
  • Go SDK (github.com/beava-dev/beava/sdk/go) — communicate-only.
  • Schema evolution flags: force=True for destructive register with a structured diff payload, and dry_run=True for preview without applying.
  • Three demo workloads bundled with the Python SDK: bv.demo("adtech"), bv.demo("fraud"), bv.demo("ecommerce").
  • beava bench CLI with four modes: throughput, mixed, memory, fsync.
  • Global aggregations (no key= form) and bv.lit(value) literal expression.
  • Cold-entity TTL — opt-in lazy eviction via @bv.event(cold_after='<duration>').
  • Memory governance — every lifetime aggregation operator declares a finite per-entity ceiling at register time; unbounded ops are rejected at register.
  • In-memory persistence backend for ephemeral and test workloads.
  • OP_RESET (0x0040) opcode and POST /reset route, gated on test-mode (environment, config, or SDK kwarg).
  • Documentation site at beava.dev — quickstart, operator catalogue, wire spec, Python / TypeScript / Go SDK reference, and ADR archive.

Architecture

  • Hand-rolled mio-only single-threaded data plane (Redis-shaped); horizontal scaling via multi-instance entity-key sharding.
  • Apache-2.0 OSS license.

Removed

  • Tables (upsert / delete / retract API), MVCC temporal store, joins, unions, session windows, and event-time / watermark semantics — descoped to v0.1+.
  • event_time_ms field on the wire — Beava is processing-time only in v0.

Security

  • OWASP Top-10 and LLM Top-10 review against the v0 attack surface.
  • ASVS L1 threat model — single-tenant; operator owns network isolation; no auth in v0.
  • cargo audit + cargo deny policy enforced at workspace root.

Performance

  • ≥3M events/sec/core sustained on representative fraud workloads (single-thread mio data plane).
  • P99 batch-get latency under 10 ms warm.
  • Per-entity memory budget under 7 KB for a typical 30-feature pack.