Changelog¶

All notable changes to this project will be documented in this file.

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

[Unreleased]¶

Dependency floors raised to what the test suite proves: pico-ioc >= 2.3.3 (was 2.2.0; below 2.3.0 the settings prefix fails to resolve). A new CI job runs the suite with every declared floor pinned, so a floor that installs but does not work can no longer ship.

__all__ declares the public API and tests/test_exports.py pins it, per the ecosystem stability policy (ADR-014 in pico-ioc).

stop() claims the loop and thread atomically under the registrar lock, so concurrent stops (ASGI lifespan plus a manual container.shutdown()) can no longer interleave — the loser sees no loop and returns. This was the root cause of the level-2 teardown hang that 0.1.3 only bounded; defense in depth alongside pico-ioc 2.3.3's idempotent container shutdown. Announced in 0.1.4 but that build shipped without the code change; the regression test now drives a real loop and fails against the 0.1.4 stop().

Concurrent-stop regression test. The atomic stop() it was meant to pin missed this build (the test passed vacuously without a running loop); fixed for real in 0.1.5.

stop() is best-effort: if the clean shutdown exceeds its deadline, it logs a warning and forces the loop down instead of raising through container.shutdown(). An unclean broker disconnect beats a bricked application exit.

Shutdown uses consumer_start_timeout_seconds as its deadline too: stopping real consumers (group leave + offset commit) under load exceeded the 10s produce timeout and made container.shutdown() raise. Found by the flagship level-2 teardown.

Consumer startup no longer reuses produce_timeout_seconds as its deadline: new kafka.consumer_start_timeout_seconds setting (default 60s). Cold brokers (fresh KRaft clusters) exposing the group coordinator slowly used to kill the whole container boot at 10s.

@kafka_consumer(topic, group_id=) marker for component methods (sync and async); JSON record values decoded before dispatch.
@kafka_producer / @produce(topic) declarative producing clients (sync stubs block on ack, async stubs await).
KafkaRegistrar: dedicated background asyncio loop; one aiokafka consumer task per subscribed method; skip-on-failure policy so poison records never stall a partition; everything stopped on container shutdown.
Settings under the kafka prefix: bootstrap_servers, enabled, group_id, produce_timeout_seconds (zero-config).
Auto-discovery via the pico_boot.modules entry point.