Skip to content

Changelog

0.4.0 (2026-09-14)

Features

  • a container-owned channel pool for the Dishka providers, shared by every upstream told to borrow it (ecdc4e2), closes #31

0.3.0 (2026-09-14)

Features

  • Prometheus collector for the metrics protocol and Dishka providers for the client lifecycle (6dfdb44), closes #28

0.2.0 (2026-09-14)

Features

  • settings models for the environment, mirroring the dataclasses (edffb5f)

0.1.2 (2026-09-07)

Bug Fixes

  • an around_call teardown cannot change a call's outcome, and shares its setup's context (#20) (27b5c83)

0.1.1 (2026-09-06)

Bug Fixes

  • missing root exports, and three guides that described the wrong behaviour (#17) (b43d8cb)

0.1.0 (2026-08-13)

Initial release of grpc-client-kit.

Features

  • ChannelPool — channels pooled by their full identity (target, security, credentials, options, compression, interceptor chain); idle connections parked by gRPC core, health verdicts flag pooled channels
  • GrpcClient — a cheap stub factory over the pool: picks a target, asks for the matching channel, hands out a stub
  • GrpcClientFactory — maps one settings object onto the pool, balancer, health checker and per-target interceptor chains; eager settings validation, first-health-pass gate on entry, graceful shutdown_grace on exit
  • Streaming-aware interceptors — one around_call async-generator seam expands into all four grpc.aio adapter classes, and write()-style streaming calls return promptly
  • Canonical chain: logging, tracing, metrics, timeout, deadline budget, wait-for-ready, retry, circuit breaker — ordered so every layer sees what it needs
  • Timeout as the budget of the whole call, divided between retry attempts rather than granted afresh to each one
  • Deadline-budget propagation (deadline extra) — outgoing calls trimmed to what the caller's request has left
  • Wait-for-ready policy per method, with an optional deadline requirement
  • Retries with budget arithmetic, idempotency whitelists, jitter and retry metrics — and a warning when a native retryPolicy would multiply them
  • Circuit breaker per method and per target — half-open trials, bounded LRU state with OPEN circuits pinned, rejections reported as status="rejected", never error
  • Client-side load balancing — round-robin, random, weighted; health-aware narrowing and passive quarantine that takes a target out of rotation on the first real transport failure
  • Active grpc.health.v1 monitoring (health extra) on the checker's own channels, with wait_until_ready and per-service probing
  • Uniform observability — structured logs with sensitive-metadata redaction, Prometheus metrics, OpenTelemetry spans that parent correctly, all four RPC kinds covered
  • GrpcClientKitError family separating the kit's own refusals from server failures, while RPC-shaped ones remain AioRpcErrors
  • Connectivity tuning in seconds — keepalive and reconnect backoff, part of channel identity
  • TLS by default; zero-dependency core (grpcio only) — every integration is an opt-in extra