Skip to content

v2.0.0

Latest
Latest

Choose a tag to compare

View all tags
@mpangrazzi mpangrazzi released this 01 Oct 09:59
v2.0.0
01b9f7f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Hayhooks 2.0 introduces durable execution: Haystack Pipelines and Agents that keep running after the HTTP request returns, survive restarts and deploys, and can pause for a human. State is checkpointed in Redis, and any replica can pick the work back up.

pip install "hayhooks[durable]"

Durable execution requires Haystack 3.1+ and Redis 6.2+ (or Valkey 7.2+). Without it, Hayhooks works as before.

⏳ Durable Execution

  • Detached runs. Submitting returns 202 with an execution ID while workers carry on in the background.
  • Restart recovery. If a worker dies, its run is picked up again once the lease expires, and fencing stops a stale worker from writing results.
  • Pipeline and Agent checkpoints. Pipelines resume from a chosen component boundary or a Haystack failure snapshot. Agents restore their state between LLM and tool steps, so finished work isn't repeated.
  • Human-in-the-loop. A run can suspend with a public wait reason and continue later with typed resume input.
  • Retries and cancellation. Crash recovery and application-requested retries have separate budgets, and cancellation is cooperative with a durable terminal result.
  • Idempotent submissions. A caller-supplied idempotency key lets clients retry a submission safely and rejects conflicting reuse.
  • Live streaming. SSE streams can be reattached with Last-Event-ID. They deliver chunks as workers flush them and end with the authoritative result.

🧭 Durable Mode for the Hayhooks Server

  • Set HAYHOOKS_DURABLE_MODE=true to serve wrappers that implement run_durable / run_durable_async. Each one gets submit, inspect, cancel, resume and stream routes.
  • Durable mode loads a fixed set of pipelines at startup. Live deploy/undeploy routes and MCP tools are not available in this mode, and any startup failure stops the server.
  • Optional owner scoping restricts each execution to the caller that submitted it.
  • GET /status reports durable_mode and durable store health for readiness checks.
  • HAYHOOKS_DURABLE_* settings control Redis, capacity, payload and stream limits, leases, retries and shutdown handoff.

🧩 Portable Engine

  • Use the engine in any asyncio or FastAPI app without the Hayhooks server: DurableDeployment, DurableRuntime, RedisExecutionStore / MemoryExecutionStore, and create_durable_router.
  • hayhooks.durable only needs Pydantic and Loguru. FastAPI, Haystack and Redis are loaded only when you use them.

⚠️ Upgrade Notes

  • Default (live) mode now rejects durable wrappers. Serve them with HAYHOOKS_DURABLE_MODE=true.
  • In durable mode, hayhooks run --workers must be 1. Scale out with replicas instead.
  • Several internal server helpers in hayhooks.server.utils (MCP, A2A and module loading) now take an explicit pipeline registry.

📚 Documentation

What's Changed

  • feat(durable): portable durable execution engine with Lua-guarded Redis protocol by @mpangrazzi in #267
  • feat(durable): fixed runtime lifecycle, drained shutdown, and portable imports by @mpangrazzi in #268
  • feat(durable): durable mode with a fixed pipeline registry and managed durable hosting by @mpangrazzi in #270
  • fix(durable): harden recovery, streaming, and owner access by @mpangrazzi in #271

Full Changelog: v1.25.0...v2.0.0

Contributors

mpangrazzi
Assets 2
Loading