Skip to content

Release v19.130.0

Choose a tag to compare

@github-actions github-actions released this 05 Nov 23:07
916a8db

Client reactors now receive events as pages that span partitions, live, during catch-up and during replay, so a batch reactor is called once per page instead of once per partition. The kernel also stops waking healthy observers every minute, survives transient storage errors on background tasks, answers repeated projection registrations at once, and refuses to start on System data it is not looking at.

Added

  • Reactors whose client declares page delivery receive one call per page, with each partition settled on its own. A failing partition is parked at its first unhandled event and the rest of the page still counts (#5679)
  • Catch-up and replay of a page-declaring reactor run as one ordered sweep over the whole observer, using the same pages. Clients that do not declare page delivery are unaffected.
  • Reducer state caching. A client that opts in at registration (the .NET client does) keeps the state it last returned per partition, and the kernel tells it the state is unchanged instead of serializing the whole read model for every reduce. Memory and CPU per reduce no longer grow with the size of the read model while it is unchanged between reduces. Clients that do not opt in, and older kernels, behave as before (#5669)
  • The kernel refuses to start, naming both database names, when the System event store is empty under the configured name but holds events under the other one, instead of creating a fresh administrator roster in an empty store. The downgrade path is documented in the storage configuration page (#5690)

Changed

  • An observer registers its alert reminder only while there is something to report, and the first tick is spread over the following minute so reminders no longer tick in the same second (#5685)
  • The repeat move of the System event store databases is versioned 19.128.2, so it runs on the next 19.x release instead of never (#5689)
  • A projection registration that repeats one that just completed is answered at once instead of queueing behind subscription work and timing out against its own grain (#5688)
  • The observer interface gains a page method under a new version, so rolling upgrades inside the major keep working.

Fixed

  • A transient storage error in a fire-and-forget background task is logged and no longer terminates the kernel. Unobserved task exceptions are logged once per owner and failure with the owning task named, and an unhandled crash exits with a distinct code (#5687)
  • A cold start no longer wakes every healthy observer through its alert reminder at the same moment (#5685)

Reducer state is still sent in full to clients that do not opt in to state caching, and the full state is still read from the sink for every reduce. A joining silo still counts calls into a busy sibling towards its retry attempts (part of #5676).