Skip to content

Release v19.124.0

Choose a tag to compare

@github-actions github-actions released this 04 Nov 11:08
a672a5b

Added

  • External token authority options for the .NET client. (part of #4350)
  • Benchmarks for decision reads, guarded appends and concurrent decision reads (#4856)
  • Failed partition groups page in the Workbench groups current failed partitions by observer and failure kind (part of #5539)
  • Added observers.maxConcurrentRegistrations to limit how many reactor and reducer registrations a kernel silo processes at once, so rolling client updates no longer exhaust the storage connection pool (part of #5514)

Changed

  • A configured Authentication:Authority is now used to obtain client tokens (through OIDC discovery). To keep the previous behaviour, remove Authority from the client configuration. (part of #4350)
  • New PostgreSQL read-model tables for schemas whose key is spelled ID get a single ID key column; existing tables are unchanged (#4590)
  • Publish workflows for the Kotlin and TypeScript clients validate the version input before publishing.
  • Workflow inputs are passed to scripts through environment variables.
  • Chronicle's telemetry setup (AddChronicleTelemetry) no longer exports Orleans runtime and grain-call spans by default; set Cratis:Chronicle:Telemetry:OrleansTracing=true to restore them. Trace context propagation remains enabled, and the public AddChronicleInstrumentation() still adds the Orleans sources. With Orleans tracing off, untraced grain calls, the default and including appends, are 4-10x cheaper, and sampled calls are about 20% cheaper. (#4607)
  • Verify-mode SQL replicas now refuse to start until the chronicle_sink_cycle_receipts table exists and is current for every event store and namespace. Run --migrate-only with the new version before upgrading Verify replicas. Run it again for a namespace created later. (#5526)
  • Document JSON naming policies for object-valued migration defaults and value maps, including migration hashes and mixed-client rollout considerations. (#5581)
  • Changed the MongoDB health check to read the driver's cluster state without taking a pooled connection, and storage operations now retry when the MongoDB connection wait queue is full (part of #5514)
  • CI test retries now stop when a retry fails exactly the same tests as the attempt before it, instead of re-running against the same stuck fixture (part of #4057)
  • Chronicle now depends on Cratis.Orleans 1.34.0, Cratis.Fundamentals 7.32.0, Cratis.Arc 22.156.4 and Microsoft.Orleans 10.4.2 or later (#5532)
  • MongoDB index provisioning also creates the index of the new job-step-plans collection on the jobs database (#5532)

Fixed

  • Token failure logs no longer include response bodies. (part of #4350)
  • A job start that is delayed or retried can no longer run after the observer has handed its work to a newer job, so a stale catch-up or replay no longer releases, overwrites or quarantines the work of the job that replaced it. (#5382)
  • SQL read models (PostgreSQL, SQL Server, SQLite) now apply a projection event that writes several rows in one transaction. A failure or a retried commit no longer leaves the read model partly updated or applies the change twice. (part of #5229)
  • Reading a named occurrence of a SQL read model, such as a revert backup, no longer creates tables or databases; occurrence tables with an unexpected shape are treated as absent instead of returning partial records
  • SQL read models whose identifier is spelled ID can now be stored on SQLite and SQL Server, instead of failing because a second Id key column was added (#4590)
  • Kotlin client: token request failures no longer include the token endpoint's response or other server-supplied text in the exception message
  • Catch-up jobs for large bursts no longer exceed MongoDB's 16 MB document limit and stall the observer. On MongoDB, large job step plans are now kept outside the job document (#5532)
  • A Given event that the in-process kernel refuses now fails the specification with GivenEventAppendFailed, naming the event type and the reasons, instead of being silently dropped. Specifications that passed only because a Given event was silently refused will now fail and need their seeded history corrected. (#5618)
  • A kernel silo no longer crash-loops when registering its own projections times out during startup, for example while many clients re-register after a rolling restart. It starts, keeps retrying the registration in the background with a jittered, capped backoff, and reports its health as degraded until the registration succeeds, so it stays reachable for clients. (part of #5627)

Security

  • Update Cratis.Choreography to 2.1.2