You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Fixed
A client's reactors and reducers no longer lose their subscription when the kernel runs several replicas and a client's keep-alives or observer streams reach a different silo than its connection. Every silo now finds the connection wherever it is registered, and a keep-alive that lands on the wrong silo is forwarded to the silo that holds the connection, so the client is not expired while alive. (#5665)
A silo of 19.123.1 or later that finds an earlier release in the cluster now stops with a clear message to stop every silo and start the new release, instead of crash-looping with TypeLoadException on the Cratis.Orleans to Cratis.Choreography rename. The rolling-deploy guide lists the boundary. (#5666)
A replay job that the observer watchdog finds missing is no longer restarted every minute without limit. Restarts back off from one minute up to one hour and reset once a replay job is running. A job that stays in preparation without a single step for 15 minutes is now deleted so the work can start again. (#5667)
Reducer state sent to the client with every reduce is now compact JSON instead of indented JSON, and a warning names reducers whose state exceeds 8 MiB. The state is still sent in full with every reduce, so a very large read model remains expensive. (part of #5669)