Skip to content

Release v19.136.2

Choose a tag to compare

@github-actions github-actions released this 07 Nov 14:27
6ac9aca

Changed

  • An observer's subscription, its reader scope, and event types are handed between grains of the same silo as they are instead of being copied through JSON, and a job step shares one copy of the event types its observer reads instead of keeping its own. A wave of thousands of job steps, such as the catch-up of an observer with thousands of partitions, held a copy of the same few dozen event types and subscription per step.
  • The converter between dynamic objects and JSON and the type formats it uses are shared instead of created again for every grain that needs one. Each job step grain held four of them, each with its own tables.

Fixed

  • Starting the kernel on a database with a large catch-up job no longer runs out of memory. Looking up the retry or replay jobs of an observer listed every job, each with its whole plan of thousands of steps, and copied each of them through JSON on the way. Only the jobs of the type asked for are listed now, and a job state is not copied within a silo.
  • Reading a stored job state no longer writes the whole document, plan included, out a second time before it is deserialized.
  • A failed partition whose recovery job could not be started, because its grain could not activate while the silo was busy, no longer ends the observer's recovery turn. It is tried again a few seconds later, instead of the partitions queued behind it waiting for the next subscription.