Release v19.32.3

Observers no longer get stuck behind a catch-up that is finishing, observers of a new namespace no longer share partition state, and MongoDB projections no longer lose a write that arrives at the end of a catch-up.

Changed

• Upgrade every kernel silo in a cluster together rather than one at a time. While silos run different versions, an observer that finishes catching up may not be notified and may deliver some events again once it is recovered.

Fixed

• An observer no longer stays behind the tail when an event is appended while its catch-up job is finishing. A new catch-up starts for that event instead of waiting for the finishing job, and overlapping catch-ups share a single job. (#4548)
• Observers created without stored state no longer share their catching-up, replaying and in-flight partition sets, so one observer's catch-up no longer makes the others drop live events. (#4558)
• A reducer batch that fails part-way is retried from the start of the batch, so events folded before the failure are no longer skipped. (#4540)
• Webhooks in event stores other than the System store now keep their authorization and their updates to event types, target URL and headers, and adding, updating or removing a webhook now fails when one of its events could not be appended. (#4567)
• MongoDB: a projection write that arrives while a catch-up's final bulk flush is in progress is now written instead of dropped, so the read model no longer stays behind an observer that reports the event as handled. (#4598)