Stratara 4.4.2

A release about a flag that cost too much. Every host sharing the replay coordination state over Redis
asked Redis, with a blocking round trip on the calling thread, whether a replay was active — and the
Orleans execution model asked on every hot path, eight to ten times per command on a host with three
projections, inside grain turns. On a two-core host that was enough to starve the thread pool and time
the Redis connection out, surfacing as failed catch-ups, lagging read models and failed grain
registrations although no replay ever ran. The answer now comes from memory. No migration is needed;
one optional setting is new.

Fixed

• No path waits on Redis to learn whether a replay is active. The Redis-backed replay state answers
  IsReplayActive from a field. A change made in the host itself is seen there at once; a change made
  elsewhere — a replay that began or ended in another host, or a marking that lapsed — is announced over
  a channel beside the replay-request channel and seen within its latency, and every host re-reads the
  marking every ProjectionReplay:RefreshSeconds regardless, as the safety net for a lost announcement.
  The Orleans execution model's command dispatcher, bundle dispatcher, store readers and drain, and the
  bus dispatchers, make no Redis call for it any more. Nothing changes for a host without Redis, whose
  state was in process already, nor for a host that registers its own IProjectionReplayState.

Added

• ProjectionReplayOptions.RefreshSeconds (default 5), read from the ProjectionReplay section
  like the lease. It bounds how long a host answers from a replay marking that changed elsewhere where the
  announcement was lost. Refused when the host starts at zero or below, and at or above
  LeaseSeconds, with an OptionsValidationException naming ProjectionReplay:RefreshSeconds.
• Log events 104_015 to 104_018. A host that cannot refresh the replay marking from Redis says
  so once at Warning (104_015) and keeps its last answer; 104_016 at Information says it reads the
  marking again. A host that cannot subscribe to the marking's announcements says so once at Warning
  (104_017), sees a change on the refresh period only, and 104_018 says it is subscribed again.

Changed

• A host that cannot reach Redis keeps its last answer about the replay. Before, every dispatch and
  every catch-up on such a host failed with the connection's exception. Now they proceed on what the host
  last learned — initially that no replay is active — and the loss is logged once. The replay stays
  correct regardless: on the Orleans execution model the readers are paused by the rebuild and the
  replay themselves, and apply the store from the beginning once more when the replay ends.
• Rolling upgrade. A host on a release before this one announces nothing when it starts or ends a
  replay, so a host on this release sees that replay at the refresh period rather than at once. The
  marking, its key and its lease are unchanged, so the two releases share one replay state.