FleetGrain

Remarks

⚠ It is a register and an alert channel, not a controller. The decisions about one map belong to that map's grain, which is where the roster and the arrival rate are; collecting them here would make one grain the serialisation point for every placement in a region, which is precisely the bottleneck the per-map keying exists to avoid.

What it does own is the version every new shard is started on — because that is a decision about the region rather than about a map — and the escalation that a drain which cannot finish arrives at. Doc 27 § Drain: nothing is force-disconnected, and the escalation ends in a person deciding rather than in a timeout deciding for them.

Methods (7)

• public FleetGrain(ILogger<FleetGrain> log)

  Stands one up.

• public Task<ShardReport[]> Shards()

  Every shard in the region.

• public Task ShardChanged(ShardReport report)

  Told by a map grain when one of its shards changes.

• public Task Escalate(ShardId shard, string reason)

  A drain that could not finish.

• public Task<RealmVersion> Target()

  The version new shards are started on.

• public Task SetTarget(RealmVersion version)

  Points the fleet at a version. Rolling back is the same call with the old pair.

• public Task<double> VersionSpread()

  The fraction of shards not on Target — doc 27 § Upgrades' watched number.