Minor items deferred from PR #27's correctness review.
-
DurationStore.recordInvocation ignores firstOnShard. Every other duration path excludes a first-on-shard result from the estimates that drive slowest-first ordering, because the first unit on a cold JVM absorbs the whole per-JVM setup cost (~17.6s for the first consumer). For distributed invocations that cost lands on whichever position happened to run first on a slot, skewing that position's estimate upward for good. Scheduling quality only — never correctness — but the exclusion already exists everywhere else and this path was missed.
-
Stray blank line in CoordinatorCore.sweepSilentShards.
Minor items deferred from PR #27's correctness review.
DurationStore.recordInvocationignoresfirstOnShard. Every other duration path excludes a first-on-shard result from the estimates that drive slowest-first ordering, because the first unit on a cold JVM absorbs the whole per-JVM setup cost (~17.6s for the first consumer). For distributed invocations that cost lands on whichever position happened to run first on a slot, skewing that position's estimate upward for good. Scheduling quality only — never correctness — but the exclusion already exists everywhere else and this path was missed.Stray blank line in
CoordinatorCore.sweepSilentShards.