prisma-risk / tsoracle / 26369003684

95%

Build Type

push

github

Committed by web-flow

Commit Message

feat(stress): detect non-overlapping cross-client real-time monotonicity (#135) (#297)

* feat(stress): detect non-overlapping cross-client real-time monotonicity (#135)

The supervisor enforced per-client monotonicity, cross-client uniqueness,
per-batch ordering, and (only in a chaos window's post-grace tail) fence
freshness. The strictly cross-client real-time case slipped through: client
A's RPC completes before client B's RPC starts, yet B receives a ts no
greater than A's. That is a real happens-before violation a monotonic TSO
must never exhibit, and a steady-state regression (e.g. a clean elective
failover that re-allocates a physical-ms window) would never correlate with
a ChaosKind window, so fence freshness could miss it entirely.

Add a fifth invariant in on_issued: track a single O(1) completed_high_water
tuple (highest ts seen, plus the recv_time of the RPC that produced it),
advanced only when ts strictly increases. A sample whose issued_at is at or
after that completion but whose ts does not advance past it is flagged as the
new CrossClientRealtimeMonotonicity violation kind.

The check keys off the per-sample issued_at/recv_time the client driver
records, not mpsc arrival order, so the existing tolerance of concurrent
arrival reordering still holds: overlapping RPCs never satisfy
issued_at >= prior.recv_time. Two gates keep the signal clean: a sample
already named by a per-client or uniqueness violation is not double-reported,
and the check is suppressed while a leadership-changing window is open, where
fence freshness is the authoritative cross-leadership check (failpoint
windows do not suppress, since real-time monotonicity must hold across them).

stress emits the new kind in its JSON output, so stress-auto-issue dedups it
under a distinct fingerprint with no workflow change.

monotonicity_tolerates_cross_client_arrival_reorder previously used
zero-duration, sequentially increasing instants that did not model the
concurrency it clai... (continued)

Coverage Stats

162 of 163 new or added lines in 3 files covered. (99.39%)

2 existing lines in 1 file now uncovered.

11267 of 11873 relevant lines covered (94.9%)

485264.13 hits per line