prisma-risk / tsoracle / 26255833479

95%

push

github

fix(stress): meaningful --ci-smoke output (counter, params, precision) (#55)

The CI smoke landed by ci/stress-pipeline-and-docs reported
`outcome=Ok, timestamps=10, all latency 0` for `--topology process
--ci-smoke --nodes 1` -- a misleading pass.  Diagnosis surfaced four
distinct defects, fixed here:

1. `events_observed` counted every SupervisorEvent variant (Issued,
   Chaos, Liveness, End), but `lib.rs` reported it as the user-visible
   "timestamps" field.  Under killer-loop, this inflated the count
   with chaos events while masking that zero real batches completed.
   Renamed to `issued_observed` and moved the increment into the
   Issued arm.  New unit test covers the mixed-event case.

2. `--ci-smoke` forced clients=16 + warmup=1000, requiring 16,000
   successful warmup RPCs per run.  Achievable on mem/raft but not on
   process+1-node under continuous kills, where healthy windows are
   hundreds of milliseconds.  Reduced to clients=8 + warmup=100,
   matching the known-good unit-test smoke shape.

3. `{:.0}` on the throughput format strings truncated sub-1.0 rates
   (e.g., `0.1` -> `"0"`), hiding the symptom from item 2.  Switched
   to `{:.2}`.

4. Both READMEs claimed latency "currently reads as zeros" with a
   "per-client histogram is a tracked follow-up" -- but the recording
   path has been in supervisor.rs since the first stress crate.
   Rewrote both notes to describe the real failure mode (empty
   histogram from no completed batches, e.g. continuous kills against
   an undersized topology).

After these fixes, the same command that previously reported
`client_calls=2 timestamps=10` now reports `client_calls=12216
timestamps=48864` with real p50/p90/p99 latency and 216 transient
retries (proving the killer-loop chaos actually landed), still
`violations: 0`.

50 of 50 new or added lines in 3 files covered. (100.0%)

6239 of 6484 relevant lines covered (96.22%)

89455.59 hits per line