feat(helexa-bench): percentiles + prefill/decode split in store & report (#86) #101
Reference in New Issue
Block a user
Delete Branch "feat/86-bench-percentiles"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Part of the Performance observability epic (#83); builds on #85 (now merged).
Bench reported only a single median per cell, hiding tail latency and unable to record the server-measured prefill/decode split now emitted on
usage.helexa_timing(#85).usage.helexa_timingintoScenarioMetrics(prefill_ms, decode_ms, prefill_tokens).ensure_columnsmigration (pre-#85 DBs backfill as NULL); aggregation now emits p50/p95/p99 for TTFT and total (nearest-rank) plus a prefill-tok/s median derived from the split.Tests: nearest-rank percentile, idempotent backfill migration, and a report cell asserting percentiles + prefill split. All four validation jobs green.
Closes #86.