Repository navigation
ci: e2e engine package runs at the 15m timeout (TestReseedHorizonRate under -race -coverpkg) #750
Description
The flake
The e2e workflow runs go test -timeout 15m -race -v -coverpkg=./... ./.... The engine package now runs within seconds of that timeout:
- main after fix(bucketindex): scope block identity to the tenure and fence commits on it #747 (run 36487241735):
enginetook 874.9s.TestReseedHorizonRatealone took 714s. - fix(bucketindex): keep split lineage and rollup layouts in the index #749 (run 36493407676):
enginehit the 900s timeout whileTestReseedHorizonRatewas still running (11m48s in).
This isn't a regression. On the same machine, with -race, the test takes 13.3s on both origin/main and #749. The CI cost comes from -coverpkg=./..., which instruments every package, on a slower runner. The test runs 300 flush+merge cycles over 400 series.
Direction
Cut the cost of that one test under the e2e profile rather than raising the timeout for everything. Options:
- skip it under
-raceortesting.Short, since it is a measurement that logs its numbers; - fewer cycles when instrumented;
- move it to the benchmark job.
Alternatively, give the engine package its own e2e step with its own timeout.
🤖 Generated with Claude Code
Activity
tdakkota commentedon Sep 29, 2026
Cause: CPU, not waiting
TestReseedHorizonRatehas no sleeps or timers and runs against the in-memory backend. It is a measurement test:MaxPartBytes, so parts turn over constantly.A CPU profile of the plain run spreads across the codecs (float/timestamp encode and decode, part cursors, zstd) and GC. There is no hotspot and no idle time.
-race-race -coverpkg=./...(the e2e job), measured locally on mainCoverage multiplies the race time by about 12×. Under
-race, coverage runs in atomic mode, so every basic block in every package does an atomic increment. This test's hot path is tight per-value codec loops. The CI runner is about 4× slower again.Suggested fix
Keep a small run in normal test runs, for example 20 cycles, which is enough for the two assertions (merge-consumed parts stay accounted for; live parts stay under
MaxWants). Run the full 300-cycle measurement only without race and coverage, or behind an opt-in flag.TestReseedHorizonWithRetentionhas the same shape and needs the same treatment.🤖 Generated with Claude Code
https://claude.ai/code/session_01KcLq1G8KpWBvM53JaQb1tS