Repository navigation
test: SIGKILL crash test must verify committed data, not only that the file opens #384
Description
Problem
tests/crash_kill_test.rs kills a child process during transact and checkpoint. After the kill it checks only two things:
Minigraf::opensucceeds.[:find ?e :where [?e :k ?v]]returns without an error.
It never checks that the data is correct. The query uses the AEVT path. #370 broke the EAVT (entity-bound) path and left AEVT intact. So this test passed every night while #370 was live.
The child also writes one fact per transact, so it cannot hit #371.
Ask
- The child records each transaction it has confirmed as committed. For example, it writes the
tx_countto a side file afterexecutereturns and fsyncs that file. - After the kill and reopen, the parent checks that every confirmed transaction is visible:
- through an entity-bound query (EAVT),
- through an attribute scan (AEVT),
- through a full scan.
All three must return the same result.
- No data from a transaction that was never confirmed may be visible.
- Vary the workload: multi-valued batches, batched retracts, several checkpoints before the kill, and kills timed inside
save(). - Run again after a second reopen, to catch problems that
checkpoint()copies forward.
Why
This is the cheapest way to catch the class of bug that shipped in #370 and #371.
Activity
ci(crash-kill): alternate the nightly run between main and v3 (#384)
ci(crash-kill): alternate the nightly run between main and v3 (#384)
Merge pull request #484 from project-minigraf/test/384-crash-kill-data
Merge pull request #485 from project-minigraf/test/384-crash-kill-dat…
release: v2.0.4
adityamukho commentedon Oct 14, 2026
Done: #484 on
v3and the v2.x backport #485, released in v2.0.4. The nightly crash-kill job alternates betweenmainandv3.