Description
Describe the bug
The unit_union_blobref_invalidation test failed with a "reference never released" exception.
To Reproduce
Unfortunately, I can't reproduce the bug locally. The failing test can be run using cabal run lsm-tree-test -- -p "unit_union_blobref_invalidation"
Expected behavior
Every reference should eventually be released, so such an exception should not occur.
Output
The test output can be found here: https://github.com/IntersectMBO/lsm-tree/actions/runs/15437001994/job/43445623120?pr=751
unit_union_blobref_invalidation: FAIL
Exception: Exceptions thrown while committing an action registry.
A registered action threw an error:
The error:
Reference is never released: RefId 488704
Allocation site: CallStack (from HasCallStack):
newRefWithTracker, called at src-control/Control/RefCount.hs:316:5 in lsm-tree-0.1.0.0-inplace-control:Control.RefCount
dupRef, called at src/Database/LSMTree/Internal/MergingRun.hs:279:47 in lsm-tree-0.1.0.0-inplace:Database.LSMTree.Internal.MergingRun
Registration site:
CallStack (from HasCallStack):
mkAction, called at src-control/Control/ActionRegistry.hs:644:40 in lsm-tree-0.1.0.0-inplace-control:Control.ActionRegistry
delayedCommit, called at src/Database/LSMTree/Internal/MergeSchedule.hs:300:7 in lsm-tree-0.1.0.0-inplace:Database.LSMTree.Internal.MergeSchedule
Desktop (please complete the following information):
- OS: Ubuntu
- OS version: 24.04.2
- GHC: ghc-9.12
Additional context
Since the bug can't be reproduced locally, I think this failure relies on a very specific scheduling of the garbage collector. My reason for thinking is that references are only marked as forgotten while garbage collecting, after which a "reference never released" exception can be thrown.