[jira] [Created] (OAK-12400) Online revision cleanup can delete referenced segments on a store not written in compaction order (e.g. a promoted cold standby)

["Rishabh Daim (Jira)" <ji...@apache.org>]

Rishabh Daim created OAK-12400:
----------------------------------

             Summary: Online revision cleanup can delete referenced segments on a store not written in compaction order (e.g. a promoted cold standby)
                 Key: OAK-12400
                 URL: https://issues.apache.org/jira/browse/OAK-12400
             Project: Jackrabbit Oak
          Issue Type: Bug
          Components: segment-tar
            Reporter: Rishabh Daim


{{DefaultCleanupContext.isDanglingFutureSegment}} detects unused (aborted-compaction) segments by
scan position: during cleanup, TAR entries are marked newest-first and every compacted segment is
treated as reclaimable until the last compacted root recorded in {{gc.log}} is reached (the
{{aheadOfRoot}} flag). This assumes segments are physically laid out in compaction write order
(the compacted root written last).

That assumption does not hold for a segment store assembled out of compaction order — e.g. a cold
standby, whose segments are written via {{FileStore.writeSegment}} in sync-arrival order. There the
recorded compacted root is not the newest segment, so the reverse scan reaches live, head-referenced
compacted segments while {{aheadOfRoot}} is still true and marks them reclaimable. Online revision
cleanup then removes referenced segments, and a subsequent read/compaction fails with
{{SegmentNotFoundException}}.

h4. Steps to reproduce
# Promote a cold standby to primary (its segment store was assembled via standby sync).
# Run online (full) revision cleanup.

Cleanup removes still-referenced compacted segments; the following compaction throws
{{SegmentNotFoundException}} while traversing the head. {{oak-run check}} passes immediately before
cleanup.

h4. Suggested fix
Make the dangling-segment test generation-based rather than positional: treat a compacted segment as
an unused future segment only if its generation is strictly newer than the head generation. This is
independent of physical segment write order and consistent for all store types.




--
This message was sent by Atlassian Jira
(v8.20.10#820010)