[jira] [Created] (OAK-12404) CompactionWriter.flush() persists partial writer before target writer

["Julian Sedding (Jira)" <ji...@apache.org>]

Julian Sedding created OAK-12404:
------------------------------------

             Summary: CompactionWriter.flush() persists partial writer before target writer
                 Key: OAK-12404
                 URL: https://issues.apache.org/jira/browse/OAK-12404
             Project: Jackrabbit Oak
          Issue Type: Bug
          Components: segment-tar
    Affects Versions: 2.6.0
            Reporter: Julian Sedding


A partially compacted node references the fully compacted nodes below it, so this persists a segment before the segment it depends on. Two consequences:

* The head's segment becomes durable before the records it references. The journal is written after both flushes (TarRevisions.doFlush), so a plain crash is not corrupting, but the ordering invariant is inverted.
* {{DefaultCleanupContext.isDanglingFutureSegment}} assumes compacted segments are unused iff they are persisted after the last compacted root, and reclaims them without checking reachability. After a soft-cancelled compaction the compacted root is the partial root, so the target segment flushed afterwards is reclaimed while still reachable from the head.

Only observable with soft cancellation; otherwise the partial writer has nothing buffered. {{AbstractGarbageCollectionStrategy}} currently supplies no-op soft cancellers, so this is not reachable from production GC today, but must be fixed before partial compaction is enabled.

Reproduction: {{CompactionWriterFlushOrderTest}} (see linked PR) drives a {{ClassicCompactor}} with a soft canceller and asserts that the segment holding the compacted root references no segment persisted after it. Fails on trunk with:

{noformat}
the segment holding the compacted root 2156cdbf-... (GCGeneration{generation=0, fullGeneration=0, isCompacted=true}) references 1 segment(s) that are only persisted after it: [5fc4f9d2-... GCGeneration{generation=1, fullGeneration=1, isCompacted=true}]
{noformat}

Fix: flush in reference order.

{code:java}
public void flush() throws IOException {
    targetWriter.flush();
    partialWriter.flush();
}
{code}




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