[jira] [Comment Edited] (HBASE-27519) Another case for FNFE on StoreFileScanner after a flush followed by a compaction

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

    [ https://issues.apache.org/jira/browse/HBASE-27519?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17645547#comment-17645547 ] 

chenglei edited comment on HBASE-27519 at 12/10/22 4:52 AM:
------------------------------------------------------------

Pushed to 2.4+, thanks [~wchevreuil] for reviewing!


was (Author: comnetwork):
pushed to 2.4+, thanks [~wchevreuil] for reviewing!

> Another case for FNFE on StoreFileScanner after a flush followed by a compaction
> --------------------------------------------------------------------------------
>
>                 Key: HBASE-27519
>                 URL: https://issues.apache.org/jira/browse/HBASE-27519
>             Project: HBase
>          Issue Type: Bug
>    Affects Versions: 2.6.0, 3.0.0-alpha-3, 2.4.15, 2.5.2
>            Reporter: chenglei
>            Assignee: chenglei
>            Priority: Major
>
> Besides the fix for {{HStore.getScanners}} for FNFE on StoreFileScanner after a flush followed by a compaction in HBASE-27484, there is a another case would cause FNFE:
> # When {{StoreScanner}} is scanning, there is a flush followed by a compaction. When the flushed is completed in {{HStore.completeFlush}} and after it adds the new {{HStoreFile}} in {{StoreEngine}}, it would notify the {{StoreScanner.updateReaders}} by {{HStore.notifyChangedReadersObservers}} with the new flushed {{HStoreFile}}：
> {code:java}
>     private boolean completeFlush(List<HStoreFile> sfs, long snapshotId) throws IOException {
>     // NOTE:we should keep clearSnapshot method inside the write lock because clearSnapshot may
>     // close {@link DefaultMemStore#snapshot}, which may be used by
>     // {@link DefaultMemStore#getScanners}.
>     storeEngine.addStoreFiles(sfs,
>       snapshotId > 0 ? () -> this.memstore.clearSnapshot(snapshotId) : () -> {
>       });
>     // notify to be called here - only in case of flushes
>     notifyChangedReadersObservers(sfs);
>     if (LOG.isTraceEnabled()) {
>       long totalSize = getTotalSize(sfs);
>       String traceMessage = "FLUSH time,count,size,store size,store files ["
>         + EnvironmentEdgeManager.currentTime() + "," + sfs.size() + "," + totalSize + ","
>         + storeSize + "," + storeEngine.getStoreFileManager().getStorefileCount() + "]";
>       LOG.trace(traceMessage);
>     }
>     return needsCompaction();
>   }
> {code}
> # But if before entering {{StoreScanner.updateReaders}} , the compaction task starts and  it compacts the new flushed {{HStoreFile}} and then {{HStore.closeAndArchiveCompactedFiles}} is called by {{CompactedHFilesDischarger}} to delete the {{HStoreFile}}. 
> # When {{StoreScanner.updateReaders}} continues to execute, it would read already deleted {{HStoreFile}} and throws FNFE.



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