[
https://issues.apache.org/jira/browse/HBASE-26026?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17370389#comment-17370389
]
chenglei edited comment on HBASE-26026 at 7/2/21, 9:32 AM:
-----------------------------------------------------------
[~zhangduo]，would please help me have a review at your convenience ? thanks.
was (Author: comnetwork):
[~zhangduo]，would please help me have a review at your convenienc ? thanks.
> HBase Write may be stuck forever when using CompactingMemStore
> --------------------------------------------------------------
>
> Key: HBASE-26026
> URL: https://issues.apache.org/jira/browse/HBASE-26026
> Project: HBase
> Issue Type: Bug
> Components: in-memory-compaction
> Affects Versions: 3.0.0-alpha-1, 2.3.0, 2.4.0
> Reporter: chenglei
> Assignee: chenglei
> Priority: Critical
>
> Sometimes I observed that HBase Write might be stuck in my hbase cluster
> which enabling {{CompactingMemStore}}. I have simulated the problem by unit
> test in my PR.
> The problem is caused by {{CompactingMemStore.checkAndAddToActiveSize}} :
> {code:java}
> 425 private boolean checkAndAddToActiveSize(MutableSegment currActive, Cell
> cellToAdd,
> 426 MemStoreSizing memstoreSizing) {
> 427 if (shouldFlushInMemory(currActive, cellToAdd, memstoreSizing)) {
> 428 if (currActive.setInMemoryFlushed()) {
> 429 flushInMemory(currActive);
> 430 if (setInMemoryCompactionFlag()) {
> 431 // The thread is dispatched to do in-memory compaction in the
> background
> ......
> }
> {code}
> In line 427, if {{currActive.getDataSize}} adding the size of {{cellToAdd}}
> exceeds {{CompactingMemStore.inmemoryFlushSize}}, then {{currActive}} should
> be flushed, {{MutableSegment.setInMemoryFlushed()}} is invoked in above line
> 428 ：
> {code:java}
> public boolean setInMemoryFlushed() {
> return flushed.compareAndSet(false, true);
> }
> {code}
> After set {{currActive.flushed}} to true, in above line 429
> {{flushInMemory(currActive)}} invokes
> {{CompactingMemStore.pushActiveToPipeline}} :
> {code:java}
> protected void pushActiveToPipeline(MutableSegment currActive) {
> if (!currActive.isEmpty()) {
> pipeline.pushHead(currActive);
> resetActive();
> }
> }
> {code}
> In above {{CompactingMemStore.pushActiveToPipeline}} method , if the
> {{currActive.cellSet}} is empty, then nothing is done. Due to concurrent
> writes and because we first add cell size to {{currActive.getDataSize}} and
> then actually add cell to {{currActive.cellSet}}, it is possible that
> {{currActive.getDataSize}} could not accommodate {{cellToAdd}} but
> {{currActive.cellSet}} is still empty if pending writes which not yet add
> cells to {{currActive.cellSet}}.
> So if the {{currActive.cellSet}} is empty now, then no {{ActiveSegment}} is
> created, and new writes still continue target to {{currActive}}, but
> {{currActive.flushed}} is true, {{currActive}} could not enter
> {{flushInMemory(currActive)}} again，and new {{ActiveSegment}} could not be
> created forever ! In the end all writes would be stuck.
> In my opinion , once {{currActive.flushed}} is set true, it could not
> continue use as {{ActiveSegment}} , and because of concurrent pending writes,
> only after {{currActive.updatesLock.writeLock()}} is acquired(i.e.
> {{currActive.waitForUpdates}} is called) in
> {{CompactingMemStore.inMemoryCompaction}} ,we can safely say {{currActive}}
> is empty or not.
> My fix is remove the {{if (!currActive.isEmpty())}} check here and left the
> check to background {{InMemoryCompactionRunnable}} after
> {{currActive.waitForUpdates}} is called. An alternative fix is we use
> synchronization mechanism in {{checkAndAddToActiveSize}} method to prevent
> all writes , wait for all pending write completed(i.e.
> currActive.waitForUpdates is called) and if {{currActive}} is still empty
> ,then we set {{currActive.flushed}} back to false,but I am not inclined to
> use so heavy synchronization in write path, and I think we would better
> maintain lockless implementation for {{CompactingMemStore.add}} method just
> as now and {{currActive.waitForUpdates}} would better be left in background
> {{InMemoryCompactionRunnable}}.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
