[jira] [Commented] (SOLR-8687) Race condition with RTGs during soft commit
[
https://issues.apache.org/jira/browse/SOLR-8687?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15153422#comment-15153422
]
Ishan Chattopadhyaya commented on SOLR-8687:
bq. it's also disconcerting if it's not air-tight and our tests don't catch it
With [~sar...@syr.edu]'s help, I stress tested 1000 rounds (16 at a time) of
the StressReorderTest and it didn't fail.
However, the above mentioned test, which is similar to StressReorderTest, but
on a 3 node cluster instead of a simulated replica, failed around 10 times
(each with this exact failure). Also, I had increased the number of read
operations within each test from 50k to 200k. At this time, I am reasonably
sure that the test had nothing to do with my other changes. Next up, I shall
isolate the test from the other changes and try to run it on a fresh master so
as to be sure I can reproduce.
> Race condition with RTGs during soft commit
> ---
>
> Key: SOLR-8687
> URL: https://issues.apache.org/jira/browse/SOLR-8687
> Project: Solr
> Issue Type: Bug
>Reporter: Ishan Chattopadhyaya
>
> I am facing a problem with stress testing SOLR-5944, even though I think this
> problem persists in Solr even without my changes.
> The symptom is that during a stress test (similar to TestStressReorder), RTG
> gets a document which is older version than that of the last acknowledged
> write.
> Possible reason:
> {code}
> (DUH2's commit())
> ...
> 1: if (cmd.softCommit) {
> 2:// ulog.preSoftCommit();
> 3:synchronized (solrCoreState.getUpdateLock()) {
> 4: if (ulog != null) ulog.preSoftCommit(cmd);
> 5: core.getSearcher(true, false, waitSearcher, true);
> 6: if (ulog != null) ulog.postSoftCommit(cmd);
> 7:}
> 8:callPostSoftCommitCallbacks();
> 9: }
> ...
> {code}
> * Before line 1, there was an update (say id=2) which was in ulog's map. Maps
> are, say, map=\{2=LogPtr(1234)\} , prevMap=\{...\} , prevMap2=\{...\}
> * Due to line 4 (ulog.preSoftCommit()), the maps were rotated. Now, the id=2
> is in prevMap: map={}, prevMap=\{2=LogPtr(1234)\}, prevMap2=\{...\} . Till
> now RTG for id=2 will work.
> * Due to line 5, a new searcher is due to be opened. But this is
> asynchronous, and lets assume this doesn't complete before few more lines are
> executed.
> * Due to line 6 (ulog.postSoftCommit()), the previous maps are cleared out.
> Now the maps are: map={}, prevMap=null, prevMap2=null
> * If there's an RTG for id=2, it will not work from the ulog's maps, so it
> will fall through to be searched using the last searcher. But, the searcher
> due to be opened in line 5 hasn't yet been opened. In this case, the returned
> document will be whatever version of id=2 that was present in the previous
> searcher.
> Can someone please confirm if this is a potential problem? If so, any
> suggestions for a fix, please? I tried opening a ulog.openRealtimeSearcher()
> in the above synchronized block, but the problem still persists, but I
> haven't looked into why that could be.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
-
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (SOLR-8687) Race condition with RTGs during soft commit
[
https://issues.apache.org/jira/browse/SOLR-8687?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15153413#comment-15153413
]
Yonik Seeley commented on SOLR-8687:
Hmmm, RTG should be air-tight (I always hate changes to synchronization in this
area because it's hard to re-validate). We have stress tests for this, so it's
also disconcerting if it's not air-tight and our tests don't catch it.
I'll review the related code again...
> Race condition with RTGs during soft commit
> ---
>
> Key: SOLR-8687
> URL: https://issues.apache.org/jira/browse/SOLR-8687
> Project: Solr
> Issue Type: Bug
>Reporter: Ishan Chattopadhyaya
>
> I am facing a problem with stress testing SOLR-5944, even though I think this
> problem persists in Solr even without my changes.
> The symptom is that during a stress test (similar to TestStressReorder), RTG
> gets a document which is older version than that of the last acknowledged
> write.
> Possible reason:
> {code}
> (DUH2's commit())
> ...
> 1: if (cmd.softCommit) {
> 2:// ulog.preSoftCommit();
> 3:synchronized (solrCoreState.getUpdateLock()) {
> 4: if (ulog != null) ulog.preSoftCommit(cmd);
> 5: core.getSearcher(true, false, waitSearcher, true);
> 6: if (ulog != null) ulog.postSoftCommit(cmd);
> 7:}
> 8:callPostSoftCommitCallbacks();
> 9: }
> ...
> {code}
> * Before line 1, there was an update (say id=2) which was in ulog's map. Maps
> are, say, map=\{2=LogPtr(1234)\} , prevMap=\{...\} , prevMap2=\{...\}
> * Due to line 4 (ulog.preSoftCommit()), the maps were rotated. Now, the id=2
> is in prevMap: map={}, prevMap=\{2=LogPtr(1234)\}, prevMap2=\{...\} . Till
> now RTG for id=2 will work.
> * Due to line 5, a new searcher is due to be opened. But this is
> asynchronous, and lets assume this doesn't complete before few more lines are
> executed.
> * Due to line 6 (ulog.postSoftCommit()), the previous maps are cleared out.
> Now the maps are: map={}, prevMap=null, prevMap2=null
> * If there's an RTG for id=2, it will not work from the ulog's maps, so it
> will fall through to be searched using the last searcher. But, the searcher
> due to be opened in line 5 hasn't yet been opened. In this case, the returned
> document will be whatever version of id=2 that was present in the previous
> searcher.
> Can someone please confirm if this is a potential problem? If so, any
> suggestions for a fix, please? I tried opening a ulog.openRealtimeSearcher()
> in the above synchronized block, but the problem still persists, but I
> haven't looked into why that could be.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
-
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (SOLR-8687) Race condition with RTGs during soft commit
[
https://issues.apache.org/jira/browse/SOLR-8687?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15153329#comment-15153329
]
Shalin Shekhar Mangar commented on SOLR-8687:
-
This does look like a genuine bug, Ishan. We wait for the new searcher outside
of the synchronized block and so as you said, a concurrent RTG request can miss
updates. Perhaps [~ysee...@gmail.com] has a suggestion for us?
> Race condition with RTGs during soft commit
> ---
>
> Key: SOLR-8687
> URL: https://issues.apache.org/jira/browse/SOLR-8687
> Project: Solr
> Issue Type: Bug
>Reporter: Ishan Chattopadhyaya
>
> I am facing a problem with stress testing SOLR-5944, even though I think this
> problem persists in Solr even without my changes.
> The symptom is that during a stress test (similar to TestStressReorder), RTG
> gets a document which is older version than that of the last acknowledged
> write.
> Possible reason:
> {code}
> (DUH2's commit())
> ...
> 1: if (cmd.softCommit) {
> 2:// ulog.preSoftCommit();
> 3:synchronized (solrCoreState.getUpdateLock()) {
> 4: if (ulog != null) ulog.preSoftCommit(cmd);
> 5: core.getSearcher(true, false, waitSearcher, true);
> 6: if (ulog != null) ulog.postSoftCommit(cmd);
> 7:}
> 8:callPostSoftCommitCallbacks();
> 9: }
> ...
> {code}
> * Before line 1, there was an update (say id=2) which was in ulog's map. Maps
> are, say, map=\{2=LogPtr(1234)\} , prevMap=\{...\} , prevMap2=\{...\}
> * Due to line 4 (ulog.preSoftCommit()), the maps were rotated. Now, the id=2
> is in prevMap: map={}, prevMap=\{2=LogPtr(1234)\}, prevMap2=\{...\} . Till
> now RTG for id=2 will work.
> * Due to line 5, a new searcher is due to be opened. But this is
> asynchronous, and lets assume this doesn't complete before few more lines are
> executed.
> * Due to line 6 (ulog.postSoftCommit()), the previous maps are cleared out.
> Now the maps are: map={}, prevMap=null, prevMap2=null
> * If there's an RTG for id=2, it will not work from the ulog's maps, so it
> will fall through to be searched using the last searcher. But, the searcher
> due to be opened in line 5 hasn't yet been opened. In this case, the returned
> document will be whatever version of id=2 that was present in the previous
> searcher.
> Can someone please confirm if this is a potential problem? If so, any
> suggestions for a fix, please? I tried opening a ulog.openRealtimeSearcher()
> in the above synchronized block, but the problem still persists, but I
> haven't looked into why that could be.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
-
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (SOLR-8687) Race condition with RTGs during soft commit
[
https://issues.apache.org/jira/browse/SOLR-8687?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15151059#comment-15151059
]
Ishan Chattopadhyaya commented on SOLR-8687:
My stress test is here:
https://github.com/chatman/lucene-solr/blob/627b9ac9b46796f20be78b04ebbdfa4299b96ab7/solr/core/src/test/org/apache/solr/cloud/TestStressInPlaceUpdates.java
> Race condition with RTGs during soft commit
> ---
>
> Key: SOLR-8687
> URL: https://issues.apache.org/jira/browse/SOLR-8687
> Project: Solr
> Issue Type: Bug
>Reporter: Ishan Chattopadhyaya
>
> I am facing a problem with stress testing SOLR-5944, even though I think this
> problem persists in Solr even without my changes.
> The symptom is that during a stress test (similar to TestStressReorder), RTG
> gets a document which is older version than that of the last acknowledged
> write.
> Possible reason:
> {code}
> (DUH2's commit())
> ...
> 1: if (cmd.softCommit) {
> 2:// ulog.preSoftCommit();
> 3:synchronized (solrCoreState.getUpdateLock()) {
> 4: if (ulog != null) ulog.preSoftCommit(cmd);
> 5: core.getSearcher(true, false, waitSearcher, true);
> 6: if (ulog != null) ulog.postSoftCommit(cmd);
> 7:}
> 8:callPostSoftCommitCallbacks();
> 9: }
> ...
> {code}
> * Before line 1, there was an update (say id=2) which was in ulog's map. Maps
> are, say, `map={2=LogPtr(1234)}, prevMap={...}, prevMap2={...}`
> * Due to line 4 (ulog.preSoftCommit()), the maps were rotated. Now, the id=2
> is in prevMap: `map={}, prevMap={2=LogPtr(1234)}, prevMap2={...}`. Till now
> RTG for id=2 will work.
> * Due to line 5, a new searcher is due to be opened. But this is
> asynchronous, and lets assume this doesn't complete before few more lines are
> executed.
> * Due to line 6 (ulog.postSoftCommit()), the previous maps are cleared out.
> Now the maps are: `map={}, prevMap=null, prevMap2=null`
> * If there's an RTG for id=2, it will not work from the ulog's maps, so it
> will fall through to be searched using the last searcher. But, the searcher
> due to be opened in line 5 hasn't yet been opened. In this case, the returned
> document will be whatever version of id=2 that was present in the previous
> searcher.
> Can someone please confirm if this is a potential problem? If so, any
> suggestions for a fix, please? I tried opening a ulog.openRealtimeSearcher()
> in the above synchronized block, but the problem still persists, but I
> haven't looked into why that could be.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
-
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
