[
https://issues.apache.org/jira/browse/PHOENIX-5080?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16729385#comment-16729385
]
Monani Mihir commented on PHOENIX-5080:
---------------------------------------
{quote} I think for rebuilds (code path following rebuildIndices()), maybe we
can skip updating the index state, and that way MetadataRegionObserver can
handle the state transitions.
{quote}
With checks present in MetaDataEngPointImpl, like 1) if INDEX_STATE has been
changed during rebuild 2) INDEX_STATE is INACTIVE to successfully transition
into ACTIVE etc, Once rebuild is finished index will again transit into
DISABLE/PENDING_DISABLE state via MetaDataRegionObserver and we would again run
partial rebuilder. This cycle could repeat so many times.
I think the more optimal solution would be that for each mutation until all
retires are exhausted (considering Partial Rebuilder as client) , we should not
attempt to update INDEX_STATE , from INACTIVE to PENDING_DISABLE.
{code:java}
PhoenixIndexFailurePolicy#doBatchWithRetries :-
while (canRetryMore(numRetry++, maxTries, canRetryUntil)) {
try {
Thread.sleep(ConnectionUtils.getPauseTime(pause, numRetry)); // HBase's
exponential backoff
mutateCommand.doMutation();
// success - change the index state from PENDING_DISABLE back to ACTIVE
handleIndexWriteSuccessFromClient(iwe, connection);
return;
...{code}
In above code block when we call mutateCommand.doMutation(), we should not
attempt to update INDEX_STATE when failure occurs. Instead , once all retries
are exhausted in while loop, we should DISABLE the index directly.
What do you think [~vincentpoon] ?
I don't have much idea about UpsertSelect logic and how it works on server side.
> Index becomes Active during Partial Index Rebuilder if Index Failure happens
> ----------------------------------------------------------------------------
>
> Key: PHOENIX-5080
> URL: https://issues.apache.org/jira/browse/PHOENIX-5080
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.14.1
> Reporter: Monani Mihir
> Priority: Blocker
>
> After PHOENIX-4130 and PHOENIX-4600 , If there is Index failure during
> Partial Index Rebuild, Rebuilder will try again to write Index updates. If it
> succeeds then it will transition Index from INACTIVE to ACTIVE, even before
> Rebuilder finishes.
> Here is where it goes wrong, I think :-
> {code:java}
> PhoenixIndexFailurePolicy.java :-
> public static void doBatchWithRetries(MutateCommand mutateCommand,
> IndexWriteException iwe, PhoenixConnection connection,
> ReadOnlyProps config) throws IOException {
> ....
> while (canRetryMore(numRetry++, maxTries, canRetryUntil)) {
> ...
> handleIndexWriteSuccessFromClient(iwe, connection);
> ...
> }
> }
> ....
> private static void handleIndexWriteSuccessFromClient(IndexWriteException
> indexWriteException, PhoenixConnection conn) {
> handleExceptionFromClient(indexWriteException, conn,
> PIndexState.ACTIVE);
> }
> {code}
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
