phoenix-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <>
Subject [jira] [Commented] (PHOENIX-4130) Avoid server retries for mutable indexes
Date Fri, 26 Jan 2018 21:01:00 GMT


ASF GitHub Bot commented on PHOENIX-4130:

Github user JamesRTaylor commented on a diff in the pull request:
    --- Diff: phoenix-core/src/main/java/org/apache/phoenix/hbase/index/write/
    @@ -70,13 +70,13 @@
        public static final String HTABLE_KEEP_ALIVE_KEY = "hbase.htable.threads.keepalivetime";
        public static final String INDEX_WRITER_RPC_RETRIES_NUMBER = "phoenix.index.writes.rpc.retries.number";
    -   /**
    -    * Based on the logic in HBase's AsyncProcess, a default of 11 retries with a pause
of 100ms
    -    * approximates 48 sec total retry time (factoring in backoffs).  The total time should
be less
    -    * than HBase's rpc timeout (default of 60 sec) or else the client will retry before
    -    * the response
    -    */
    -   public static final int DEFAULT_INDEX_WRITER_RPC_RETRIES_NUMBER = 11;
    +    /**
    +     * Retry server-server index write rpc only once, and let the client retry the data
    +     * instead to avoid typing up the handler
    +     */
    +   // note in HBase 2+, numTries = numRetries + 1
    +   // in prior versions, numTries = numRetries
    +   public static final int DEFAULT_INDEX_WRITER_RPC_RETRIES_NUMBER = 1;
    --- End diff --
    I was thinking it'd be ok if the onus was on the operator. I think we're in the minority
in having a mix of old/new clients with zero downtime. If you think it's worth checking the
client version (or checking if upgraded yet) and can come up with a reasonable way of doing
that, that's fine too. Not sure it's worth the effort, though. Here's some ideas:
    * add client version to IndexMaintainer (which is protobuf-ed, so it'd be doable). Then
use the "correct" IndexBuilder (or config?) based on the client version pulled out of the
IndexMaintainer. If you have separate IndexBuilder instances, then you'd get duplicate thread
pools to do the writing too which isn't ideal. Maybe within the parallel writer you could
use the correct config?
    * less dynamic, but check the timestamp of system.catalog to know what version you're
at. Somewhat scary to have an RPC in the start path of the coprocessor, though, as if system.catalog
RS can't be reached, then you're kind of screwed (but many you are already anyway).
    * least dynamic - put onus on operator and change to setIfUnset.

> Avoid server retries for mutable indexes
> ----------------------------------------
>                 Key: PHOENIX-4130
>                 URL:
>             Project: Phoenix
>          Issue Type: Improvement
>            Reporter: Lars Hofhansl
>            Assignee: Vincent Poon
>            Priority: Major
>             Fix For: 4.14.0
>         Attachments: PHOENIX-4130.v1.master.patch, PHOENIX-4130.v2.master.patch, PHOENIX-4130.v3.master.patch
> Had some discussions with [~jamestaylor], [~samarthjain], and [~vincentpoon], during
which I suggested that we can possibly eliminate retry loops happening at the server that
cause the handler threads to be stuck potentially for quite a while (at least multiple seconds
to ride over common scenarios like splits).
> Instead we can do the retries at the Phoenix client that.
> So:
> # The index updates are not retried on the server. (retries = 0)
> # A failed index update would set the failed index timestamp but leave the index enabled.
> # Now the handler thread is done, it throws an appropriate exception back to the client.
> # The Phoenix client can now retry. When those retries fail the index is disabled (if
the policy dictates that) and throw the exception back to its caller.
> So no more waiting is needed on the server, handler threads are freed immediately.

This message was sent by Atlassian JIRA

View raw message