[
https://issues.apache.org/jira/browse/HBASE-25108?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17837046#comment-17837046
]
Istvan Toth edited comment on HBASE-25108 at 4/15/24 4:00 AM:
--------------------------------------------------------------
Maybe we could add an option to auto-disable retries for non-idempotent
operations ?
Only increment and checkAndMutate variants come to my mind, though anything
implemented in coprocessors could also be non-idempotent.
was (Author: stoty):
Maybe we could add an option to auto-disable retries for non-idempotent
operations ?
Only increment and checkAndPut come to my mind, though anything implemented in
coprocessors could also be non-idempotent.
> checkAndPut (or checkAndMutate) might return false when the row is mutated
> successfully
> ---------------------------------------------------------------------------------------
>
> Key: HBASE-25108
> URL: https://issues.apache.org/jira/browse/HBASE-25108
> Project: HBase
> Issue Type: Bug
> Affects Versions: 2.1.0, 1.2.11
> Reporter: Murilo Giacometti Rocha
> Priority: Major
>
> In the client, when the MutateRequest times out, we retry the operation in
> org.apache.hadoop.hbase.client.RpcRetryingCallerImpl#callWithRetries. If the
> server received the request but the client failed to get a response, the
> server returns processed=false because the value is already there. So the
> value false is returned, even though the checkAndPut was successful in the
> first attempt. It should return true if processed=false because it's already
> been processed AND it is a retry operation.
>
> Example RpcRetryingCallerImpl inside checkAndPut:
> request
> client -------{-}{{-}}o{{-}}{-}--------> server
>
> client --------------------> server (processing)
>
> client --------------------> server (processed)
> (timed out)
>
> retry request
> client -------{-}{{-}}o{{-}}{-}--------> server
>
> client --------------------> server (processing)
>
> client ------------------> server (already processed)
>
> response (processed = false)
> client <-----{-}{{-}}o{{-}}{-}------- server
>
> checkAndPut returns false, even though it's successful.
>
> In 2.1.0, I could only reproduce it three times by accident, but it always
> happened in 1.2.11. In 2.1.0, I could only reproduce it systematically by
> cleaning the response with the debugger before it got to the hconnection
> thread.
> Repro steps
> * Create a breakpoint in the exception in
> org.apache.hadoop.hbase.client.RpcRetryingCallerImpl#callWithRetries to make
> sure we get the exception and retry.
> * Create an infinite loop to create different rows with checkAndPut.
> * Start running with the disabled breakpoints.
> * Enable the breakpoint.
> * Pause all threads and verify that we are waiting for a response in the IPC
> thread. Wait for 1-2 minutes. This will cause a timeout.
> * Continue and verify that an exception is triggered.
> * Add a breakpoint to verify the response.
> * Continue and check the response and the returned value.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
