phoenix-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Vincent Poon (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (PHOENIX-5266) Client can only write on Index Table and skip data table if failure happens because of region split/move etc
Date Tue, 07 May 2019 23:09:00 GMT

    [ https://issues.apache.org/jira/browse/PHOENIX-5266?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16835189#comment-16835189
] 

Vincent Poon commented on PHOENIX-5266:
---------------------------------------

[~abhishek.chouhan] Looks like we are doing it for the first batch because given the way the
code is right now, that's the only batch we can be sure was written to the data table. One
option is to do what you said, and change the code so we only set the replay attribute on
the batch that failed the index write.  We also only need to continue from the batch that
failed, instead of redoing the entire batchList.

An alternative is to just remove the use of the REPLAY attribute there.  It's a small optimization
that probably isn't worth the code complexity and risk to correctness otherwise.

I'll leave it up to you, [~mihir6692].

> Client can only write on Index Table and skip data table if failure happens because of
region split/move etc
> ------------------------------------------------------------------------------------------------------------
>
>                 Key: PHOENIX-5266
>                 URL: https://issues.apache.org/jira/browse/PHOENIX-5266
>             Project: Phoenix
>          Issue Type: Bug
>    Affects Versions: 4.15.0, 4.14.1, 5.1.0, 4.14.2
>            Reporter: Mihir Monani
>            Assignee: Mihir Monani
>            Priority: Blocker
>             Fix For: 4.14.2
>
>         Attachments: PHOENIX-5266-4.x-HBase-1.3.01.patch, PHOENIX-5266.patch, PHOENIX-5266.patch
>
>
> With Phoenix 4.14.1 client, There is a scenario where client would skip data table write
but do successful index table write. In this case, we should treat it as Data loss scenario.
>  
> Relevant code path :-
> [https://github.com/apache/phoenix/blob/4.x-HBase-1.3/phoenix-core/src/main/java/org/apache/phoenix/execute/MutationState.java#L994-L1043]
> [https://github.com/apache/phoenix/blob/4.x-HBase-1.3/phoenix-core/src/main/java/org/apache/phoenix/execute/MutationState.java#L1089-L1109]
>  
> Here is what happens :-
>  * Consider below assumptions for scenario :- 
>  ** max no row in single batch = 100
>  ** max size of batch = 2 MB
>  * When client faces SQLException Code 1121, it sets variable shouldRetryIndexedMutation=true.
>  * In scenarios where client sends batch of 100 rows only as per configuration, but batch
size is >2 MB, MutationState.java#991 will split this 100 row batch into multiple smaller
batches which are <2MB.
>  ** MutationState.java#991 :- [https://github.com/apache/phoenix/blob/4.x-HBase-1.3/phoenix-core/src/main/java/org/apache/phoenix/execute/MutationState.java#L991]
>  * Suppose there are 5 batches of 20 rows but client faces 1121 SQLExceptionCode on 2nd
batch , then it will set shouldRetryIndexedMutation=true and it will retry all 5 batches again
with only Index updates. This will results in rows missing from Data table.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message