[jira] [Commented] (BEAM-2439) Datastore writer can fail to progress if Datastore is slow

["ASF GitHub Bot (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/BEAM-2439?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16091737#comment-16091737 ] 

ASF GitHub Bot commented on BEAM-2439:
--------------------------------------

GitHub user cph6 opened a pull request:

    https://github.com/apache/beam/pull/3585

    [BEAM-2439] Dynamic sizing of Datastore write RPCs.

    This implements the same behaviour recently added to Java SDK:
    - start at 200 entities per RPC;
    - size subsequent requests based on observed latency of previous requests.
    Includes a MovingSum class to track recent latency.
    Report RPC success & failure counts as metrics (again, as in the Java SDK).
    
    
    R: @vikkyrk
    R: @ssisk


You can merge this pull request into a Git repository by running:

    $ git pull https://github.com/cph6/beam datastore_batching_py

Alternatively you can review and apply these changes as the patch at:

    https://github.com/apache/beam/pull/3585.patch

To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:

    This closes #3585
    
----

----


> Datastore writer can fail to progress if Datastore is slow
> ----------------------------------------------------------
>
>                 Key: BEAM-2439
>                 URL: https://issues.apache.org/jira/browse/BEAM-2439
>             Project: Beam
>          Issue Type: Bug
>          Components: sdk-java-gcp
>            Reporter: Colin Phipps
>            Assignee: Colin Phipps
>            Priority: Minor
>              Labels: datastore
>             Fix For: 2.1.0
>
>
> When writing to Datastore, Beam groups writes into large batches (usually 500 entities per write, the maximum permitted by the API). If these writes are slow to commit on the serving side, the request may time out before all of the entities are written.
> When this happens, it loses any progress that has been made on those entities (the connector uses non-transactional writes, so some entities might have been written, but partial results are not returned to the connector so it has to assume that all entities need rewriting). It will retry the write with the same set of entities, which may time out in the same way repeatedly. This can be influenced by factors on the Datastore serving side, some of which are transient (hotspots) but some of which are not.
> We (Datastore) are developing a fix for this.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)