[
https://issues.apache.org/jira/browse/KAFKA-5330?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16033578#comment-16033578
]
ASF GitHub Bot commented on KAFKA-5330:
---------------------------------------
GitHub user tbcdns opened a pull request:
https://github.com/apache/kafka/pull/3196
KAFKA-5330: Use per-task converters
Instead of sharing the same converter instance within the worker, use a
converter per task.
More details:
- https://github.com/confluentinc/schema-registry/issues/514
- https://issues.apache.org/jira/browse/KAFKA-5330
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/tbcdns/kafka KAFKA-5330
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/kafka/pull/3196.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 #3196
----
commit 43f38a5a4e0a7eeff233afe90a7893f252c6edac
Author: Thibaud Chardonnens <thibaud.chardonn...@swisscom.com>
Date: 2017-06-01T19:38:16Z
KAFKA-5330: Use per-task converters
----
> Use per-task converters in Connect
> ----------------------------------
>
> Key: KAFKA-5330
> URL: https://issues.apache.org/jira/browse/KAFKA-5330
> Project: Kafka
> Issue Type: Improvement
> Components: KafkaConnect
> Affects Versions: 0.11.0.0
> Reporter: Ewen Cheslack-Postava
> Original Estimate: 24h
> Remaining Estimate: 24h
>
> Because Connect started with a worker-wide model of data formats, we
> currently allocate a single Converter per worker and only allocate an
> independent one when the user overrides the converter.
> This can lead to performance problems when the worker-level default converter
> is used by a large number of tasks because converters need to be threadsafe
> to support this model and they may spend a lot of time just on
> synchronization.
> We could, instead, simply allocate one converter per task. There is some
> overhead involved, but generally it shouldn't be that large. For example,
> Confluent's Avro converters will each have their own schema cache and have to
> make their on calls to the schema registry API, but these are relatively
> small, likely inconsequential compared to any normal overhead we would
> already have for creating and managing each task.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
