flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Till Rohrmann (JIRA)" <j...@apache.org>
Subject [jira] [Created] (FLINK-3885) Revisit parallelism logic
Date Mon, 09 May 2016 10:22:12 GMT
Till Rohrmann created FLINK-3885:
------------------------------------

             Summary: Revisit parallelism logic
                 Key: FLINK-3885
                 URL: https://issues.apache.org/jira/browse/FLINK-3885
             Project: Flink
          Issue Type: Bug
          Components: Streaming
    Affects Versions: 1.1.0
            Reporter: Till Rohrmann


Flink has multiple ways to specify the parallelism of programs. One can set it directly at
the operator or job-wide via the {{ExecutionConfig}}. Then it is also possible to define a
default parallelism in the {{flink-conf.yaml}}. This default parallelism is only respected
on the client side.

For batch jobs, the default parallelism is consistently set in the {{Optimizer}}. However,
for streaming jobs the default parallelism is only set iff the {{StreamExecutionEnvironment}}
has been created via {{StreamExecutionEnvironment.getExecutionEnvironment()}}. If one creates
a {{RemoteStreamEnvironment}}, then the default parallelism won't be respected. Instead the
parallelism is {{-1}}. Also the {{JobManager}} does not respect the default parallelism. But
whenever it sees a parallelism of {{-1}} it sets it to {{1}}. This behaviour is not consistent
with the batch behaviour and the behaviour described in the documentation of {{parallelism.default}}.

On top of that we also have the {{PARALLELISM_AUTO_MAX}} value for the parallelism. This value
tells the system to use all available slots. However, this feature is nowhere documented in
our docs.

I would propose to get rid of {{PARALLELISM_AUTO_MAX}} and to make the default parallelism
behaviour in the batch and streaming API consistent.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message