[
https://issues.apache.org/jira/browse/FLINK-9897?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16554727#comment-16554727
]
ASF GitHub Bot commented on FLINK-9897:
---------------------------------------
GitHub user glaksh100 opened a pull request:
https://github.com/apache/flink/pull/6408
[FLINK-9897] Make adaptive reads depend on run loop time instead of
fetchintervalmillis
## What is the purpose of the change
[FLINK-9692](https://github.com/apache/flink/pull/6300) introduced the
feature of adapting `maxNumberOfRecordsPerFetch` based on the average size of
Kinesis records. The PR assumed a maximum of `1/fetchIntervalMillis`
reads/second. However, in the case that the run loop of the `ShardConsumer`
takes more than `fetchIntervalMillis` to process records, the
`maxNumberOfRecordsPerFetch` is still sub-optimal. The purpose of this change
is to make the adaptive reads more efficient by using the actual run loop
frequency to determine the number of reads/second and
`maxNumberOfRecordsPerFetch`. The change also re-factors the run loop to be
more modular.
## Brief change log
- `processingStartTimeNanos` records start time of loop
- `processingEndTimeNanos` records end time of loop
- `adjustRunLoopFrequency()` adjusts end time depending on
`sleepTimeMillis` (if any).
- `runLoopTimeNanos` records actual run loop time.
- `adaptRecordsToRead` calculates `maxNumberOfRecordsPerFetch` based on
`runLoopTimeNanos`
- Unused method `getAdaptiveMaxRecordsPerFetch` is removed.
## Verifying this change
This change is already covered by existing tests, such as
`ShardConsumerTest`
This has also been tested against a stream with the following configuration
```
Number of Shards: 512
Parallelism: 128
```
## Does this pull request potentially affect one of the following parts:
- Dependencies (does it add or upgrade a dependency): (yes / **no**)
- The public API, i.e., is any changed class annotated with
`@Public(Evolving)`: (yes / **no**)
- The serializers: (yes / **no** / don't know)
- The runtime per-record code paths (performance sensitive): (yes /
**no** / don't know)
- Anything that affects deployment or recovery: JobManager (and its
components), Checkpointing, Yarn/Mesos, ZooKeeper: (yes / **no** / don't know)
- The S3 file system connector: (yes / **no** / don't know)
## Documentation
- Does this pull request introduce a new feature? (yes / **no**)
- If yes, how is the feature documented? (**not applicable** / docs /
JavaDocs / not documented)
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/lyft/flink FLINK-9897.AdaptiveReadsRunLoop
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/flink/pull/6408.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 #6408
----
commit 786556b9a9a509051a14772fbbd282db73e65252
Author: Lakshmi Gururaja Rao <glaksh100@...>
Date: 2018-07-24T18:44:08Z
[FLINK-9897] Make adaptive reads depend on run loop time instead of fetch
interval millis
----
> Further enhance adaptiveReads in Kinesis Connector to read more records in
> the case of long running loops
> ---------------------------------------------------------------------------------------------------------
>
> Key: FLINK-9897
> URL: https://issues.apache.org/jira/browse/FLINK-9897
> Project: Flink
> Issue Type: Improvement
> Components: Kinesis Connector
> Reporter: Lakshmi Rao
> Priority: Major
> Labels: pull-request-available
>
> In FLINK-9692, we introduced the ability for the shardConsumer to adaptively
> read more records based on the current average record size to optimize the 2
> Mb/sec shard limit. The feature maximizes maxNumberOfRecordsPerFetch of 5
> reads/sec (as prescribed by Kinesis limits). In the case where applications
> take more time to process records in the run loop, they are no longer able to
> read at a frequency of 5 reads/sec (even though their fetchIntervalMillis
> maybe set to 200 ms). In such a scenario, the maxNumberOfRecordsPerFetch
> should be calculated based on the time that the run loop actually takes as
> opposed to fetchIntervalMillis.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
