GitHub user xuchuanyin opened a pull request:
https://github.com/apache/carbondata/pull/1261
[CARBONDATA-1373] Enhance update performance by increasing parallelism
# Scenario
Recently I have tested the update feature provided in Carbondata and found
its poor performance.
I had a table containing about 14 million records with about 370 columns(no
dictionary columns) and the data files are about 3.8 GB in total. All the data
files were in one segment.
I performed an update SQL which update a column for all the records and the
SQL looked like `UPDATE myTable SET (col1)=(col1+1000) WHERE TRUE`. In my
environment, the update job failed with 'executor lost errors'. And I found
'spill data' related messages in the container logs.
# Analyze
I've read about the implementation of update-delete in Carbondata in
ISSUE#440. The update consists a delete and an insert operation. And the error
occurred during the insert operation.
After studying the code, I have found that while doing inserting, the
updated records are grouped by the `segmentId`, which means all the recoreds in
one segment will be processed in only one task, thus will cause task failure
when the amount of input data is quite large.
# Solution
We should improve the parallelism when doing update for a segment.
I append a random key to the `segmentId` to increase the partition number
before doing the insertion stage and then remove the suffix when doing the real
insertion.
# Modification
+ Increase parallelism while processing one segment in update, this is
achieved by distributing records to different partitions by using a customized
partitioner.
+ Add a property to configure the parallelism.
+ Clean up local files after update (previous bugs)
+ Remove useless imports
+ Add tests
+ Add related documents
# Notes
I have tested in my example and the job finished in about 13 minutes
successfully. The records were updated as expected.
Comparing to the previous implementation, the update performance has been
enhanced:
Origin(Parallelism(1) + GroupBy): Update **FAILED**
Adding Parallelism(6) + GroupBy: Update **SUCCESSFULLY** using **13mins**
Parallelism(1) + PartitionBy: Update **SUCCESSFULLY** using **21mins**
Adding Parallelism(6) + PartitionBy: Update **SUCCESSFULLY** using **5mins**
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/xuchuanyin/carbondata enhance_update_perf
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/carbondata/pull/1261.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 #1261
----
commit ebfe1ca2a125c0b736e917d8a7956f5e39dedc50
Author: xuchuanyin <[email protected]>
Date: 2017-08-11T15:00:20Z
Enhance update performance by increasing parallelism
+ Increase parallelism while processing one segment in update
+ Use partitionBy instead of groupby
+ Add a property to configure the parallelism
+ Clean up local files after update (previous bugs)
+ Remove useless imports
----
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
