[
https://issues.apache.org/jira/browse/HUDI-494?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17101055#comment-17101055
]
Yanjia Gary Li edited comment on HUDI-494 at 5/8/20, 1:38 AM:
--------------------------------------------------------------
-Ok, I see what happened here. Root cause is
[https://github.com/apache/incubator-hudi/blob/master/hudi-client/src/main/java/org/apache/hudi/index/bloom/HoodieBloomIndex.java#L214]-
So basically commit 1 wrote a very small file(let's say 200 records) to a new
partition day=05. And then when commit 2 was trying to write, it looks back to
commit 1 to get an estimated size of every record, but because commit 1 has too
little records so it's inaccurate and way too big. So Hudi will calculate
record/file using the big record size number and get a very small record/file.
This lead to many small files.
was (Author: garyli1019):
Ok, I see what happened here. Root cause is
[https://github.com/apache/incubator-hudi/blob/master/hudi-client/src/main/java/org/apache/hudi/index/bloom/HoodieBloomIndex.java#L214]
So basically commit 1 wrote a very small file(let's say 200 records) to a new
partition day=05. And then when commit 2 trying to write to day=05, it will
look up the affected partition and use the Bloom index range from the existing
files, so it will use 200 here. Commit 2 has much more records than 200, so it
will create tons of files since the Bloom index range is too small.
I am not really familiar with the indexing part of the code. Please let me know
if I understand this correctly and we can figure out a fix. [~lamber-ken]
[~vinoth]
> [DEBUGGING] Huge amount of tasks when writing files into HDFS
> -------------------------------------------------------------
>
> Key: HUDI-494
> URL: https://issues.apache.org/jira/browse/HUDI-494
> Project: Apache Hudi (incubating)
> Issue Type: Test
> Reporter: Yanjia Gary Li
> Assignee: Yanjia Gary Li
> Priority: Major
> Attachments: Screen Shot 2020-01-02 at 8.53.24 PM.png, Screen Shot
> 2020-01-02 at 8.53.44 PM.png, example2_hdfs.png, example2_sparkui.png,
> image-2020-01-05-07-30-53-567.png
>
>
> I am using the manual build master after
> [https://github.com/apache/incubator-hudi/commit/36b3b6f5dd913d3f1c9aa116aff8daf6540fed65]
> commit. EDIT: tried with the latest master but got the same result
> I am seeing 3 million tasks when the Hudi Spark job writing the files into
> HDFS. It seems like related to the input size. With 7.7 GB input it was 3.2
> million tasks, with 9 GB input it was 3.7 million. Both with 10 parallelisms.
> I am seeing a huge amount of 0 byte files being written into .hoodie/.temp/
> folder in my HDFS. In the Spark UI, each task only writes less than 10
> records in
> {code:java}
> count at HoodieSparkSqlWriter{code}
> All the stages before this seem normal. Any idea what happened here? My
> first guess would be something related to the bloom filter index. Maybe
> somewhere trigger the repartitioning with the bloom filter index? But I am
> not really familiar with that part of the code.
> Thanks
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
