[jira] [Commented] (SPARK-24914) totalSize is not a good estimate for broadcast joins

["Igor Uchôa (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-24914?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17574409#comment-17574409 ] 

Igor Uchôa commented on SPARK-24914:
------------------------------------

Hi [~bersprockets],

I just wanted to know if you guys have any updates regarding this issue. Right now, I'm facing a very similar behavior; some Hive tables have a very small number, or even zero, for `rawDataSize`. For the ones which have zero value, I don't think they are problematic because Spark uses the totalSize, but the ones with a very small number (eg 192bytes of `rawDataSize` against 200Mb of `totalSize`), they are falling into BroadcastHashJoin and we are getting OOM issues. I thought of doing a similar solution to what you have proposed. The difference is that I'm not going to use a factor for each table, but a global one and a spark property indicating whether to use `rawDataSize` or not.

But before doing that, I would like to be most aligned as possible with the community, so any future changes will be easily migrated.

> totalSize is not a good estimate for broadcast joins
> ----------------------------------------------------
>
>                 Key: SPARK-24914
>                 URL: https://issues.apache.org/jira/browse/SPARK-24914
>             Project: Spark
>          Issue Type: Improvement
>          Components: SQL
>    Affects Versions: 3.1.0
>            Reporter: Bruce Robbins
>            Priority: Major
>
> When determining whether to do a broadcast join, Spark estimates the size of the smaller table as follows:
>  - if totalSize is defined and greater than 0, use it.
>  - else, if rawDataSize is defined and greater than 0, use it
>  - else, use spark.sql.defaultSizeInBytes (default: Long.MaxValue)
> Therefore, Spark prefers totalSize over rawDataSize.
> Unfortunately, totalSize is often quite a bit smaller than the actual table size, since it represents the size of the table's files on disk. Parquet and Orc files, for example, are encoded and compressed. This can result in the JVM throwing an OutOfMemoryError while Spark is loading the table into a HashedRelation, or when Spark actually attempts to broadcast the data.
> On the other hand, rawDataSize represents the uncompressed size of the dataset, according to Hive documentation. This seems like a pretty good number to use in preference to totalSize. However, due to HIVE-20079, this value is simply #columns * #rows. Once that bug is fixed, it may be a superior statistic, at least for managed tables.
> In the meantime, we could apply a configurable "fudge factor" to totalSize, at least for types of files that are encoded and compressed. Hive has the setting hive.stats.deserialization.factor, which defaults to 1.0, and is described as follows:
> {quote}in the absence of uncompressed/raw data size, total file size will be used for statistics annotation. But the file may be compressed, encoded and serialized which may be lesser in size than the actual uncompressed/raw data size. This factor will be multiplied to file size to estimate the raw data size.
> {quote}
> Also, I propose a configuration setting to allow the user to completely ignore rawDataSize, since that value is broken (due to HIVE-20079). When that configuration setting is set to true, Spark would instead estimate the table as follows:
> - if totalSize is defined and greater than 0, use totalSize*fudgeFactor.
>  - else, use spark.sql.defaultSizeInBytes (default: Long.MaxValue)
> Caveat: This mitigates the issue only for Hive tables. It does not help much when the user is reading files using {{spark.read.parquet}}, unless we apply the same fudge factor there.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org