[jira] [Commented] (SPARK-21184) QuantileSummaries implementation is wrong and QuantileSummariesSuite fails with larger n

["Timothy Hunter (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-21184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16149510#comment-16149510 ] 

Timothy Hunter commented on SPARK-21184:
----------------------------------------

[~a1ray] thank you for the report, someone should investigate about these given values.

You raise some valid questions about the choice of data structures and algorithm, which were discussed during the implementation and that can certainly be revisited:

- tree structures: the major constraint here is that this structure gets serialized often, due to how UDAFs work. This is why the current implementation is amortized over multiple records. Edo Liberty has published some recent work that is relevant in that area.

- algorithm: we looked at t-digest (and q-digest). The main concern back then was that there was no published worst-time guarantee given a target precision. This is still the case to my knowledge. Because of that, it is hard to understand what could happen in some unusual cases - which tend to be not so unusual in big data. That being said, it looks like it is a popular and well-maintained choice now, so I am certainly open to relaxing this constraint.

> QuantileSummaries implementation is wrong and QuantileSummariesSuite fails with larger n
> ----------------------------------------------------------------------------------------
>
>                 Key: SPARK-21184
>                 URL: https://issues.apache.org/jira/browse/SPARK-21184
>             Project: Spark
>          Issue Type: Bug
>          Components: SQL
>    Affects Versions: 2.1.1
>            Reporter: Andrew Ray
>
> 1. QuantileSummaries implementation does not match the paper it is supposed to be based on.
> 1a. The compress method (https://github.com/apache/spark/blob/master/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/util/QuantileSummaries.scala#L240) merges neighboring buckets, but thats not what the paper says to do. The paper (http://infolab.stanford.edu/~datar/courses/cs361a/papers/quantiles.pdf) describes an implicit tree structure and the compress method deletes selected subtrees.
> 1b. The paper does not discuss merging these summary data structures at all. The following comment is in the merge method of QuantileSummaries:
> {quote}      // The GK algorithm is a bit unclear about it, but it seems there is no need to adjust the
>       // statistics during the merging: the invariants are still respected after the merge.{quote}
> Unless I'm missing something that needs substantiation, it's not clear that that the invariants hold.
> 2. QuantileSummariesSuite fails with n = 10000 (and other non trivial values)
> https://github.com/apache/spark/blob/master/sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/util/QuantileSummariesSuite.scala#L27
> One possible solution if these issues can't be resolved would be to move to an algorithm that explicitly supports merging and is well tested like https://github.com/tdunning/t-digest



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

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