[
https://issues.apache.org/jira/browse/SPARK-25206?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16592445#comment-16592445
]
Dongjoon Hyun commented on SPARK-25206:
---------------------------------------
[~yucai]. First of all, I know your intention and support both backports (one
bug + one improvement). Let me rephrase your words and add some.
1. Vanilla Spark 2.2.0 ~ 2.3.1 always returns NULL for Parquet mismatched
columns due to case sensitivity. It's not related to filters at all.
2. Spark 2.4 can handle this by SPARK-24716(Improvement, 2018-07-04) and
SPARK-25132(Bug, 2018-08-21). So, we need both of them in `branch-2.3`.
I'm saying (1) like you. Is (1) wrong? Does Spark 2.2.0 ~ 2.3.1 have some
inconsistency between `on` and `off` for parquet predicate pushdowns?
The reason why SPARK-25132 is not complete to your situations is simply that
it's based on `master` branch. It depends on lots of improvements only in
`master` branch. Now now, we only have one (SPARK-24716). But, for some other
corner cases, we may require more `master` patches.
> Wrong data may be returned for Parquet
> --------------------------------------
>
> Key: SPARK-25206
> URL: https://issues.apache.org/jira/browse/SPARK-25206
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.2.2, 2.3.1
> Reporter: yucai
> Priority: Blocker
> Labels: correctness
> Attachments: image-2018-08-24-18-05-23-485.png,
> image-2018-08-24-22-33-03-231.png, image-2018-08-24-22-34-11-539.png,
> image-2018-08-24-22-46-05-346.png, image-2018-08-25-09-54-53-219.png,
> image-2018-08-25-10-04-21-901.png, pr22183.png
>
>
> In current Spark 2.3.1, below query returns wrong data silently.
> {code:java}
> spark.range(10).write.parquet("/tmp/data")
> sql("DROP TABLE t")
> sql("CREATE TABLE t (ID LONG) USING parquet LOCATION '/tmp/data'")
> scala> sql("select * from t").show
> +----+
> | ID|
> +----+
> |null|
> |null|
> |null|
> |null|
> |null|
> |null|
> |null|
> |null|
> |null|
> |null|
> +----+
> scala> sql("select * from t where id > 0").show
> +---+
> | ID|
> +---+
> +---+
> scala> sql("set spark.sql.parquet.filterPushdown").show
> +--------------------+-----+
> | key|value|
> +--------------------+-----+
> |spark.sql.parquet...| true|
> +--------------------+-----+
> scala> sql("set spark.sql.parquet.filterPushdown=false").show
> +--------------------+-----+
> | key|value|
> +--------------------+-----+
> |spark.sql.parquet...|false|
> +--------------------+-----+
> scala> sql("select * from t where id > 0").show
> +---+
> | ID|
> +---+
> +---+
> {code}
>
> *Root Cause*
> Spark pushdowns FilterApi.gt(intColumn("{color:#ff0000}ID{color}"), 0:
> Integer) into parquet, but {color:#ff0000}ID{color} does not exist in
> /tmp/data (parquet is case sensitive, it has {color:#ff0000}id{color}
> actually).
> So no records are returned.
> In Spark 2.1, the user will get Exception:
> {code:java}
> Caused by: java.lang.IllegalArgumentException: Column [ID] was not found in
> schema!{code}
> But in Spark 2.3, they will get the wrong results sliently.
>
> Since SPARK-24716, Spark uses Parquet schema instead of Hive metastore schema
> to do the pushdown, perfect for this issue.
> [~yumwang], [~cloud_fan], [~smilegator], any thoughts? Should we backport it?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
