[SPARK-16628][SQL] Translate file-based relation schema when file schema is inconsistent with catalog schema by viirya · Pull Request #14365 · apache/spark

viirya · 2016-07-26T08:36:49Z

What changes were proposed in this pull request?

We will convert Metastore Orc tables (represented by MetastoreRelation) to datasource tables (represented by HadoopFsRelation) if spark.sql.hive.convertMetastoreOrc is enabled for better performance.

However, due to a Hive issue, an Orc table created by Hive does not store column name correctly in the Orc files. For these Orc tables, the converted relation has wrong schema.

To fix this, we assume the metastore schema dataSchema can match to physicalSchema by each column disregarding the column names. If not, we throw an exception that suggests users to disable the conversion of Hive Orc tables.

How was this patch tested?

Jenkins tests.

… for file-based data source relation.

SparkQA · 2016-07-26T09:59:19Z

Test build #62875 has finished for PR 14365 at commit f9fab59.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-07-26T11:43:33Z

Test build #62880 has finished for PR 14365 at commit cc953a3.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

viirya · 2016-07-28T07:33:05Z

cc @cloud-fan @yhuai @liancheng @rxin Please review this change. Thanks.

cloud-fan · 2016-07-28T07:47:15Z

+/**
+ * An interface for mapping two different schemas. For the relations that have are backed by files,
+ * the inferred schema from the files might be different with the schema stored in the catalog. In
+ * such case, the interface helps mapping inconsistent schemas.


Can you put the detailed description of this mapping in the doc here? thanks.

I've added more detailed document. Please take a look. Thanks.

SparkQA · 2016-07-29T05:41:35Z

Test build #62993 has finished for PR 14365 at commit 0c17748.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

viirya · 2016-08-02T10:04:45Z

cc @yhuai @liancheng @rxin

cloud-fan · 2016-08-02T11:20:06Z

@rxin @yhuai , is it worth adding this complex logic to work around this hive bug? This feature(convert Metastore Orc tables to data source table) is not a major feature and the broken case seems a corner case. I think it's ok to disable this feature when we meet this hive bug, what do you think?

viirya · 2016-08-02T14:12:48Z

@cloud-fan I've submitted a PR #14282 previously that disables the conversion if the schema is inconsistent.

…sistent-schema Conflicts: sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/FileSourceStrategy.scala

SparkQA · 2016-08-04T07:56:58Z

Test build #63214 has finished for PR 14365 at commit 5cfdff1.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

viirya · 2016-08-12T06:30:41Z

@cloud-fan So do we have decision on this? A simpler approach to disable the conversion if the schema is inconsistent, or a complex one to work around this Hive bug?

viirya · 2016-10-06T02:44:22Z

ping @cloud-fan Can you have a decision about this? Are we going to have complex logic for this issue? Or just disable it?

cloud-fan · 2016-10-09T08:54:32Z

I noticed that spark.sql.hive.convertMetastoreOrc is false by default, so this bug is not so critical.

And we may stop inferring file schema when reading ORC tables, for performance reasons(#16980), then we don't have the chance to detect the mismatch and disable this feature.

We may remove this feature entirely, or thinking harder to come up with a better fix. cc @yhuai

yhuai · 2016-10-10T18:22:59Z

@cloud-fan actually, this conversion was disabled because of this bug.

btw, pr that @cloud-fan mentioned is #14690.

I think it is better to hold this change before we finish #14690.

dongjoon-hyun · 2016-12-02T18:33:58Z

Hi, @viirya , @cloud-fan , @yhuai .
#14690 and the parent issue SPARK-17861 seems to be resolved.
What about restarting this issue?

viirya · 2016-12-07T08:20:30Z

As we replace schema inferring with metastore schema completely by #14690 for converted Hive tables, we may not have the chance to detect the mismatch between Orc file's physical schema and metastore schema.

dongjoon-hyun · 2016-12-07T13:29:03Z

It seems to exist some cases like the following (on the current master).

On hive,

CREATETABLEt1 (a string) PARTITIONED BY (b string) STORED AS ORC;
INSERT INTO TABLE t1 PARTITION (b='01') VALUES ('1');

On spark-shell,

scala> sql("select * from t1").show
+---+---+| a| b|+---+---+|1|01|+---+---+
scala> sql("set spark.sql.hive.convertMetastoreOrc=true").show
+--------------------+-----+| key|value|+--------------------+-----+|spark.sql.hive.co...|true|+--------------------+-----+
scala> sql("select * from t1").show
16/12/0705:17:19ERRORExecutor:Exception in task 0.0 in stage 2.0 (TID2)
java.lang.IllegalArgumentException:Field"a" does not exist.

viirya · 2016-12-08T04:48:20Z

@dongjoon-hyun yeah, I see. Because we directly use metastore schema of converted Orc table, when the physical schema in Orc file and metastore schema mismatch, this issue happens.

viirya · 2016-12-08T04:53:41Z

We have two options. First one is to map metastore schema to physical Orc schema like this. But we don't infer physical schema of Orc file now. I will update this to have this mapping in OrcFileFormat.

Another one is like #14282. But as we don't infer schema from Orc file now, we can't disable the conversion when the mismatch is detected. One possible is to throw exception in OrcFileFormat when detecting the mismatch before reading and show message to ask user to disable spark.sql.hive.convertMetastoreOrc.

@cloud-fan @yhuai @dongjoon-hyun What do you think?

…sistent-schema

SparkQA · 2016-12-08T07:02:45Z

Test build #69852 has started for PR 14365 at commit 2bb8368.

viirya · 2016-12-08T07:04:30Z

@cloud-fan @yhuai @dongjoon-hyun I've updated this as:

Assume metastore schema matches with physical Orc schema by column, disregarding column names.
Mapping required schema to columns in physical Orc schema.
If the length or data types of metastore schema and physical schema is not matched, throw an exception suggesting users to disable spark.sql.hive.convertMetastoreOrc.

Please let me know what you think about this approach. Thanks.

viirya · 2016-12-08T08:24:09Z

retest this please.

SparkQA · 2016-12-08T10:08:18Z

Test build #69856 has finished for PR 14365 at commit 2bb8368.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

dongjoon-hyun · 2016-12-08T20:10:49Z

 }

+ def isMismatchSchema(physicalSchema: StructType, requestedSchema: StructType): Boolean = {
+ requestedSchema.forall(a => physicalSchema.getFieldIndex(a.name).isEmpty)


Maybe, the following?
!requestedSchema.forall(a => physicalSchema.getFieldIndex(a.name).isDefined)

Oh, never mind. It's really about checking whether all requested columns are not matched.

dongjoon-hyun · 2016-12-08T22:04:48Z

It looks good to me. Thank you for updating, @viirya !

viirya · 2016-12-19T04:53:51Z

ping @cloud-fan @yhuai May you take a look? Thanks.

SparkQA · 2016-12-19T07:00:19Z

Test build #70336 has finished for PR 14365 at commit 141cb1d.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

viirya added 2 commits July 26, 2016 16:28

Fix the possible inconsistency between file schema and catalog schema…
f9fab59
… for file-based data source relation.

Fix test with correct require condition.
cc953a3

cloud-fan reviewed Jul 28, 2016
View reviewed changes

Add document.
0c17748

Merge remote-tracking branch 'upstream/master' into fix-file-ds-incon…
5cfdff1
…sistent-schema Conflicts: sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/FileSourceStrategy.scala

gatorsmile mentioned this pull request Aug 24, 2016
[SPARK-16948][SQL] Use metastore schema instead of inferring schema for ORC in HiveMetastoreCatalog #14537
Closed

viirya added 3 commits December 8, 2016 06:20

Do schema translation in OrcFileFormat.
4125e6f

Merge remote-tracking branch 'upstream/master' into fix-file-ds-incon…
9cbb406
…sistent-schema

Refine code comment.
2bb8368

dongjoon-hyun reviewed Dec 8, 2016
View reviewed changes

Add more test.
141cb1d

viirya closed this Feb 23, 2017

viirya deleted the fix-file-ds-inconsistent-schema branch December 27, 2023 18:20

Uh oh!

Conversation

viirya commented Jul 26, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

SparkQA commented Jul 26, 2016

Uh oh!

SparkQA commented Jul 26, 2016

Uh oh!

viirya commented Jul 28, 2016

Uh oh!

cloud-fanJul 28, 2016

Choose a reason for hiding this comment

Uh oh!

viiryaJul 29, 2016

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jul 29, 2016

Uh oh!

viirya commented Aug 2, 2016

Uh oh!

cloud-fan commented Aug 2, 2016

Uh oh!

viirya commented Aug 2, 2016

Uh oh!

SparkQA commented Aug 4, 2016

Uh oh!

viirya commented Aug 12, 2016

Uh oh!

viirya commented Oct 6, 2016

Uh oh!

cloud-fan commented Oct 9, 2016

Uh oh!

yhuai commented Oct 10, 2016

Uh oh!

dongjoon-hyun commented Dec 2, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

viirya commented Dec 7, 2016

Uh oh!

dongjoon-hyun commented Dec 7, 2016

Uh oh!

viirya commented Dec 8, 2016

Uh oh!

viirya commented Dec 8, 2016

Uh oh!

SparkQA commented Dec 8, 2016

Uh oh!

viirya commented Dec 8, 2016

Uh oh!

viirya commented Dec 8, 2016

Uh oh!

SparkQA commented Dec 8, 2016

Uh oh!

dongjoon-hyunDec 8, 2016

Choose a reason for hiding this comment

Uh oh!

dongjoon-hyunDec 8, 2016

Choose a reason for hiding this comment

Uh oh!

dongjoon-hyun commented Dec 8, 2016

Uh oh!

viirya commented Dec 19, 2016

Uh oh!

SparkQA commented Dec 19, 2016

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

viirya commented Jul 26, 2016•
edited
Loading

dongjoon-hyun commented Dec 2, 2016•
edited
Loading