[SPARK-20594][SQL]The staging directory should be a child directory starts with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory. #17858

zuotingbing · 2017-05-04T07:25:47Z

JIRA Issue: https://issues.apache.org/jira/browse/SPARK-20594

What changes were proposed in this pull request?

The staging directory should be a child directory starts with "." to avoid being deleted before moving staging directory to table directory if we set hive.exec.stagingdir under the table directory.

How was this patch tested?

Added unit tests

…ging" to avoid being deleted if we set hive.exec.stagingdir under the table directory without start with "."

zuotingbing · 2017-05-04T08:13:32Z

In this case, Hive will create the staging directory under the table directory, and when moving staging directory to table directory, Hive will still empty the table directory, but will exclude the staging directory which start with "." or "_"

public static final PathFilter HIDDEN_FILES_PATH_FILTER = new PathFilter() { public boolean accept(Path p) { String name = p.getName(); return !name.startsWith("_") && !name.startsWith("."); } };

srowen · 2017-05-04T10:01:59Z

Do you really need to force this? or, is it just that any path relative to the output dir has to be a hidden directory starting with "." or "_"? For example, right now this prevents me from making the staging dir "/foo/bar" but I don't see a reason to disallow that.

gatorsmile · 2017-05-04T23:09:30Z

This sounds a bug in Hive metastore. Could you try the same thing in Hive? Do you hit the same error? Let us see how Hive behaves and then we can decide what is the best way to handle it. Thanks!

BTW, you need to create a test case. For example, InsertIntoHiveTableSuite.scala.

cloud-fan · 2017-05-05T15:27:14Z

if this is a hive bug, this patch seems a valid workaround for Spark SQL.

zuotingbing · 2017-05-08T06:01:56Z

yes i tried the same thing in Hive(version 2.10), got the same error:
`2017-05-08T13:48:04,634 ERROR exec.Task (:()) - Failed with exception Unable to move source hdfs://nameservice/hive/test_table1/test_hive_2017-05-08_13-47-40_660_5235248825413690559-1/-ext-10000 to destination hdfs://nameservice/hive/test_table1
org.apache.hadoop.hive.ql.metadata.HiveException: Unable to move source hdfs://nameservice/hive/test_table1/test_hive_2017-05-08_13-47-40_660_5235248825413690559-1/-ext-10000 to destination hdfs://nameservice/hive/test_table1
at org.apache.hadoop.hive.ql.metadata.Hive.moveFile(Hive.java:2959)
at org.apache.hadoop.hive.ql.metadata.Hive.replaceFiles(Hive.java:3198)
at org.apache.hadoop.hive.ql.metadata.Hive.loadTable(Hive.java:1805)
at org.apache.hadoop.hive.ql.exec.MoveTask.execute(MoveTask.java:355)
at org.apache.hadoop.hive.ql.exec.Task.executeTask(Task.java:197)
at org.apache.hadoop.hive.ql.exec.TaskRunner.runSequential(TaskRunner.java:100)
at org.apache.hadoop.hive.ql.Driver.launchTask(Driver.java:1917)
at org.apache.hadoop.hive.ql.Driver.execute(Driver.java:1586)
at org.apache.hadoop.hive.ql.Driver.runInternal(Driver.java:1331)
at org.apache.hadoop.hive.ql.Driver.run(Driver.java:1092)
at org.apache.hadoop.hive.ql.Driver.run(Driver.java:1080)
at org.apache.hadoop.hive.cli.CliDriver.processLocalCmd(CliDriver.java:232)
at org.apache.hadoop.hive.cli.CliDriver.processCmd(CliDriver.java:183)
at org.apache.hadoop.hive.cli.CliDriver.processLine(CliDriver.java:399)
at org.apache.hadoop.hive.cli.CliDriver.executeDriver(CliDriver.java:776)
at org.apache.hadoop.hive.cli.CliDriver.run(CliDriver.java:714)
at org.apache.hadoop.hive.cli.CliDriver.main(CliDriver.java:641)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:57)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:606)
at org.apache.hadoop.util.RunJar.run(RunJar.java:221)
at org.apache.hadoop.util.RunJar.main(RunJar.java:136)
Caused by: java.io.FileNotFoundException: File hdfs://nameservice/hive/test_table1/test_hive_2017-05-08_13-47-40_660_5235248825413690559-1/-ext-10000 does not exist.
at org.apache.hadoop.hdfs.DistributedFileSystem.listStatusInternal(DistributedFileSystem.java:697)
at org.apache.hadoop.hdfs.DistributedFileSystem.access$600(DistributedFileSystem.java:105)
at org.apache.hadoop.hdfs.DistributedFileSystem$15.doCall(DistributedFileSystem.java:755)
at org.apache.hadoop.hdfs.DistributedFileSystem$15.doCall(DistributedFileSystem.java:751)
at org.apache.hadoop.fs.FileSystemLinkResolver.resolve(FileSystemLinkResolver.java:81)
at org.apache.hadoop.hdfs.DistributedFileSystem.listStatus(DistributedFileSystem.java:751)
at org.apache.hadoop.fs.FileSystem.listStatus(FileSystem.java:1485)
at org.apache.hadoop.fs.FileSystem.listStatus(FileSystem.java:1525)
at org.apache.hadoop.hive.ql.metadata.Hive.moveFile(Hive.java:2896)
... 22 more

2017-05-08T13:48:04,635 ERROR ql.Driver (:()) - FAILED: Execution Error, return code 1 from org.apache.hadoop.hive.ql.exec.MoveTask. Unable to move source hdfs://nameservice/hive/test_table1/test_hive_2017-05-08_13-47-40_660_5235248825413690559-1/-ext-10000 to destination hdfs://nameservice/hive/test_table1`

gatorsmile · 2017-05-08T07:30:42Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/InsertIntoHiveTable.scala

@@ -222,7 +222,7 @@ case class InsertIntoHiveTable(
    val externalCatalog = sparkSession.sharedState.externalCatalog
    val hiveVersion = externalCatalog.asInstanceOf[HiveExternalCatalog].client.version
    val hadoopConf = sessionState.newHadoopConf()
-    val stagingDir = hadoopConf.get("hive.exec.stagingdir", ".hive-staging")
+    val stagingDir = hadoopConf.get("hive.exec.stagingdir", ".hive-staging") + "/.hive-staging"


How about?

// SPARK-20594 After Hive 1.1, Hive will create the staging directory under the table directory, // and when moving staging directory to table directory, Hive will still empty the table // directory, but will exclude the staging directory, which start with "." or "_". val stagingDir = new Path(hadoopConf.get("hive.exec.stagingdir", ".hive-staging"), ".hive-staging").toString

gatorsmile · 2017-05-08T07:32:34Z

This will not pass the test cases, because we only deleted the child directory .hive-staging of the stagingDir. We should clean it up. Let me trigger the test to show you the test results.

gatorsmile · 2017-05-08T07:32:43Z

ok to test

SparkQA · 2017-05-08T07:39:30Z

Test build #76566 has finished for PR 17858 at commit de938ed.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds no public classes.

zuotingbing · 2017-05-08T08:50:44Z

@gatorsmile it seems my mistake, i will try to fix this.

… with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory.

gatorsmile · 2017-05-09T04:25:11Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/InsertIntoHiveTable.scala

+    // SPARK-20594: The staging directory should be a child directory starts with "." to avoid
+    // being deleted if we set hive.exec.stagingdir under the table directory.
+    if (FileUtils.isSubDir(new Path(stagingPathName), inputPath, fs)
+      && !stagingPathName.stripPrefix(inputPathName).startsWith(".")) {


This is just to hide the issue and make the test cases passed, right?

We need to drop the created staging directory no matter what is the value users set.

Sorry i do not follow your logic. Correct me if I'm wrong, but isn't the logic of dropping the created staging directory was already there before with fs.deleteOnExit(dir)?
As @cloud-fan said this patch seems a valid workaround in Spark SQL for this case.

fs.deleteOnExit(dir) deletes dir, but the parent directory is still there.

SparkQA · 2017-05-09T05:53:59Z

Test build #76617 has finished for PR 17858 at commit 6b22d3e.

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

SparkQA · 2017-05-09T06:44:37Z

Test build #76644 has finished for PR 17858 at commit 6b1b153.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2017-05-09T08:52:26Z

Test build #76655 has finished for PR 17858 at commit 2a542e4.

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

gatorsmile · 2017-05-09T19:57:09Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/InsertIntoHiveTable.scala

+    // SPARK-20594: The staging directory should be a child directory starts with "." to avoid
+    // being deleted if we set hive.exec.stagingdir under the table directory.
+    if (FileUtils.isSubDir(new Path(stagingPathName), inputPath, fs)
+      && !stagingPathName.stripPrefix(inputPathName).stripPrefix(File.separator).startsWith(".")) {


fs.deleteOnExit(dir) deletes dir, but the parent directory is still there.

The dir with executionId is the staging dir which seems to me should be deleted exactly.
BTW, if we set hive.exec.stagingdir=/test/a/b , the dir /test/a will also not be deleted by the logic before.
Should i need to delete the parent directory only for this patch? and it seems not safe to do this since the parent directory could be used by other processes. Thanks.

Ok. This fix does not make things worse. I might accept it. cc @cloud-fan

gatorsmile · 2017-05-10T07:27:35Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/InsertIntoHiveTable.scala

      if (inputPathName.indexOf(stagingDir) == -1) {
        new Path(inputPathName, stagingDir).toString
      } else {
        inputPathName.substring(0, inputPathName.indexOf(stagingDir) + stagingDir.length)
      }
+
+    // SPARK-20594: The staging directory should be a child directory starts with "." to avoid
+    // being deleted if we set hive.exec.stagingdir under the table directory.


How about?

SPARK-20594: This is a walk-around fix to resolve a Hive bug. Hive requires that the staging directory needs to avoid being deleted when users set hive.exec.stagingdir under the table directory.

ok, i will update it . Thanks!

gatorsmile · 2017-05-10T07:32:11Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/InsertIntoHiveTableSuite.scala

+      |to avoid being deleted if we set hive.exec.stagingdir under the table directory
+      |without start with "."""".stripMargin) {
+
+    dropTables("test_table", "test_table1")


withTable("test_table", "test_table1") { ... }

yes you are right. :)

gatorsmile · 2017-05-10T07:33:01Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/InsertIntoHiveTableSuite.scala

+    tableNames.foreach { name =>
+      sql(s"DROP TABLE IF EXISTS $name")
+    }
+  }


This is not needed. You can call withTable.

gatorsmile · 2017-05-10T07:34:41Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/InsertIntoHiveTableSuite.scala

+    sql("CREATE TABLE test_table (key int, value string)")
+
+    // Add some data.
+    testData.write.mode(SaveMode.Append).insertInto("test_table")


You can simplify the above two lines by spark.range(1).write.saveAsTable("test_table")

no, as i tested we must create table rather than simplify the above two lines by spark.range(1).write.saveAsTable("test_table").

SparkQA · 2017-05-10T10:46:19Z

Test build #76744 has finished for PR 17858 at commit 9f41436.

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

gatorsmile · 2017-05-10T21:27:04Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/InsertIntoHiveTableSuite.scala

+      sql("CREATE TABLE test_table1 (key int)")
+
+      // Set hive.exec.stagingdir under the table directory without start with ".".
+      sql("set hive.exec.stagingdir=./test")


Could you set it back after this test case.

gatorsmile · 2017-05-10T23:56:30Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/InsertIntoHiveTable.scala

+      && !stagingPathName.stripPrefix(inputPathName).stripPrefix(File.separator).startsWith(".")) {
+      logDebug(s"The staging dir '$stagingPathName' should be a child directory starts " +
+        s"with '.' to avoid being deleted if we set hive.exec.stagingdir under the table " +
+        s"directory.")


Nit: please remove the above TWO string Interpolation s

gatorsmile · 2017-05-10T23:57:49Z

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/InsertIntoHiveTable.scala

+    // staging directory needs to avoid being deleted when users set hive.exec.stagingdir
+    // under the table directory.
+    if (FileUtils.isSubDir(new Path(stagingPathName), inputPath, fs)
+      && !stagingPathName.stripPrefix(inputPathName).stripPrefix(File.separator).startsWith(".")) {


Nit: Please move && in line 110

gatorsmile · 2017-05-11T00:01:11Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/InsertIntoHiveTableSuite.scala

+      sql("set hive.exec.stagingdir=./test")
+
+      // Now overwrite.
+      sql("INSERT OVERWRITE TABLE test_table1 SELECT * FROM test_table")


We do not need test_table, right?

INSERT OVERWRITE TABLE test_table1 SELECT 1

SparkQA · 2017-05-11T04:16:11Z

Test build #76765 has finished for PR 17858 at commit bf1b4ec.

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

gatorsmile · 2017-05-11T08:14:58Z

sql/hive/src/test/scala/org/apache/spark/sql/hive/InsertIntoHiveTableSuite.scala

+  test(
+    """SPARK-20594: This is a walk-around fix to resolve a Hive bug. Hive requires that the
+      |staging directory needs to avoid being deleted when users set hive.exec.stagingdir
+      |under the table directory.""".stripMargin) {


Without your fix, this test case still passes. Could you please check it?

Hmm, as i tested again, we must create table rather than simplify by spark.range(1).write.saveAsTable("test_table"). Thanks again. :)

uh, because that is not to create a Hive table. How about simplifying the test case to?

test("SPARK-20594: hive.exec.stagingdir was deleted by Hive") { // Set hive.exec.stagingdir under the table directory without start with ".". withSQLConf("hive.exec.stagingdir" -> "./test") { withTable("test_table") { sql("CREATE TABLE test_table (key int)") sql("INSERT OVERWRITE TABLE test_table SELECT 1") checkAnswer(sql("SELECT * FROM test_table"), Row(1)) } } }

good choice. thanks for your time!

SparkQA · 2017-05-11T13:44:04Z

Test build #76806 has finished for PR 17858 at commit 4e1b6a0.

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

SparkQA · 2017-05-12T03:10:10Z

Test build #76839 has finished for PR 17858 at commit 639d63a.

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

gatorsmile · 2017-05-12T18:23:24Z

LGTM

…starts with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory. JIRA Issue: https://issues.apache.org/jira/browse/SPARK-20594 ## What changes were proposed in this pull request? The staging directory should be a child directory starts with "." to avoid being deleted before moving staging directory to table directory if we set hive.exec.stagingdir under the table directory. ## How was this patch tested? Added unit tests Author: zuotingbing <[email protected]> Closes #17858 from zuotingbing/spark-stagingdir. (cherry picked from commit e3d2022) Signed-off-by: Xiao Li <[email protected]>

gatorsmile · 2017-05-12T18:26:09Z

Thanks! Merging to master/2.2

zuotingbing · 2017-05-15T01:42:59Z

Thank you all. Delete the branch.

…starts with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory. JIRA Issue: https://issues.apache.org/jira/browse/SPARK-20594 ## What changes were proposed in this pull request? The staging directory should be a child directory starts with "." to avoid being deleted before moving staging directory to table directory if we set hive.exec.stagingdir under the table directory. ## How was this patch tested? Added unit tests Author: zuotingbing <[email protected]> Closes apache#17858 from zuotingbing/spark-stagingdir.

[SPARK-20594]The staging directory should be appended with ".hive-sta…

c154f3a

…ging" to avoid being deleted if we set hive.exec.stagingdir under the table directory without start with "."

gatorsmile reviewed May 8, 2017

View reviewed changes

add a test case for SPARK-20594

de938ed

[SPARK-20594]The staging directory should be a child directory starts…

6b22d3e

… with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory.

gatorsmile reviewed May 9, 2017

View reviewed changes

should add stripPrefix(File.separator)

6b1b153

Fix scala style checks

2a542e4

gatorsmile reviewed May 9, 2017

View reviewed changes

gatorsmile reviewed May 10, 2017

View reviewed changes

Fix code review

9f41436

gatorsmile reviewed May 10, 2017

View reviewed changes

gatorsmile reviewed May 11, 2017

View reviewed changes

Fix code review

bf1b4ec

gatorsmile reviewed May 11, 2017

View reviewed changes

Fix test case

4e1b6a0

simplify the test case

639d63a

asfgit closed this in e3d2022 May 12, 2017

zuotingbing deleted the spark-stagingdir branch May 15, 2017 01:51

[SPARK-20594][SQL]The staging directory should be a child directory starts with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory. #17858

[SPARK-20594][SQL]The staging directory should be a child directory starts with "." to avoid being deleted if we set hive.exec.stagingdir under the table directory. #17858

Conversation

zuotingbing commented May 4, 2017 • edited Loading

What changes were proposed in this pull request?

How was this patch tested?

zuotingbing commented May 4, 2017

srowen commented May 4, 2017

gatorsmile commented May 4, 2017

cloud-fan commented May 5, 2017

zuotingbing commented May 8, 2017 • edited Loading

Choose a reason for hiding this comment

gatorsmile commented May 8, 2017 • edited Loading

gatorsmile commented May 8, 2017

SparkQA commented May 8, 2017

zuotingbing commented May 8, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SparkQA commented May 9, 2017

SparkQA commented May 9, 2017

SparkQA commented May 9, 2017

gatorsmile May 9, 2017 • edited Loading

Choose a reason for hiding this comment

zuotingbing May 10, 2017 • edited Loading

Choose a reason for hiding this comment

gatorsmile May 10, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SparkQA commented May 10, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SparkQA commented May 11, 2017

Choose a reason for hiding this comment

zuotingbing May 11, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SparkQA commented May 11, 2017

SparkQA commented May 12, 2017

gatorsmile commented May 12, 2017

gatorsmile commented May 12, 2017

zuotingbing commented May 15, 2017

zuotingbing commented May 4, 2017 •

edited

Loading

zuotingbing commented May 8, 2017 •

edited

Loading

gatorsmile commented May 8, 2017 •

edited

Loading

gatorsmile May 9, 2017 •

edited

Loading

zuotingbing May 10, 2017 •

edited

Loading

gatorsmile May 10, 2017 •

edited

Loading

zuotingbing May 11, 2017 •

edited

Loading