[DRILL-5612] Random failure in TestMergeJoinWithSchemaChanges - ASF JIRA

Attach files

Attach Screenshot

Voters

Watch issue

Watchers

Create sub-task

Link

Clone

Update Comment Author

Replace String in Comment

Update Comment Visibility

Delete Comments

XML

Word

Printable

JSON

Details

Type: Bug
Status: Resolved
Priority: Major
Resolution: Fixed
Affects Version/s: 1.11.0
Fix Version/s: 1.20.0
Component/s: None
Labels:
None

Description

The unit test org.apache.drill.exec.physical.impl.join.TestMergeJoinWithSchemaChanges#testMissingAndNewColumns is subject to random failures, perhaps due to changes in file order in readers.

The test builds a number of input files, then executes queries against them. On most runs, the output is fine:

Running org.apache.drill.exec.physical.impl.join.TestMergeJoinWithSchemaChanges#testMissingAndNewColumns
/home/.../target/1498606483211-0/mergejoin-schemachanges-left
/home/.../target/1498606483211-1/mergejoin-schemachanges-right

But, on occasion, the query fails:

org.apache.drill.exec.physical.impl.join.TestMergeJoinWithSchemaChanges
testMissingAndNewColumns(org.apache.drill.exec.physical.impl.join.TestMergeJoinWithSchemaChanges)  Time elapsed: 0.569 sec  <<< ERROR!
...: UNSUPPORTED_OPERATION ERROR: Sort doesn't currently support sorts with changing schemas

Fragment 0:0

  (org.apache.drill.exec.exception.SchemaChangeException) Sort currently only supports a single schema.
    org.apache.drill.exec.physical.impl.sort.SortRecordBatchBuilder.build():152
    org.apache.drill.exec.physical.impl.xsort.ExternalSortBatch.innerNext():476
...

The line in the exception above:

  public void build(VectorContainer outputContainer) throws SchemaChangeException {
    outputContainer.clear();
    if (batches.keySet().size() > 1) {
      throw new SchemaChangeException("Sort currently only supports a single schema.");
    }

The above code has not changed in quite some time. The failure is in the "legacy" external sort.

Although the external sort does support schema changes, it only does so in the form of a union vector, which must be enabled. (Other tests validate that schema changes work.)

What is likely happening here is that the sort sometimes sees two files with differing schemas, sometimes multiple threads run so that a single sort sees only one file. This speculation can be verified by looking at a log file (not available in the test run that failed) to see if the scan under the sort read more than one file.

Or, perhaps the order of the JSON files matters. Perhaps file order varies across machines (since the Linux command to list directories does not guarantee order.)