[FLINK-9603] Incorrect indexing of part files, when part suffix is specified (FileAlreadyExistsException)

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.5.0, 1.6.0
• Fix Version/s: 1.5.1, 1.6.0
• Component/s: Connectors / FileSystem
• Labels: pull-request-available
• Assignee: Kostas Kloudas
• Reporter: Rinat Sharipov
• Created: 16/Jun/18 14:28
• Updated: 12/Jul/18 11:50
• Resolved: 12/Jul/18 11:50

Description

Hi mates, since 1.5 release, BucketingSink has ability to configure suffix of the part file. It’s very useful, when it’s necessary to set specific extension of the file.
 
During the usage, I’ve found the issue - when new part file is created, it has the same part index, as index of just closed file. 
So, when Flink tries to move it into final state, we have a FileAlreadyExistsException.
 
This problem is related with the following code:
Here we are trying to find the max index of part file, that doesn’t exist in bucket directory, the problem is, that the partSuffix is not involved into path assembly. This means, that path always doesn’t exist
and partCounter wouldn’t be ever incremented.
 

Path partPath = new Path(bucketPath, partPrefix + "-" + subtaskIndex + "-" + bucketState.partCounter);
while (fs.exists(partPath) ||
      fs.exists(getPendingPathFor(partPath)) ||
      fs.exists(getInProgressPathFor(partPath))) {
   bucketState.partCounter++;
   partPath = new Path(bucketPath, partPrefix + "-" + subtaskIndex + "-" + bucketState.partCounter);
}

bucketState.creationTime = processingTimeService.getCurrentProcessingTime();

Before creating of writer, we appending the partSuffix here, but it should be already appended, before index checks

if (partSuffix != null) {
   partPath = partPath.suffix(partSuffix);
}

Issue Links

• links to: GitHub Pull Request #6176