[SPARK-24351] offsetLog/commitLog purge thresholdBatchId should be computed with current committed epoch but not currentBatchId in CP mode

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.3.0
• Fix Version/s: 2.4.0
• Component/s: Structured Streaming
• Assignee: Tengfei Huang
• Reporter: Tengfei Huang
• Created: 22/May/18 15:45
• Updated: 01/Jun/18 17:48
• Resolved: 01/Jun/18 17:48

Description

In structured streaming, there is a conf spark.sql.streaming.minBatchesToRetain which is set to specify 'The minimum number of batches that must be retained and made recoverable' as described in SQLConf. In continuous processing, the metadata purge is triggered when an epoch is committed in ContinuousExecution.
Since currentBatchId increases independently in cp mode, the current committed epoch may be far behind currentBatchId if some task hangs for some time. It is not safe to discard the metadata with thresholdBatchId computed based on currentBatchId because we may clean all the metadata in the checkpoint directory.

Issue Links

• links to: [Github] Pull Request #21400 (ivoson)