[ https://issues.apache.org/jira/browse/FLINK-11116?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Kostas Kloudas updated FLINK-11116: ----------------------------------- Description: In order to guarantee exactly-once semantics, the streaming file sink is implementing a two-phase commit protocol when writing files to the filesystem. Initially data is written to in-progress files. These files are then put into "pending" state when they are completed (based on the rolling policy), and they are finally committed when the checkpoint that put them in the "pending" state is acknowledged as complete. The above shows that in the case that we have: 1) checkpoints A, B, C coming 2) checkpoint A being acknowledged and 3) failure Then we may have files that do not belong to any checkpoint (because B and C were not considered successful). These files are currently not cleaned up. In order to reduce the amount of such files created, we removed the random suffix from in-progress temporary files, so that the next in-progress file that is opened for this part, overwrites them. was: In order to guarantee exactly-once semantics, the streaming file sink is implementing a two-phase commit protocol when writing files to the filesystem. Initially data is written to in-progress files. These files are then put into "pending" state when they are completed (based on the rolling policy), and they are finally committed when the checkpoint that put them in the "pending" state is acknowledged as complete. The above shows that in the case that we have: 1) checkpoints A, B, C coming 2) checkpoint A being acknowledged and 3) failure Then we may have files that do not belong to any checkpoint (because B and C were not considered successful). These files are currently not cleaned up. This issue aims at cleaning up these files. > Overwrite outdated in-progress files in StreamingFileSink. > ---------------------------------------------------------- > > Key: FLINK-11116 > URL: https://issues.apache.org/jira/browse/FLINK-11116 > Project: Flink > Issue Type: Improvement > Components: filesystem-connector > Affects Versions: 1.7.0 > Reporter: Kostas Kloudas > Assignee: Kostas Kloudas > Priority: Major > Fix For: 1.7.2 > > > In order to guarantee exactly-once semantics, the streaming file sink is > implementing a two-phase commit protocol when writing files to the filesystem. > Initially data is written to in-progress files. These files are then put into > "pending" state when they are completed (based on the rolling policy), and > they are finally committed when the checkpoint that put them in the "pending" > state is acknowledged as complete. > The above shows that in the case that we have: > 1) checkpoints A, B, C coming > 2) checkpoint A being acknowledged and > 3) failure > Then we may have files that do not belong to any checkpoint (because B and C > were not considered successful). These files are currently not cleaned up. > In order to reduce the amount of such files created, we removed the random > suffix from in-progress temporary files, so that the next in-progress file > that is opened for this part, overwrites them. -- This message was sent by Atlassian JIRA (v7.6.3#76005)