Checkpoint files (offset log files, state store files) in Structured Streaming must be written atomically such that no partial files are generated (would break fault-tolerance guarantees). Currently, there are 3 locations which try to do this individually, and in some cases, incorrectly.
- HDFSOffsetMetadataLog - This uses a FileManager interface to use any implementation of `FileSystem` or `FileContext` APIs. It preferably loads `FileContext` implementation as FileContext of HDFS has atomic renames.
- HDFSBackedStateStore (aka in-memory state store)
- Writing a version.delta file - This uses FileSystem APIs only to perform a rename. This is incorrect as rename is not atomic in HDFS FileSystem implementation.
- Writing a snapshot file - Same as above.
- State Store behavior is incorrect -
- Inflexible - Some file systems provide mechanisms other than write-to-temp-file-and-rename for writing atomically and more efficiently. For example, with S3 you can write directly to the final file and it will be made visible only when the entire file is written and closed correctly. Any failure can be made to terminate the writing without making any partial files visible in S3. The current code does not abstract out this mechanism enough that it can be customized.
- Introduce a common interface that all 3 cases above can use to write checkpoint files atomically.
- This interface must provide the necessary interfaces that allow customization of the write-and-rename mechanism.