- Generalized BloomIndex to work with file ids instead of paths
- Abstracted away Bloom filter checking into HoodieLookupHandle
- Abstracted away range information retrieval into HoodieRangeInfoHandle
1. Go through dependencies list one round to ensure compliance. Generated current NOTICE list in all submodules (other apache projects like flink does this).
To be on conservative side regarding licensing, NOTICE.txt lists all dependencies including transitive. Pending Compliance questions reported in https://issues.apache.org/jira/browse/LEGAL-461
2. Automate generating NOTICE.txt files to allow future package compliance issues be identified early as part of code-review process.
3. Added NOTICE.txt and LICENSE.txt to all HUDI jars
- Join based on HoodieKey and not RecordKey during tagging
- Unit tests changed to run with duplicate keys
- Special casing GlobalBloom to still join by recordkey
1. Parquet Avro schema mismatch errors when ingesting are sometimes silently ignored due to race-condition in BoundedInMemoryExecutor. This was reproducible when running long-running delta-streamer with wrong schema and it caused data-loss
2. Fix behavior of Delta-Streamer to error out by default if there are any error records
3. Fix a bug in tracking write errors in WriteStats. Earlier the write errors were tracking sampled errors as opposed to total errors.
4. Delta Streamer does not commit the changes done as part of inline compaction as auto-commit is force disabled. Fix this behavior to always auto-commit inline compaction as it would not otherwise commit.
- Tackles the skew seen in sort based partitioning/checking
- Parameterized the HoodieBloomIndex test
- Config to turn on/off (on by default)
- Unit tests & also tested at scale
- Users using spark.executor.cores > 1 used to fail due to "FileSystem closed"
- This is due to HoodieWrapperFileSystem closing the wrapped filesytem obj
- FileSystem.getInternal caching code races threads and closes the extra fs instance(s)
- Bumped up num cores in tests to 8, speeds up tests by 3-4 mins
- For implicit indexes (e.g BloomIndex), don't buffer up written records
- By default, only collect 10% of failing records to avoid OOMs
- Improves debuggability via above, since data errors can now show up in collect()
- Unit tests & fixing subclasses & adjusting tests
- Added a precondition check, otherwise incr pull could miss commits
- Lowered default cleaner retention to 10, to enable simpler understanding for newbies
- Bumped down min/max instants to retain as well
(1) Apply transformation when using delta-streamer to ingest data.
(2) Add Hudi Incremental Source for Delta Streamer
(3) Allow delta-streamer config-property to be passed as command-line
(4) Add Hive Integration to Delta-Streamer and address Review comments
(5) Ensure MultiPartKeysValueExtractor handle hive style partition description
(6) Reuse same spark session on both source and transformer
(7) Support extracting partition fields from _hoodie_partition_path for HoodieIncrSource
(8) Reuse Binary Avro coders
(9) Add push down filter for Incremental source
(10) Add Hoodie DeltaStreamer metrics to track total time taken
