Details
Major Description
This is umbrella jira to track hive streaming ingest improvements. At a high level following are the improvements
- Support for dynamic partitioning
- API changes (simple streaming connection builder)
- Hide the transaction batches from clients (client can tune the transaction batch but doesn't have to know about the transaction batch size)
- Support auto rollover to next transaction batch (clients don't have to worry about closing a transaction batch and opening a new one)
- Record writers will all be strict meaning the schema of the record has to match table schema. This is to avoid the multiple serialization/deserialization for re-ordering columns if there is schema mismatch
- Automatic distribution for non-bucketed tables so that compactor can have more parallelism
- Create delta files with all ORC overhead disabled (no index, no compression, no dictionary). Compactor will recreate the orc files with index, compression and dictionary encoding.
- Automatic memory management via auto-flushing (will yield smaller stripes for delta files but is more scalable and clients don't have to worry about distributing the data across writers)
- Support for more writers (Avro specifically. ORC passthrough format?)
- Support to accept input stream instead of record byte[]
- Removing HCatalog dependency (old streaming API will be in the hcatalog package for backward compatibility, new streaming API will be in its own hive module)