...
- Apache Hudi - Release Guide (Pre Graduation)
- Apache Hudi Community Bi-Weekly Sync
- Committer On-boarding Guide
- Community Support
Roadmap
Under construction
, early 2021 unveiling
Writing data & Indexing
Below is a tentative roadmap for 2021 (in no particular order; since that is determined by Release Management process)
Integrations
Spark SQL with Merge/Delete statements support (RFC - 25: Spark SQL Extension For Hudi)
Trino integration with support for querying/writing Hudi table using SQL statements
Kinesis/Pulsar integrations with DeltaStreamer
Kafka Connect Sink for Hudi
- Dremio integration
Interops with other table formats
- ORC Support
Writing
Indexing
MetadataIndex implementation that servers bloom filters/key ranges from metadata table, to speed up bloom index on cloud storage.
Addition of record level indexes for fast CDC (RFC-08 Record level indexing mechanisms for Hudi datasets)
Range index to maintain column/field value ranges, to help file skipping for query performance
Addition of more auxiliary indexing structures - bitmaps, ..
- Improving indexing speed for time-ordered keys/small updates
- leverage parquet record indexes,
- serving bloom filters/ranges from timeline server/consolidate metadata
- Indexing the log file, moving closer to scalable 1-min ingests
- Improving indexing speed for uuid-keys/large update spreads
global/hash based index to faster point-in-time lookup
- Incrementalize & standardize all metadata operations e.g cleaning based on timeline metadata
- Auto tuning
- Auto tune bloom filter entries based on records
- Partitioning based on historical workload trend
- Determination of compression ratio
Reading data
Concurrency Control
- Addition of optimistic concurrency control, with pluggable locking services.
Non-blocking clustering implementation w.r.t updates
- Multi-writer support with fully non-blocking log based concurrency control.
- Multi table transactions
- Performance
- Integrate row writer with all Hudi writer operations
Self Managing
Clustering based on historical workload trend
- On-fly data locality during write time (HUDI-1628)
Auto Determination of compression ratio
Querying
Performance
- Complete integration with metadata table.
- Realtime view performance/memory footprint reduction.
- PrestoDB
Incremental Query support on Presto
- Hive
- Storage handler to leverage metadata table for partition pruning
- Incremental Pull natively via Spark Datasource
- Real-time view support on Presto
- Spark SQL
Hardening incremental pull via Realtime view
- Realtime view performance/memory footprint reduction.
- Support for Streaming style batch programs via Beam/Structured Streaming integration
Storage
- ORC Support
- Support for collapsing and splitting file groups
- Custom strategies for data clustering
- Columnar stats collection to power better query planning
- Object storage
Usability
- Spark Datasource redesign around metadata table
- Streaming ETL via Structured Streaming
- Flink
Support for end-end streaming ETL pipelines
- Materialized view support via Flink/Calcite SQL
Mutable, Columnar Cache Service
- File group level caching to enable real-time analytics (backed by Arrow/AresDB)
- Painless migration of historical data, with safe experimentation
- Hudi on Flink
- Hudi for ML/Feature stores
Metadata Management
- Standalone timeline server to handle
- Serves interactive query planning performance: schema, DFS listings, statistics, timeline requests
- High availability/sharding
- Pluggable backing stores including rocksDB, Dynamo, Spanner
- Hudi timeline is a log. if we compact it we get a snapshot of the table