- If migrating from release older than 0.5.3, please also check the upgrade instructions for each subsequent release below.
- With 0.6.0 Hudi is moving from list based rollback to marker based rollbacks. To smoothly aid this transition a
new property called
hoodie.table.versionis added to
hoodie.propertiesfile. Whenever Hudi is launched with newer table version i.e 1 (or moving from pre 0.6.0 to 0.6.0), an upgrade step will be executed automatically. This automatic upgrade step will happen just once per Hudi table as the
hoodie.table.versionwill be updated in property file after upgrade is completed.
- Similarly, a command line tool for Downgrading (command -
downgrade) is added if in case some users want to downgrade Hudi from table version 1 to 0 or move from Hudi 0.6.0 to pre 0.6.0
- If you were using a user defined partitioner with bulkInsert() RDD API, the base interface has changed to
BulkInsertPartitionerand will need minor adjustments to your existing implementations.
- Bootstrapping existing parquet datasets : Adds support for bootstrapping existing datasets into Hudi, via both Spark datasource writer and deltastreamer tool, with support for reading from Hive, SparkSQL, AWS Athena (prestoDB support coming soon). See RFC-15 for technical details. Note that this is an experimental feature, which will be improved upon further in the 0.6.x versions.
- Native row writing for bulk_insert : Avoids any dataframe-rdd conversion for bulk_insert path, which can improve performance of initial bulk loads. Although, this is typically not the bottleneck for upsert/deletes, subsequent releases in 0.6.x versions will expand this to other write operations to make reasoning about schema management easier, avoiding the spark-avro conversion totally.
- Bulk insert sort modes : Hudi bulk_insert sorts the input globally to optimize file sizes and avoid out-of-memory issues encountered when writing parallely to multiple DFS partitions.
For users who want to prepare the dataframe for writing outside of Hudi, we have made this configurable using
- Cleaning can now be run concurrently with writing, using
hoodie.clean.async=truewhich can speed up time taken to finish committing.
- Async compaction for spark streaming writes to hudi table, is now self managed by default, controlling
- Rollbacks no longer perform full table listings, by leveraging marker files. To enable, set
- Added a new index
hoodie.index.type=SIMPLEwhich can be faster than
BLOOM_INDEXfor cases where updates/deletes spread across a large portion of the table.
- Hudi now supports
Azure Data Lake Storage V2,
Tencent Cloud Object Storagestorages.
- HoodieMultiDeltaStreamer adds support for ingesting multiple kafka streams in a single DeltaStreamer deployment, effectively reducing operational burden for using delta streamer as your data lake ingestion tool (Experimental feature)
- Added a new tool - InitialCheckPointProvider, to set checkpoints when migrating to DeltaStreamer after an initial load of the table is complete.
- Delta Streamer tool now supports ingesting CSV data sources, chaining of multiple transformers to build more advanced ETL jobs.
- Introducing a new
CustomKeyGeneratorkey generator class, that provides flexible configurations to provide enable different types of key, partition path generation in single class. We also added support for more time units and date/time formats in
TimestampBasedKeyGenerator. See docs for more.
- Starting 0.6.0, snapshot queries are feasible on MOR tables using spark datasource. (experimental feature)
- In prior versions we only supported
HoodieCombineHiveInputFormatfor CopyOnWrite tables to ensure that there is a limit on the number of mappers spawned for any query. Hudi now supports Merge on Read tables also using
- Speedup spark read queries by caching metaclient in HoodieROPathFilter. This helps reduce listing related overheads in S3 when filtering files for read-optimized queries.
- Spark DAGs are named to aid better debuggability.
- Support pluggable metrics reporting by introducing proper abstraction for user defined metrics. Console, JMX, Prometheus and DataDog metric reporters have been added.
- A new utility called Data snapshot exporter has been added. Latest table snapshot as of a certain point in time can be exported as plain parquet files with this tool.
- Introduce write committed callback hooks for incremental pipelines to be notified and act on new commits in the timeline. For e.g, Apache Airflow jobs can be triggered as new commits arrive.
- Added support for deleting savepoints via CLI
- Added a new command -
export instants, to export metadata of instants
The raw release notes are available here