* [HUDI-2795] Add mechanism to safely update,delete and recover table properties
- Fail safe mechanism, that lets queries succeed off a backup file
- Readers who are not upgraded to this version of code will just fail until recovery is done.
- Added unit tests that exercises all these scenarios.
- Adding CLI for recovery, updation to table command.
- [Pending] Add some hash based verfication to ensure any rare partial writes for HDFS
* Fixing upgrade/downgrade infrastructure to use new updation method
- Metadata table today has virtual keys disabled, thereby populating the metafields
for each record written out and increasing the overall storage space used. Hereby
adding virtual keys support for metadata table so that metafields are disabled
for metadata table records.
- Adding a custom KeyGenerator for Metadata table so as to not rely on the
default Base/SimpleKeyGenerators which currently look for record key
and partition field set in the table config.
- AbstractHoodieLogRecordReader's version of processing next data block and
createHoodieRecord() will be a generic version and making the derived class
HoodieMetadataMergedLogRecordReader take care of the special creation of
records from explictly passed in partition names.
* Enabling timeline server based markers
* Enabling timeline server based markers and marker based rollback
* Removing constraint that timeline server can be enabled only for hdfs
* Fixing tests
- ExternalSpillableMap does the payload/value size estimation on the first put to
determine when to spill over to disk map. The payload size re-estimation also
happens after a minimum threshold of puts. This size re-estimation goes my the
current in-memory map size for calculating average payload size and does attempts
divide by zero operation when the map is size is empty. Avoiding the
ArithmeticException during the payload size re-estimate by checking the map size
upfront.
* [HUDI-2634] Improved the metadata table bootstrap for very large tables.
Following improvements are implemented:
1. Memory overhead reduction:
- Existing code caches FileStatus for each file in memory.
- Created a new class DirectoryInfo which is used to cache a director's file list with parts of the FileStatus (only filename and file len). This reduces the memory requirements.
2. Improved parallelism:
- Existing code collects all the listing to the Driver and then creates HoodieRecord on the Driver.
- This takes a long time for large tables (11million HoodieRecords to be created)
- Created a new function in SparkRDDWriteClient specifically for bootstrap commit. In it, the HoodieRecord creation is parallelized across executors so it completes fast.
3. Fixed setting to limit the number of parallel listings:
- Existing code had a bug wherein 1500 executors were hardcoded to perform listing. This leads to exception due to limit in the spark's result memory.
- Corrected the use of the config.
Result:
Dataset has 1299 partitions and 12Million files.
file listing time=1.5mins
HoodieRecord creation time=13seconds
deltacommit duration=2.6mins
Co-authored-by: Sivabalan Narayanan <n.siva.b@gmail.com>
