You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@parquet.apache.org by Ryan Blue <rb...@netflix.com.INVALID> on 2016/09/08 18:13:58 UTC
Notes from the parquet sync-up
Hi everyone! Here are the notes from today's sync-up. Thanks to everyone
that attended!
Topics:
* Sort order for min/max stats
Parquet MR and Parquet Format sorting: PARQUET-686
PRs ready to go
Negative bytes are causing problems in the binary sort.
Create a new field for unsigned comparison. Should this be considered a
bug?
* Dependency switch arrow->parquet to parquet->arrow
Sort of like the Java side, add an arrow object model to C++
Need to decide if Arrow is optional or not
Python will depend on Parquet, maybe Arrow as well
Circular dependencies are bad, need to figure out a clean solution
* Compatibility checks between Java and C++
Brought up by signed byte comparison, C++ uses unsigned comparisons
We should have test suites to make sure they are readable
Look at the old parquet-compatibility lib
Piyush will open a JIRA and start looking at it
Java is reading files in the repository to make sure they're still correct
Would create a C++ version of this
* ByteBuffer: setup performance benchmarks and run them regularly
Ran into issues trying to release 1.9.0 at Twitter
Release ran 5-15% slower due to encoding/decoding
We don't have continuous performance tests for each PR, so we rely on devs
Use parquet-benchmark to test each commit?
How can we do this as part of the build? Need to ignore false-positives
We should at least test performance at RCs
Treat performance regressions like bugs - add benchmarks as we would test
cases
* Parquet 1.9.0 release
Check whether the min/max fix should be included and merge
Review PARQUET-623
Get a RC out today or tomorrow
* Quarterly releases
Piyush volunteered to be release manager for next quarter
* New Interval types
Drill wanted the old interval type, but would switch to the new ones
The old type is basically all of the information for the two new types
Old type would be deprecated
* New encodings
--
Ryan Blue
Software Engineer
Netflix