DataFusion weekly project plan (Andrew Lamb) - Jan 15, 2024 #8864
Description
Follow on to Jan 8, 2024
Boilerplate Introduction
The idea of this ticket is make my plans for DataFusion visible, largely for my own personal organizational needs, but also to:
- Communicate / coordinate in the community
- Help provide an interesting summary of what is happening in DataFusion this week
It would be great for other contributors to DataFusion who plan non trivial work could try to make them visible somehow as well 🙏 (feel free to copy / modify the format)
My (personal) plans for this week
- Complete blog post about DataFusion: [Website]: DataFusion 26-34 blog post arrow-site#457
- Complete some performance work (Improve aggregate performance with specialized groups accumulator for single string group by #7064) + benchmarking (Add 'clickbench_extended' benchmark #8860 and Add test coverage for grouping on dictionary encoded columns #8791
- (Stretch) Prepare next steps for making function packages more modular [EPIC] Unify Function Interface (remove
BuiltInScalarFunction) #8045
Project Queue (list of future projects)
- More advanced pruning predicates Don't error on unknown column when pruning if predicate can still be proven false #7869
- Statistics: Epic: Statistics improvements #8227 (specifically, Introduce a way to represent constrained statistics / bounds on values in Statistics #8078)
Projects I plan to prioritize reviewing / helping
- Memory explosion in HashJoin outputs:
HashJoinStreammemory tracking insufficient #7848 from @korowa - Improve planning speed [Epic] A collection of issues to improve planning performance / speed / efficiency #5637 with @matthewmturner
- Making function packages more modular: [EPIC] Unify Function Interface (remove
BuiltInScalarFunction) #8045 (thank you @guojidan) - Supporting recursive CTEs Add support for recursive CTEs #462 with @matthewgapp and @jonahgao
Algorithm for (my) prioritizing PR reviews
Note there are many committers who can and do review and merge PRs, so this is not the priorities of the project as a whole, just the approximate algorithm I use to prioritize my own time
Priority:
- Bug fixes (where something is just incorrect), especially regressions (where it used to work and now does not)
- Improvements directly related to features needed for InfluxDB (my employer)
- Documentation and test improvements (I view these as very strategically important)
- PRs that I think are strategically important
- Other new features / additions to functionality (note this is the lowest)
The current strategically important projects in my head are:
- Making function packages modular and easier to mix/match: [EPIC] Unify Function Interface (remove
BuiltInScalarFunction) #8045 - Getting started guide for new users (who want to use DataFusion in their project)
- Improving performance of aggregation - especially high cardinality groupings, where we still lag state of the art by 2x
Thus, if you are interested in contributing to DataFusion and are interested in a fast turn around time I would recommend looking into bug fixes / test improvements / documentation or the projects named above
If you propose adding new functionality, the review cycle will likely be longer. You can make it a shorter cycle by looking at the comments on other recent PRs and following the same model (e.g. ensure there are tests in sqllogictest for example, the CI passes, includes documentation, etc)