Summarizing that discussion, we have a lot of issues/wishes. Some can be addressed as one-off and some need a unified reorganization of the runner comparison.
- Remove rows that impossible to not support (ParDo)
- Remove rows where "support" doesn't really make sense (Composite transforms)
- Deduplicate rows are actually the same model feature (all non-merging windowing / all merging windowing)
- Clearly separate rows that represent optimizations (Combine)
- Correct rows in the wrong place (Timers are actually a "what...?" row)
- Separate or remove rows have not been designed ([Meta]Data driven triggers, retractions)
- Rename rows with names that appear no where else (Timestamp control, which is called a TimestampCombiner in Java)
- Switch to a more distinct color scheme for full/partial support (currently just solid/faded colors)
- Switch to something clearer than "~" for partial support, versus ✘ and ✓ for none and full.
- Correct Gearpump support for merging windows (see BEAM-2759)
- Correct Spark support for non-merging and merging windows (see BEAM-2499)
- Lump all the basic stuff (ParDo, GroupByKey, Read, Window) into one row
- Make sections as users see them, like "ParDo" / "side Inputs" not "What?" / "side inputs"
- Add rows for non-model things, like portability framework support, metrics backends, etc
- Add versioning to the comparison, as in BEAM-166
- Find a way to fit in a plain English summary of runner's support in Beam. It should come first, as it is what new users need before getting to details.
- Find a way to describe production readiness of runners and/or testimonials of who is using it in production.
- Have a place to compare non-model differences between runners
Changes requiring engineering efforts:
- Gather and add quantitative runner metrics, perhaps Nexmark results for mid-level, smaller benchmarks for measuring aspects of specific features, and larger end-to-end benchmarks to get an idea how it might actually perform on a use case
- Tighter coupling of the matrix portion of the comparison with tags on ValidatesRunner tests
If you care to address some aspect of this, please reach out and/or just file a subtask and address it.
|Design for runner comparison with narrative intro for each runner||Open|
|Merge portability status into capability matrix||Open||Unassigned|