TimelyDataflow / differential-dataflow

An implementation of differential dataflow using timely dataflow on Rust.
MIT License
2.54k stars 183 forks source link

Demonstrate merge effort proportionality #409

Closed frankmcsherry closed 9 months ago

frankmcsherry commented 10 months ago

This PR demonstrates how one might effect "proportionality" in idle merge effort. The method reduced is use to indicate whether there is more work to do, and whether idle merging should re-schedule the arrangement operator. This PR changes the logic from "if any work at all exists" to "if any work exists between proportionally sized batches". This is not meant to be a perfect rule, nor a perfect constant, but a starting point to experiment with a reduction in CPU use when there is not a substantial amount of work to do.

There is a further reduction one could make to only look for proportionality between the largest non-empty batch and those below it, which would defer work until there was at least some proportional reduction in the total size of the arrangement, rather than among intermediate batches.

@antiguru @teskje