Clickbench Q6 and Q23 regressions

[connortsui20]

There's a noticeable improvement in several of our benchmarks because of this which is nice! Though I did notice regressions in Clickbench Q6 and Q23 that someone should probably investigate?

Originally posted by @connortsui20 in #5879 (comment)

[AdamGS]

planning to dig deeper tomorrow

[godnight10061]

Thanks for checking the benchmarks. Do you have a link to the specific run/comparison where Q6 and Q23 regress, plus the before/after numbers? If it reproduces consistently, I’m happy to dig in.

[connortsui20]

@godnight10061 If you go to https://bench.vortex.dev/?group=Clickbench and scroll down you can look at our continuous integration benchmarks for Clickbench (which is a data point per commit). You can look at the vortex-bench crate to see how to run these locally (though since @AdamGS updated the benchmarks recently I do not actually know the command to run them anymore)

[godnight10061]

Hi @connortsui20,
I authored PR #5879 that introduced these regressions. I’m currently away from my computer but wanted to acknowledge this ASAP.
Would someone be able to help investigate the Q6/Q23 slowdowns? I suspect it might be related to head-of-line blocking in LazyScanStream’s synchronous prepare() calls during MultiScan polling, but this needs proper profiling to confirm.
I’m happy to review any fixes or follow up when I’m back online. Thanks for catching this!

[connortsui20]

Oh @godnight10061 don't worry too much about the regressions! It was in general a speedup for a lot of queries, so I think we're pretty happy about that. I think @AdamGS will take a look at it sooner or later and maybe he can help you?

[AdamGS]

benchmarks can be run either using the vx-bench tool in the repo (uv run vx-bench) or by running the duckdb benchmarks binary directly (cargo run --bin duckdb-bench --release).

[connortsui20]

Closed by #5906