~18% perf delta from "-dynamic" on query of GHA sample Parquet

[philrz]

The following delta (~79.0 ms vs. 64.9 ms) is on my Intel-based Macbook, but has been similarly reproduced on AWS EC2 instances.

$ aws s3 cp s3://super-issues/super-7152/ghasample.parquet . &&
  super -version &&
  hyperfine "super -c \"SELECT count() AS count FROM 'ghasample.parquet' WHERE actor.login='johnbieren';\"" &&
  hyperfine "super -dynamic -c \"SELECT count() AS count FROM 'ghasample.parquet' WHERE actor.login='johnbieren';\""

Version: v0.3.0-280-gbc683bdec

Benchmark 1: super -c "SELECT count() AS count FROM 'ghasample.parquet' WHERE actor.login='johnbieren';"
  Time (mean ± σ):      79.0 ms ±  12.9 ms    [User: 80.4 ms, System: 24.3 ms]
  Range (min … max):    64.5 ms … 130.3 ms    21 runs

Benchmark 1: super -dynamic -c "SELECT count() AS count FROM 'ghasample.parquet' WHERE actor.login='johnbieren';"
  Time (mean ± σ):      64.9 ms ±   1.8 ms    [User: 64.5 ms, System: 22.0 ms]
  Range (min … max):    61.6 ms …  72.6 ms    44 runs

Details

Repro is with super commit bc683bd and the GitHub Archive sample data in Parquet format as stored in the S3 URL shown in the repro above.

This delta was first observed in #7102 (comment) when verifying the fix for a wider perf delta that was originally tracked in that issue. When sharing the finding with @nwt and @mccanne, there was acknowledgement that some remaining delta was unsurprising since by definition there's "less work" performed then -dynamic flag is used. However, there also was consensus that this delta is wider than we might expect and that we might want to look into that at some point.

Claude's take

I asked Claude to do some profiling and put its findings in a Gist, should that save anyone time.