My Rust as-of join is slower than Polars

TL;DR. In March 2026 I published a README table saying my Rust as-of join, FlowState, was 1.8 times faster than Polars (10 ms against 18 ms)¹. When I wrote a reproducible benchmark in October, the result reversed: in my runs on one Apple M4 laptop FlowState’s median was 13.3 ms against 9.9 ms for Polars on 1M left rows²³, and 194.9 ms against 135.8 ms on 10M left rows⁴⁵. I removed the March claim. Rebuilding the March kernel did not bring the lead back⁶, and I do not know why Polars is faster or what produced the original number.

Why this matters. A performance claim with no captured run behind it can sit in a README for months. On this workload the library was faster in batch mode, so anyone considering a custom join kernel for point-in-time ML features should measure against the library first.

What I believed

I believed FlowState would beat Polars on grouped as-of joins, and I wrote down why in the design doc on 2026-03-11: it skips the sort verification a general engine has to do, borrows symbol keys without allocating, joins symbol groups in parallel with Rayon, and splits large scans into parallel chunks⁷. I also believed pandas “falls over at 10M rows.”

What I did

The March README table was never backed by a captured run¹. On 2026-10-04 I wrote a reproducible benchmark that prints the hardware, versions and dataset, and checks every engine’s output against Polars before timing. I ran it through pow run. The first result showed Polars about 1.4 times faster, and pandas finished the 10M-row join in 0.76 s, so the pandas belief was also wrong⁸. I removed the March claim from the README. On 2026-10-30 I added the fastest and slowest run to the output and captured 10 runs⁹.

DuckDB’s ASOF JOIN did not finish at 1M rows and 1K symbols in my run, so it is opt-in in the script and not in the results here.

FlowState’s batch kernel is a merge-scan over timestamp-sorted inputs, parallelized with Rayon, exchanging data with Python as Arrow without copying. The benchmark times only the join: data generation, sorting and table construction are excluded.

What happened

Figure: median as-of join time, 10 runs per bar, whiskers are the fastest and slowest run. Apple M4, 32 GB. Run 2026-10-30-1412-asof-join-spread @ fdc202b. Generated by figure.py from the capture’s output.log.⁹

I measured the join on 1M left and 0.5M right rows, and on 10M left and 5M right rows, with 1,000 symbols each. Each figure is the median of 10 runs after 2 warmup runs, with the fastest and slowest run alongside.⁹

FlowState’s median was 1.34 times Polars’ at 1M rows¹² and 1.44 times at 10M rows¹³. In both sizes the ranges do not overlap: Polars’ slowest run was faster than FlowState’s fastest. FlowState was still well ahead of pandas.

An earlier capture on 2026-10-04, median of 5 runs with no spread recorded, shows the same ordering: 8.7 ms for Polars and 12.3 ms for FlowState at 1M rows¹⁴. The absolute times drifted between the two sessions (Polars went from 8.7 ms to 9.9 ms on identical code and data), so I only compare engines within the same run.

Why it failed

• Measured cause: none. I have not profiled either engine, and I have not tested whether any of the five reasons in my design doc hold.
• Why the March number did not reproduce: my journal suspected a stale March build of the Rust kernel or a different dataset⁸. I tested the first guess. I built the kernel and Python package at the commit that carried the claim (3c7486a) and ran today’s script against it with the same pinned libraries. FlowState was still slower: median 12.9 ms against 8.9 ms for Polars on 1M left rows⁶, 1.45 times Polars’ median with non-overlapping ranges¹⁵. A stale kernel does not explain the March figure.
• Still untested (speculation): a different dataset, and a Polars release in March that was slower than 1.38.1. The March table says only “Polars 1.x” and “Apple M-series”¹, and I have no record of how its numbers were produced. Today’s script and Rust toolchain are not the ones I used in March.
• Why Polars is faster today: I don’t know.
• What I’d test next: profile both engines on this workload, and run the same data against older Polars releases.

What I’d do instead

My view, from the README: FlowState’s value is not raw join speed against Polars. It is the pieces around the join: watermark streaming alignment, the SPSC streaming pipeline, partitioned replay and the temporal feature store. That is an opinion, and this benchmark does not test it.

Limits of this result

• What I didn’t test: other CPUs, inputs that are not already globally timestamp-sorted, real market data, a join tolerance, thread-count settings, and memory use. I recorded no energy or power data.
• Threats to validity: one laptop with no recorded power or thermal state; synthetic random-walk quotes; two sessions that disagree on absolute time.

Spec

What’s open

Whether the March figure came from a different dataset or an older, slower Polars release. I do not have the March run, so I may not be able to settle it.

Reproduce

git clone https://github.com/RyanJHamby/FlowState && cd FlowState
git checkout fdc202b
uv venv --python /opt/homebrew/bin/python3.14 .venv && . .venv/bin/activate
uv pip install maturin numpy polars==1.38.1 pandas==3.0.6 pyarrow==23.0.1
(cd flowstate-core && maturin develop --release)
uv pip install -e .
python -W ignore benchmarks/reproduce.py --pandas-scale --runs 10

Use a regular CPython 3.14 build. A free-threaded 3.14 build failed to install Polars 1.38.1 in my attempt¹⁶. I ran these commands from a fresh clone and the ordering held, with Polars’ slowest run faster than FlowState’s fastest at both sizes¹⁶. The raw log is repro-clean-env.log in the post’s draft folder.

Approximate cost to reproduce: about 1.5 minutes on the M4, including the Rust build, and no cloud spend¹⁷.

References

[1] FlowState repository, benchmarks/reproduce.py and results/ at commit fdc202b.

Reproduce: https://github.com/RyanJHamby/FlowState at fdc202b · Raw data: results/2026-10-30-1412-asof-join-spread/ at that commit; the March-build test is results/2026-10-30-1423-asof-join-march-build-3c7486a/ at 53084c2 Published: 2026-10-30 · Last updated: 2026-10-30 · Changelog AI disclosure: Drafted with AI (Claude) from my captured runs and from notes I wrote in the FlowState README, design doc and journal. The benchmarks were run on my machine through an AI assistant at my direction, and every number comes from a captured run or a cited file in that repo. Claim-by-claim verification is not recorded yet. Corrections: If something here is wrong, email ryan.j.hamby@gmail.com. I fix errors publicly in the changelog.

Changelog

• 2026-10-30: Published. Results at FlowState fdc202b and 53084c2.