Use mimalloc v3

[chirizxc]

Same as astral-sh/ruff#27586

Close #19189

[chirizxc]

It might be worth make these as features to avoid such long conditions and their duplication in the file?

[charliermarsh]

This makes sense given that we've now merged it in Ruff, but I'll let @zanieb merge since they looked at the issue last time.

[zanieb]

My prior research on this was not compelling and showed memory use regressions that were hard to offset with configuration; it's a bit messy, but see #19743

[chirizxc]

My prior research on this was not compelling and showed memory use regressions that were hard to offset with configuration; it's a bit messy, but see #19743

On my machine, memory usage for ruff actually decreased (i used this script: astral-sh/ruff#27586 (comment))

[zanieb]

Ruff is really different than uv, we'd need benchmarks for this project.

[zanieb]

I can still reproduce the cold resolve memory regressions:

• home-assistant-core: 85.23 → 113.67 MiB, +33.37%.
• airflow: 177.20 → 229.36 MiB, +29.43%.
• bio-embeddings: 93.57 → 119.23 MiB, +27.43%.
• saleor: 89.87 → 113.76 MiB, +26.59%.

[chirizxc]

I can still reproduce the cold resolve memory regressions:

* home-assistant-core: 85.23 → 113.67 MiB, +33.37%.

* airflow: 177.20 → 229.36 MiB, +29.43%.

* bio-embeddings: 93.57 → 119.23 MiB, +27.43%.

* saleor: 89.87 → 113.76 MiB, +26.59%.

Could you share your script? I'll run it locally on my computer.

[zanieb]

Sure, it's just generated by Codex https://github.com/astral-sh/uv/blob/zb/pr-21014-mimalloc-namespace-bench-20260811/.github/scripts/mimalloc-high-confidence-rss.ps1

[chirizxc]

Sure, it's just generated by Codex zb/pr-21014-mimalloc-namespace-bench-20260811/.github/scripts/mimalloc-high-confidence-rss.ps1

Yeah, I think I was able to reproduce it locally

[chirizxc]

Sure, it's just generated by Codex zb/pr-21014-mimalloc-namespace-bench-20260811/.github/scripts/mimalloc-high-confidence-rss.ps1

As I understand it, you're running this benchmark on GitHub Windows Runner?

upd. If it's not too much trouble, just for the sake of the test, could you also add https://github.com/moggedhedien/rimalloc to this bench?

[zanieb]

I was running it on a Namespace VM

[zanieb]

rimalloc doesn't have Windows support, it seems

[chirizxc]

I've also tested several Rust projects using snmalloc, but it was usually slower