Comparison of cache efficiency among conda pixi and uv

[cc636489]

Hi everyone,

I recently began exploring different Python package managers, focusing on three in particular: Conda, Pixi, and uv. I conducted an experiment comparing the total installation time with and without cache across different package size groups: small-sized packages (e.g., loguru and protobuf), medium-sized packages (e.g., numpy and pyarrow), and large-sized packages (e.g., tensorflow-gpu and pytorch-gpu). The results are as follows:

1. Conda
   

2. Pixi
   

3. uv
   

More details:

1. For Pixi and Conda: all the testing packages from both package managers are set to "conda-forge" channel. I made the installed packages from Pixi and Conda the same as much as possible.
2. For UV: Since the installed packages are from PyPI index, so they are different comparing to Conda and Pixi.

Question:

1. Based on my observation of the cache structure layout, it seems that Pixi optimizes the Conda cache rather than directly reusing it. Is this correct?
2. Is the cache efficiency of Pixi for small, medium, and large-sized packages performing as expected? I want to confirm whether I might have overlooked any configurations Pixi currently offers for optimization.
3. One thing stood out to me: a noticeable amount of time is being spent on cache validation for large-sized packages (refer to the table below). Is this behavior expected as well?
   

Many thanks in advance!

[wolfv]

Hi @cc636489 thank you so much for the benchmarks. I am missing some information: what OS are you on, and what pixi / uv / conda versions did you use? And how many cores does your system have?

We recently improved the speed dramatically by making the validation much faster and the linking of files, too. This was shipped in pixi 0.39.4.

Looking forward to hear from you!

[cc636489]

Hi @wolfv,

Happy new year!

Here is the information below:

• OS system: Rock Linux release 8.10 (Green Obsidian)
• CPU: 20 Cores, 125GB RAM, 2.8GHz (Shared compute node by around 30-40 users.)
• Pixi version: 0.39.0
• Conda version: 24.9.2
• uv version: 0.5.6

And it's very exciting to know that the overall performance of Pixi has been speeded up dramatically in the recent few releases! Do you think the latest version of Pixi would be a good candidate for me to redo some of the measurements?

Thanks,
Chen

[wolfv]

Yes, definitely :) Please try 0.39.4

[cc636489]

Hi @wolfv,

With the latest Pixi release (version: 0.59.4), here is the cache efficiency updated result:

And the time spent on cache validation is dramatically reduced, here is the breakdown:

I'm very excited about such improvement! Thanks for your tip!

I have follow-up questions:

• Based on my observation of the cache structure layout, it seems that Pixi optimizes the Conda cache rather than directly reusing it. Is it the case?
• Is there any other configuration suggestion to further improve the cache efficiency? Currently, what I use is the default configuration with pixi init followed by pixi add, while playing around the PIXI_CACHE_DIR variable to change the cache location. In the future, we are looking into integrating Pixi programmatically into our workflow, enabling users to fast create projects at runtime concurrently.

Thank you in advance for your help!