Use Mamba install

[jakirkham]

Heroku is running out of memory during the build. Switch back to Mamba for the install to see if that helps.

Checklist

• Pushed the branch to main repo
• CI passed on the branch

[jakirkham]

Here is the Heroku log

[jakirkham]

@beckermr success! 🎉

[jakirkham]

Testing a re-render to make sure things are working: conda-forge/ucx-split-feedstock#162

[jakirkham]

Seems to be working

[mbargull]

FYI, you most likely ran into conda/conda#13541 (comment) which caused a sharp memory usage spike with the latter being avoided in conda>=24.3 via conda/conda#13628 .
But even with conda=24.1 you'd probably not run into that issue -- for conda-forge at least -- since the code for repodata.json.zst support does not take that problematic code path.

IOW, the issue was most likely due to having conda<=23.11 in our installers and once we update to conda>=24.1 (or better, conda>=24.3), the peak memory usage should be about the same whether conda or mamba is used.

[hmaarrfk]

we are probably due for a new release of miniforge then.

[mbargull]

Yes, but no particular rush, IMO. The bug mentioned above has been present in conda>=4.12,<24.3, so not a particular new thing ;).

[jakirkham]

Explains why we didn't see it before. This was using 4.11.0 previously (with Mamba)

We just updated ( #580 ) 😬

A new release would be welcome, but agree not a rush. We have cobbled together a working solution to tide us over until then

[jakirkham]

Started updating Miniforge in a follow up PR ( #598 ). Though maybe we can go to 24.3 if that is ready ( conda-forge/miniforge#574 )

[jakirkham]

We have now upgraded to Miniforge 24.3.0 and are using Conda again for installs ( #598 ). Thanks all! 🙏