Open h-vetinari opened 1 year ago
We cannot unvendor jemalloc as the build is with special options. Unverdoring it would have an impact on the default allocator.
Would it be possible to just build a variant of the jemalloc
package with options needed?
Would not count on ucx
being supported on other ~architectures~ OSes. Would also like to understand why that is relevant
Would also like to understand why that is relevant
I have a maximalist approach to packaging the bindings offered by a given library -- if the library offers integration for it, why shouldn't we package it?[^1]
Users rely on conda-forge because a lot of these packages are really hard to build from source, so if a binding is not supported by us, it's effectively unusuable (even moreso if the distributed wheels also turn it off).
[^1]: whether to break a package into separate outputs is a separate question
Sorry I should clarify, what I mean is why are other OSes relevant? Asking since ucx
likely will remain Linux only (possibly macOS at some point and basically no chance of Windows support)
Sorry I should clarify, what I mean is why are other OSes relevant? Asking since
ucx
likely will remain Linux only (possibly macOS at some point and basically no chance of Windows support)
Not relevant indeed, unless they become: a) available for the respective platform & b) are supported by arrow on that platform.
From looking at the list of third-party deps, we could still do better with:
Switch to shared builds for (if/once available):
Bindings:
libtensorflow
, if at all, because it's a very heavy dependency to pull in)