search

rapidsai / integration

RAPIDS - combined conda package & integration tests for all of RAPIDS libraries
Apache License 2.0
16 stars 66 forks source link

[FEA] Nightly build of latest stable #120

Open lmeyerov opened 4 years ago

lmeyerov commented 4 years ago

Is your feature request related to a problem? Please describe.

Issue https://github.com/rapidsai/cudf/issues/6096 caused an expensive fire drill involving core members across multiple projects

Assuming the current stable release breaks again in the future for one of many reasons, and stable releases are out for ~6w, a useful thing may be something along the lines of daily build tests of last 2 releases (3mo)

Describe the solution you'd like

I'm less clear on the relative value of the full gamut of going across CUDA etc versions. AFAICT, the 80% is probably around conda: Python verison x set of RAPIDS packages

kkraus14 commented 4 years ago

@lmeyerov This is a huge ask that would require a lot of resources both in engineering hours and compute resources to running all of the builds and tests where it unfortunately isn't feasible for us currently.

I think one of the challenges here is that the number of libraries in RAPIDS is continuously growing and as new libraries are added they're not necessarily as mature as say cuDF or cuML with respect to their dependency tree. Perhaps we should define some type of "graduation" process to be included into the rapids metapackage to make it a bit more of a "controlled" experience.

In the meantime, I think a middle ground here can be the ask that we treat conda solving issues of the current stable release metapackage with more attention.

lmeyerov commented 4 years ago

Yeah I get the concern of doing a full test suite run. As a lot of it seems to be the mess that is pydata conda deps, maybe a nightly docker run - conda install rapids=x, x-1 py=3.6-3.8 can go far, and timeout the solver after say 5min.

It's problematic enough that we publish + use stable base docker rapids builds for our side ( https://hub.docker.com/orgs/graphistry/repositories ) and release them every month or two, though that still leaves us in an occasionally sensitive spot in practice when there are CVEs, updating base deps, doing clean builds, etc. For folks less deep into the ecosystem, having public stable / recommended not able to conda install isn't good, that's normally expected to be only for nightly..

EDIT: re RAPIDS is continuously growing and as new libraries are added .... that sounds more like an alarm bell calling for more automation here to support the growth, as it'll happen more, not less, and hit more devs+users...

kkraus14 commented 4 years ago

This isn't a cudf issue so moving to the integration repo as that's where the metapackage is produced.