Circumvent GWLF-E SegFaults due to NumPy

[rajadain]

Overview

A recent release of NumPy introduced a new module _multiarray_umath. Unfortunately, there are some tools that depend on NumPy and pull in the latest version as a build dependency. This latest version is then replaced with the version specified in our requirements.txt, thus the module is not available at run-time causing the segfaults.

By installing NumPy separately before requirements.txt is gathered, we ensure that whatever was using the most recent version of NumPy in its installation now uses the pre-installed one.

We'll have to ensure that the version of NumPy in requirements.txt matches the one in Ansible.

For details see:

numpy/numpy#11871
https://stackoverflow.com/q/54153886

Connects #3089

Demo

Testing Instructions

• Download and extract this file: celery-segfault.zip

• Check out this branch and destroy and rebuild the Worker VM

• Transfer the files into the Worker VM, and execute test.py

  $ python test.py
  

  • Ensure there are no segfaults

• Compare the generated output.json with the included output.verify.json:

  $ diff output.json output.verify.json
  

  • Ensure they match

• Go to :8000 and create a MapShed project.

  • Ensure the modeling stages complete successfully

[mmcfarland]

What is the numpy version where this module was introduced? I'm curious, if it within a compatible release range, that we should just upgrade? Are you concerned with introducing breaking changes?

If we stick with this approach, rather than keep the numpy version in sync with requirements.txt, can numpy just be removed from that file altogether since ansible is now installing it?

[rajadain]

What is the numpy version where this module was introduced?

According to https://stackoverflow.com/a/54153887, it was numpy-1.16.0rc2 in January. The current latest release of NumPy is 1.16.1.

I'm curious, if it within a compatible release range, that we should just upgrade? Are you concerned with introducing breaking changes?

I know that GWLF-E and Ulmo both require NumPy, and may transitively bring in their versions. I think the amount of testing required to ensure that the new version works would exhaust the available budget, and this seems the most light-weight solution.

If we stick with this approach, rather than keep the numpy version in sync with requirements.txt, can numpy just be removed from that file altogether since ansible is now installing it?

Good observation. Currently Ansible only installs NumPy for the Worker. In App, it is installed via requirements.txt. I can update the App's role definition to install NumPy and remove it from requirements.txt so we don't have to keep track.

[rajadain]

Applied the suggested fixes, confirmed that it still works:

[hectcastro]

I had pretty much the same two questions. Since I didn't have to write them out, I used that bandwidth to create an issue to remember to upgrade NumPy (#3093). 👍

[mmcfarland]

Thanks for the updates. I ran through the test script and also was able to multi-year model after rebuilding.