Brian Grant

Interesting failure mode I have seen a few times now: App runs in Docker on laptop, OOMs in #kubernetes. Turns out to be ...

... code that scales usage per-core or per-GB. It detects many cores (more than laptop) via /proc/cpuinfo, and tries to use them, but ...

... the container is limited to less-than-all the cores on the machine, so it blows up. Lesson: ...

... how to probe "available" resources is necessarily different from inside containers. This is under-recognized and under-developed ...

... We need new tools & techniques to learn what you can ACTUALLY use in your container, rather than what the machine has installed.

And unfortunately not necessarily that easy unless the container is setup in a particular way (e.g. read access to cgroup)