what workflow creates so many negative dentries?

I ran into this in production on Kubernetes. Symptoms were that pods (processes) were being evicted for no clear reason, but Kubernetes said it was due to memory pressure.

The problem? Negative dentries filling the filesystem cache!

Turns out, the blame was a health check a script for that process. It called curl every few seconds to check for health, and curl was adding 5mb to the cache (negative dentries!) every time it ran. It took a while but then suddenly poof! Process evicted due to using too much memory.

It’s a bug that still shows as open. Here’s two comments I made on it with extra details:

Report: https://github.com/kubernetes/kubernetes/issues/43916#iss...
Workaround: https://github.com/kubernetes/kubernetes/issues/43916#iss...