In multiproc mode, if one of the gunicorn workers gets killed because of a segfault or out-of-memory situation, it doesn't get a chance to clean up its .db files. These files are thus left in the multiproc folder, distorting the metrics.

Maybe this can be solved by moving the cleanup hook into the gunicorn "master" process. My suggestion is to solve this in the Python client instead, by deleting all .db files with dead PIDs before aggregating the metrics.