proxy_read_timeout was set to 3600s (1h) from 600s (10m) to handle a PAWS specific case but seems to be potentially interferring with nginx releasing workers and resources. I can't find an example of anyone setting this to greater than 10m and based on our recent experiences with being unable to run nginx reload for graceful user impact I think we should reduce.
https://gerrit.wikimedia.org/r/#/c/256882/5/modules/dynamicproxy/templates/nginx.conf
Related incident: https://wikitech.wikimedia.org/wiki/Incident_documentation/2017-04-19_ToolsProxy