Auto-restart web dynos regularly (#2074)

Procfile

@@ -1,4 +1,14 @@
 release: ./heroku-release-phase.sh
-web: gunicorn server.app:app --preload
+# gunicorn runs multiple worker processes based on the WEB_CONCURRENCY
+# environment variable (set by Heroku automatically based on dyno size).
+# https://devcenter.heroku.com/articles/python-concurrency#common-runtime
+# At the time of writing, we use Standard-2x dynos, which results in 4 workers.
+# To tame slow memory leakage that we've observed but haven't solved, we
+# configure gunicorn to restart each worker after a certain number of requests
+# (with some random jitter). This threshold was set by observing how long it
+# took for memory to rise to an unacceptable level during peak traffic, then
+# counting the number of requests in that interval and dividing by the number
+# of workers.
+web: gunicorn server.app:app --preload --max-requests 1000 --max-requests-jitter 50
 worker: python -m server.worker.worker
 slack_worker: python -m server.activity_log.slack_worker