Stop gracefully on SIGTERM about casper

[ArnaudBrousseau]

@drolando: nice find 🥇

That explains why we're seeing spurious 503 spikes. graceful_mesos_shutdown is not the root cause as I initially thought, it's just what exposes this bug.

[ArnaudBrousseau]

So far I haven't able to solve this problem, but here's a repro:

1. make dev
2. curl -o /dev/null -iv -H 'X-Smartstack-Destination: yelp-main.internalapi' -H 'X-Smartstack-Source: spectre.main' -H 'X-Ctx-Tarpit: yelp-main.internalapi' 'localhost:32927/status'. This triggers a slow request to internalapi (timeout of 15s.)
3. (in a different session): docker ps | grep abrousse | cut -d' ' -f1 | xargs docker stop --time 30. This shuts down the Casper container by sending SIGTERM first, then SIGKILL after 30s.

Expected:

The request to /status finishes then nginx/docker shuts down gracefully

Actual:

Connection is aborted and we get "curl: (52) Empty reply from server", docker shuts down at the same time

---

Some things that might be useful to note here

• process hierarchy:

dumb-init start.sh
|___start.sh
    |___nginx (master)
        |___nginx (worker)

• From dumb-init's docs: "In its default mode, dumb-init establishes a session rooted at the child, and sends signals to the entire process group. This is useful if you have a poorly-behaving child (such as a shell script) which won't normally signal its children before dying."
• From nginx's docs: SIGTERM causes a fast shutdown, and SIGQUIT causes a graceful shutdown (that's what we want)
• To check that SIGQUIT was the right thing I hoped onto the container, fired up a slow request to /status and ran /usr/local/openresty/nginx/sbin/nginx -s quit. The slow request finished, then nginx exited and the container shut down.

[drolando]

We can use dumb-init to change the SIGTERM in a SIGQUIT

dumb-init --rewrite 15:3

[ArnaudBrousseau]

We can use dumb-init to change the SIGTERM in a SIGQUIT

Nice. Just verified, it totally works (CMD ["dumb-init", "--rewrite", "15:3", "/code/start.sh"])