while (true) do; How hard can it be to keep running?

At Google we have more than a handful of servers and must leverage our administration time as effectively as possible. Between custom in-house software and off-the-shelf daemons, there are many parts to running a reliable, distributed, redundant service. Most fundamental is running the software and keeping it running. Through reboots, crashes, upgrades, downgrades, bugs, canaries and outages, myriad forces conspire to end your process and keep it stopped or worse, keep it alive but not functioning.

There exists init, upstart, rc scripts, cron, at and more that provide mechanisms to run programs unattended, but each of them can fail in different ways. When you have dozens or hundreds of servers they will fail in many different ways. I will discuss the obvious and not-so-obvious failure modes of popular packages like upstart and cron, as well as how we’ve worked with and around them to ensure that when we run a daemon it stays running. Some special emphasis will be given to how virtual hosts create new challenges that can trip up launch strategies and services written for bare metal.