Each search profile (everything, extensions, skins, etc.) listed are backed by individual instances of hound, that are controlled by
hound-$backend systemd units. Hound is a go application that is run in a Docker container using an upstream-provided image (TODO: build our own image). A flask web application proxies requests to the individual hound backends and injects in our customizations. It runs through the
hound_proxy systemd unit.
Everything is puppetized, so puppet will ensure all of the systemd units are running. Note that all of the backend units depend upon the proxy, so restarting the proxy will restart all of the backends.
The larger backends can take up to 15 minutes to start. The start up process can be rather memory and CPU intensive (especially on a truly cold start on a brand new instance), so the wait.py script tries to only allow one backend to start up at once.
The health check page should be helpful in seeing whether the hound backends are actually down, or just starting up still.
The configuration for each backend is written by the
codesearch-write-config systemd timer. Each backend's unit is configured to restart every 24 hours, so it picks up the new configuration.
Adding a new search profile/backend
write_config.pyto generate configuration for the new backend
app.pywith a label for the search profile
- Example: gerrit:565424
hieradata/labs/codesearch/common.yamlwith the new backend and the next numerical port
- Example: gerrit:567190
The codesearch patch must be merged before the puppet one. Also ensure that the
codesearch-write-config systemd timer has run before the puppet patch is merged.
Adding a new instance
Format /srv, then apply puppet role and run puppet, switch over web proxy.