Ref T4571. When a database goes down briefly, we fall back to replicas.
However, this fallback is slow (not good for users) and keeps sending a lot of traffic to the master (might be bad if the root cause is load-related).
Keep track of recent connections and fully degrade into "severed" mode if we see a sequence of failures over a reasonable period of time. In this mode, we send much less traffic to the master (faster for users; less load for the database).
We do send a little bit of traffic still, and if the master recovers we'll recover back into normal mode seeing several connections in a row succeed.
This is similar to what most load balancers do when pulling web servers in and out of pools.
For now, the specific numbers are:
- We do at most one health check every 3 seconds.
- If 5 checks in a row fail or succeed, we sever or un-sever the database (so it takes about 15 seconds to switch modes).
- If the database is currently marked unhealthy, we reduce timeouts and retries when connecting to it.
