@andreimatei writes:

I believe @asubiotto you're looking at some draining issues where the integration with some balancer isn't working right; I've stumbled upon the signals we're sending clients on draining, and I think they're wrong. Maybe this explains why the balancer isn't playing nice.
TL;DR I think we're sending the wrong message type to notify clients of the draining.

The protocol docs says this:

It is possible for NoticeResponse messages to be generated due to outside activity; for example, if the database administrator commands a “fast” database shutdown, the backend will send a NoticeResponse indicating this fact before closing the connection. Accordingly, frontends should always be prepared to accept and display NoticeResponse messages, even when the connection is nominally idle.

However, our code is this:

We're not sending a "NoticeResponse"; we're sending a plain error packet. I believe they're different.
Besides potentially not talking the language that balancers and drivers understand, I think this might also break clients - in my testing of other stuff I've noticed that lib/pq, for one, freaks out when it receives an unexpected error (i.e. when it gets an error but it's not in the context of running the specific types of commands that permit errors). On the other hand, clients apparently need to accept NoticeResponses at all times.

Does this sound right?