Our admin/internal APIs don't go via loadbalancer, so it's more likely that connection errors occur.

The change is general and not specific to our situation. But agreed, then it doesn't help us much.

Why did you opt for taking the same setting for different retries?

The option is called http_retries and the says Number of times xyzclient should retry on any failed http call..
As it stands, the code doesn't do that, it only retries on failed connections.
Separate values make more sense to me if we need the granularity to actually tune them do different values.

We wouldn't be able to disable one, if it becomes problematic.

We still can disable retries, just not separately. I would say that is good enough for fixing something which requires a fix.

By default, this only retries 503, which is probably fine for requests changing anything in Cinder, but for GET-requests, we might also want to retry on 500 - e.g. the DB restarting and thus requests failing with "internal server error". What do you think?

The retries are enabled for all requests including POST,PUT,etc. I rather would not like to retry that on the lowest level except for 503, as the APIs are not guaranteed to be idempotent.
I think, a 500 error is a "real" server error, which should be fixed on the application/db side.

E.g. fix the retry logic in the application-db api (oslo.db) to handle the restart better (more likely in a short-time frame) or fixing it on the db side that the API is zero-downtime (quite a bit of effort).