The current exponential back-off leads to subjectively long waits for the metrics to start flowing after the service has been down. (This is readily apparent during development, but should also be considered an issue for real users in the case of outages with the service.)

It might be fine just to back off linearly (1 sec, 2 secs, 3 secs, 4 secs, ...) instead of exponentially (1 sec, 2 secs, 4 secs, 8 secs, ...), or perhaps keep it exponential but just use a smaller power than 2.

An alternative might be to keep the current exponential back-off, but with a "plateau" at, say, 8 seconds, where it doesn't increment for a few minutes (1 sec, 2 secs, 4 secs, 8 secs, 8 secs, 8 secs, ...)

My preference is for the simpler options, like a linear increment, just for "guessability" in the wild. jonathan.balsano, john.morales care to weigh in?