Why a lookup table? Why not just have the client remember the last backoff, and multiply it by a constant and apply a ceiling?

def getNewBackoff( oldBackoff: int): int {

newBackoff = oldBackoff * BACKOFF_CONSTANT

if (newBackoff > BACKOFF_CEILING) { newBackoff = BACKOFF_CEILING }

return newBackoff }

if you bit align the ceiling value you can replace the conditional with a bitwise and mask. If you use 2 as your exponent you can replace the multiplication with a bit shift left.

It’s also a good idea to introduce jitter, e.g. don’t set the new backoff to the old backoff * a constant, but add a random value such that the old value gets randomized around the desired new value- this helps prevent pile-ons when a bunch of clients noticed about the same time that the server became unresponsive; if they all have the same backoff algorithm then they’re all going to hit the server again at about the same time. It’s good to spread that out to try to avoid server overloading during restarts/cold starts/whatever.

The best results I ever got were when I used server side hinting to avoid pile-ons when the retries from the client were all clustered around the same time.

When my server got in an overloaded state, the auth path for clients short circuited and delivered an overloaded message with a randomized suggested retry value. The server can’t make the clients honor it (we could in my case because we wrote the clients) but if the clients do honor it you can spread out reconnects over a period instead of make them all point-in-time and dampen or prevent oscillating overloads.