3 ms·
Pause all writes > let 16 to catch up > resume writes on 16 Isn’t that….. downtime? Unless you mean downtime to be only when reads are also not available.
by darth_avocado 2y ago
Pause all writes > let 16 to catch up > resume writes on 16
Isn’t that….. downtime? Unless you mean downtime to be only when reads are also not available.
- stopachka 2y agoWe count downtime if a service is unavailable and drops requests. In this case, since the pause took about 3.5 seconds, we were able to service all requests.
- dboreham 2y agoBy that definition you can have any length pause and still claim zero downtime.
- nijave 2y agoOnly if nothing has timeouts. Likely components somewhere in the stack have timeouts so you're bound by that
- darth_avocado 2y agoInteresting. If that was the case, I’m curious why you had to write custom code. Would RDS proxy not work? The way we usually upgrade is create a new upgraded cluster, replicate all data to new cluster, pause writes and redirect RDS proxy to the new cluster. And that usually takes a few seconds.
- dspillett 2y agoNot if the process doesn't take long, so the most the user (or consuming service) sees is a slower response than usual. That I would class as a temporary performance degradation. If the degradation is significant it might still not be acceptable, of course, but I'd not call it downtime.