Clemens Vasters  🇪🇺 ☁ 📨, Messenger

@clemensv @aphyr So it would be possible if you sacrificed availability, correct?

@kellogh @clemensv tldr is no. It's not enough for the queue system itself to be linearizable because there's no safe commit (delivery) time

@aphyr @clemensv Does this whole thing ride on ordering? Is "exactly-once somewhat-ordered" possible?

@kellogh @clemensv ignore distribution entirely for a minute. I want to call a function and know, if I crash, whether it was invoked or not.

@kellogh @clemensv to do this you need to make the function and the journaling to disk *atomic*. That's why it's so hard to do exactly once.

@aphyr @clemensv ok, that makes sense, I think. Still trying to cement that into my mind…

@kellogh @aphyr people have managed to fool themselves for long into such assurances under the assumption that "the computer" won't fail.

@clemensv @aphyr Right I get that computers fail, networks fail, GC/hiccups look like failures, etc