Question - Restoring from check-point

[andrewluetgers]

Serverless environments keep functions warm for some undefined period of time after the initial cold start. Optimizations take advantage of this time period of punctuated statefulness. For example a database connection can be initiated once and reused, an file can be loaded into memory an re-used vs being loaded for every execution.

Can your technique be used to extend this statefulness beyond this undefined warm period with some kind of check-pointing scheme? It seems it would work well for the file-io example not so much for the database connection.

[ostenbom]

Interesting, I haven't thought of that kind of checkpoint before. The checkpoints I've been thinking about are more along the lines of sharing libraries between different functions. Aka if both function A and function B use numpy, then only restore to the point where numpy was loaded after A so that loading B is faster.

In general I'd want code loaded into a function container to remain as stateless as possible, since state is expensive, but your idea of having database connections "ready" for functions is interesting!

An added point is that when testing this system I was using a warm time of around a few seconds rather than a few hours.