Ask HN: People who work on V8, how hard would it be to adapt V8 to do this?
A JIT compiler that can work across multiple machine, true Scaling.
I think we have microservices backwards. What we are doing should be the job of a program. It should be possible for a JIT compiler to automatically find code that is a bottleneck and spawn it on another server + provide remote connection to heap/stack for spawned code. In fact, have shared VRAM where all shared heap/stacks can live.
- It'd simplify debugging
- you'd get by with just monoliths
- lot less need to monitor
- no need for gRPCs, etc.
- highly simplified programs
If V8 could provide nodejs with this it'd be a game changer. A JIT compiler that can scale automatically. I believe, and I've no experience in compilers, the hardest part would be sharing stack/heap between code in efficient manner. But program should do the math and figure out if performance gained by 'outsourcing' execution is more than latency and extra processing introduced.
First off, distributed memory has been tried (e.g. in https://en.wikipedia.org/wiki/Amoeba_(operating_system) which is famous for its relationship with Python). It’s extremely hard to do in a performant and coherent way. It basically didn’t work.
Second off, automatically distributing computations implies automatically adding the necessary synchronization. That’s super hard/maybe impossible in general.
Third, IMO there’s value in human readable APIs for services. I think it makes systems easier to debug since you have clear boundaries and (should) have simple constraints for each service.
I think, excluding the JIT and optimization, you just described worker nodes and worker pools [0] (?).
Another frame, can be, openfass serverless for certains part of code.
2cents, will keep the JIT for someone else, interesting topic
[0] https://medium.com/coinmonks/visually-understanding-worker-p...