Francesco Cesarini

"When you are writing a massively complex highly distributed system, you need more just than a language, Erlang is just a programming language. You need architectural patterns, you need to start thinking in term of how you distribute your system."

"By learning them, and understanding their paradigms, and by programming in a functional way, you will end up being much more productive and youre becoming much better programmers."

"A system which is strongly available means that it is not reliable because of the reliability you might need to take down the machines. On the other hands if you dont worry about the reliability of your data, as long as youve got a value which is not too old, you can go in and distribute it, and scale that way."

"All languages will have side effects. They will have some form of mutability. The whole goal is to try to decrease this mutability as much as possible."

"What is the point with OTP? First of all, you get much more less code […] Less bugs […] Much more solid and tested code. Much more deep learning curves […] Performances will be slightly impacted due to the layers."

"What the Sunway TaihuLight, Raspberry Pi and Parallela board have in common? It is the heterogeneous multi-cores. So you have different types of cores, and thats the future of the computers we will working on today. The shift in multi-cores is inevitable."

"What you need to do is to make sure your threads become immutable. So you can change a data within a thread, but if you got two threads communicating with each other, you need to mimic this whole message passing approach. You need to copy the data from one threads to another, avoiding having shared state."

"There is two ways basically to deal with concurrency. Mutable state, you must have shared memory, you must have threads, you must have mutexes when you access critical sections. If you are dealing with immutable state, what you do is processes will not share memory. The only way for them to share data is through a message passing that needs to copy the data from one process to another. Theyll each have their own copy. So, two completely different paradigms, two completely different results."

"Even if your program is highly parallel, your bottleneck will still be the virtual machine, because you still have schedulers, you still need to process and handle and schedule your different processes."

"You distribute for scale and you replicate for availability."

"When you call a function with no side effects, what that means is, every time you call that functions with the same set of values, you always get back that same result. Thats immutable state. Mutable state, when youre dealing with shared memory, you call a function with a set of values, you dont know what value youre gonna get back. So you got determinism versus non determinism here."