I don't know what they are but I think most people want a single invocation of np.dot in a single thread to use multiple worker threads (as described in https://scipy.github.io/old-wiki/pages/ParallelProgramming) by pushing the work to a thread-aware matrix library.
Usually, this will be safe because np.dot only reads from its arguments, and produces a fresh new array to hold the result.
I just checked the documentation and found there is an out=argument to pass a destination array. (In any case there are other numpy functions that do modify their argument.) If you pass the same array to that parameter in two different threads then I don't know what happens. Maybe numpy locks the GIL to write its output. Maybe it produces C-style undefined behaviour (this would be my guess ... and my preference).
> I think most people want a single invocation of np.dot in a single thread to use multiple worker threads
Maybe? I'm not convinced that's what I want! But it's pretty irrelevant either way to my point; np.dot was just an example. My real point is that lots of CPU-heavy Python functions are implemented in C and release the GIL, so in practice parallelism with Python is often possible.
99% of people want this: np.dot and others use all the cores on your system by default. This accelerates the common use case, which is individual user on a multicore machine who wants single result ASAP. Then, there's an env var that lets you configure the underlying matrix library to use only a single (or maybe another number) of cores, and you work with the multithreaded/multiprocess libraries to run many independent calculations (possibly on shared read-only source data); then that scheduler has a tunable set to saturate the cores or IO on the system.
This is what most BLASses use and it's how data analysts and supercomputer folks can get along.
> I've worked in this field awhile.
This statement undermines your point more than it supports it. There is no "this field" for numpy. Just working in some field that uses it is no qualification for knowing whether users as a whole would like it multithreaded automatically. "Data analysts" and "supercomputer folks" no doubt include many users but certainly not all and possibly not even most (for all we know).
> 99% of people want this: ...
I can believe this might be true. I don't think I ever claimed otherwise. In fact I said "maybe"! I only claimed that I don't want this.
I will make one more claim though: I don't think there's any way that you could know this, even if it's really true.
But yes, I agree and already knew that it's how BLAS libraries work by default, including OpenBLAS which is what the prebuilt numpy wheels use. Even then, as you're probably already aware by the sounds of it, that's only for sufficiently large matrices. If you're just using small matrices and vectors (e.g. in 3D to represent real-world coordinates) then processing will be done directly in the calling thread (because the overhead of using multiple threads would be greater than the benefit).
Actually, you can use Numba to get the same effect while still writing Python. Numba allows you to simply apply a decorator to critical Python code and it is JITted into C-like performance. I do this in my code, and it has worked great for my uses.
I have definitely worked on Python projects in the past that were easily parallisable and worked absoluately fine. I simply created threads with threading.Thread, passed messages around with queue.Queue and used some standard Python modules to process those messages. The messages were independent enough from each other that I didn't have any issues with marshalling access to them, while being large enough that the overhead of mulithreading was still outweighed by its benefits. Rewriting the whole thing in C++ (etc.) would have made negligable difference to performance.
I can certainly imagine a program similar to what you describe, where there are lots of messages that operate on a single monolythic data structure, which couldn't easily be made to work with true parallelisation in Python.
You probably do not want to do that in the first place. The reason that the GIL still exists is that the overhead of fine-grained locking for Python datastructures would negate the speedup from multithreading.
Other language (like GoLang) do support parallelism but they have a more advanced runtime and don't use a GIL.
In the case of Python, the single-threaded implementation was already there, the implementation that removed the GIL in favor of fine-grained locking failed to deliver the goods.
If you want to actually take advantage of parallelism, you want data structures that are amenable to it, which is basically sections of arrays without a lot of potential read/write conflicts over multiple threads. If you're at that point, you probably want to use C/C++/Rust anyway, you don't want speed up something that's already 100x slower (the interpreter) by parallelizing it. Python offers that with C-extensions like NumPy.
Not if you're reading most of the time and are using a modern concurrent GC.
Python does support parallelism, so long as you're calling into a C library to do the work, because they (usually) release the GIL while doing the work. That sounds like a cop out but using C libraries is very common to do CPU intensive work anyway. Examples include numpy (e.g. you can call np.dot from multiple threads in Python and they will genuinely use multiple cores) and all the rest of the scipy stack, and even modules in the standard library (e.g. when using zipfile, even to process an in-memory buffer, the GIL is released).