don't release GIL in Dataset.__init__ #1193
Conversation
|
Personally, I think it's a mistake for users to rely on the GIL to make netcdf "threadsafe". |
|
Understood @dopplershift. Perhaps the GIL may have been prevented some segfaults in versions before 1.6.1, but results still may have been incorrect. |
|
What is the advantage to maintain this part of the code and let users fix it instead vs merging this to safe guard bad use? My impression is that most folks opening issues don't even know they are abusing the GIL there. |
|
Previously, the library was inconsistently releasing the GIL, as in: some methods were already releasing it and were thread unsafe. (I've personally encountered crashes as a result). So to truly solve the problem and eliminate all potential crashes due to Python-based multithreading, you'd need to not only revert #1181, but #370 as well, and make sure ALL calls hold the GIL. Personally, I'd rather have the option for better performance for use cases that aren't trying to use multiple netCDF files, but doing other things concurrently in the interpreter. I do see that this change has been a bit disruptive, so I'm not going to fight for it. It's also not my library, so my opinion counts very little. 😉 |
potential fix for issue #1192