The compressor/decompressor classes from bz2 and lzma modules rely on __init__() for initialization, but it is not guaranteed to be called. Method calls on an uninitialized object crash: >>> from bz2 import BZ2Compressor as C >>> c = C.__new__(C) >>> c.compress(b'') Segmentation fault (core dumped) I see two ways to fix this: 1) Move some initialization (notably, for "lock" field) to __new__() and add initialization checks to other methods. This should be backwards-compatible. 2) Move all initialization to __new__(). Since compressor/decompressor classes are not subclassable, it'll break only code than repeatedly calls __init__() on the same object. The simplicity of the fix might outweigh the necessity to support such code. (However, in 2.7, classes in bz2 *are* subclassable; lzma is not present in 2.7). Which way is more preferable?
I think we usually went with option 1 when we fixed similar issues in the past. See also issue 23224 for the same problem in *Decompressor classes of lzma and bz2 modules. It looks like the attached PR to that issue went with option 2: PR 7822. Perhaps we can combine this and issue 23224.
I somehow failed to notice #23224 when I searched the tracker. You're right, it's the same, and, moreover, PR 7822 fixes problem with both compressors and decompressors (though it includes tests only for the latter for some reason). I think that this report should be closed as duplicate, but should we also change the title of #23224 to be more general?
We can definitely make the title of that issue more descriptive. Feel free to change it -- IIRC, you don't need additional permissions to change the title of an issue. Thanks!
