Pandas uses _Py_HashDouble as the hash function for hashing floats. () in the 0.17.x and 0.18.x releases. This hash function is used within nogil sections within hashtable.pyx on both these releases. However the definition of this function as implemented on Python 2.x can call into the Python api depending on the values being hashed. This causes crashes if there is another thread using holding the GIL which is executing simultaneously. I see that master no longer uses this hash function. But can this fix be backported for the 0.17.x / 0.18.x branches?
Comment From: jreback
this is in 0.18.2 see the fix https://github.com/pydata/pandas/commit/f98b4b541e7a9d1b0d8f6674a84dc10080c2568b
we don't backport fixes
Comment From: deepankarsharma
Any eta on when 0.18.2 will be on pypi?
Comment From: jreback
0.18.2 was renamed to 0.19.0
0.19.0 is releasing shortly; the rc is available now https://github.com/pydata/pandas/releases/tag/v0.19.0rc1
Comment From: wesm
@deepankarsharma we are open to creating maintenance / bugfix / backport-only branches, but at the moment we do not have the committer bandwidth to support this. we would require invested corporate partners to help us do this