It seems at least on Linux, setting the environment variable PYTHONFAULTHANDLER=1 (enabling a Python-builtin package as of recent versions https://docs.python.org/3/library/faulthandler.html) causes segfaults with PyCall. The MWE I've been using is
# script.jl
using PyCall
Threads.@sync begin
Threads.@spawn GC.gc(true)
Threads.@spawn GC.gc(true)
end
Run julia --threads=2 script.jl and its fine, but run PYTHONFAULTHANDLER=1 julia --threads=2 script.jl and segfault. Reproduced across multiple systems. More impactful I would say, even in single-threaded mode, PYTHONFAULTHANDLER=1 will crash the IJulia kernel any time you try to interrupt it. This is the situation I was in, I was on a cluster where the admins enabled this by default since its a really useful Python thing, and just lived without the ability to interrupt IJulia for months without knowing why, which I can attest really sucks.
Maybe it would make sense to unset this variable in PyCall.__init__ and issue a warning that we're currently incompatible? I could submit a PR.
It seems at least on Linux, setting the environment variable
PYTHONFAULTHANDLER=1
(enabling a Python-builtin package as of recent versions https://docs.python.org/3/library/faulthandler.html) causes segfaults with PyCall. The MWE I've been using isRun
julia --threads=2 script.jl
and its fine, but runPYTHONFAULTHANDLER=1 julia --threads=2 script.jl
and segfault. Reproduced across multiple systems. More impactful I would say, even in single-threaded mode,PYTHONFAULTHANDLER=1
will crash the IJulia kernel any time you try to interrupt it. This is the situation I was in, I was on a cluster where the admins enabled this by default since its a really useful Python thing, and just lived without the ability to interrupt IJulia for months without knowing why, which I can attest really sucks.Maybe it would make sense to unset this variable in
PyCall.__init__
and issue a warning that we're currently incompatible? I could submit a PR.