Closed seibert closed 3 years ago
I should note that if I leave out --forked
, the test passes.
Some bizarre findings. The segfault is happening in the 2nd call to mlalgo.pagerank
. A single call is always fine, but the next call causes the segfault.
The problem goes away if any of the following are true:
irb.util.del_sparse_tensor(prev_score)
_pagerank = None
prior to 2nd call)scipy.sparse.csr_matrix()
anytime prior to the compilation of pagerank
--forked
is required to see the problem. Without that, test_left_bipartite_project_and_filter
is always called first and makes a scipy csr_matrix, somehow affecting the way pagerank is compiled and avoiding this issuepr = _pagerank(graph, damping, tol, maxiter)
, then simply continueSuffice it to say that I'm completely baffled. For now, I'm going to remove the del_sparse_tensor
line.
There seems to be an intermittent segfault in the pagerank test. So far I can only reproduce it on Linux with:
(
--forked
comes from thepytest-forked
plugin, which allows pytest to run each test in a sub-process for better isolation.) I get no useful output besides: