See my answer above (https://news.ycombinator.com/item?id=22031581) - the gist is that I set up GDB to break on a syscall matching the pattern of the offending close(31) (without also breaking on the zillions of legitimate close(31) calls generated by SQLite). Then I backtraced at that point and observed that this particular close wasn't from SQLite (which I'd already suspected), and the backtrace had several lines indicating that the close was from the GC finalization of a FileIO object.
The second step was to figure out which FileIO object was being GCed. I ran the program again and broke inside the FileIO __init__ function (which is written in C), conditioned on the fd being 31. From that breakpoint I did a Python backtrace (using the tools from https://github.com/python/cpython/blob/master/Misc/gdbinit) to figure out the Python code that was responsible for the allocation.
The third step was futzing around with the NamedTemporaryFile code until I could determine why it was leaking a file. This took a little while because I did not initially realize that it had been mocked. Once I did, the solution presented itself.
Comments
See my answer above (https://news.ycombinator.com/item?id=22031581) - the gist is that I set up GDB to break on a syscall matching the pattern of the offending close(31) (without also breaking on the zillions of legitimate close(31) calls generated by SQLite). Then I backtraced at that point and observed that this particular close wasn't from SQLite (which I'd already suspected), and the backtrace had several lines indicating that the close was from the GC finalization of a FileIO object.
The second step was to figure out which FileIO object was being GCed. I ran the program again and broke inside the FileIO __init__ function (which is written in C), conditioned on the fd being 31. From that breakpoint I did a Python backtrace (using the tools from https://github.com/python/cpython/blob/master/Misc/gdbinit) to figure out the Python code that was responsible for the allocation.
The third step was futzing around with the NamedTemporaryFile code until I could determine why it was leaking a file. This took a little while because I did not initially realize that it had been mocked. Once I did, the solution presented itself.