There must be something funky with the stack-installed pyarrow (not pandas). When I pip install --user --ignore-installed pyarrow, then my test runs fine; when I remove that, it hangs with the signal 11.

Sophie Reed just hit a signal 11 issue trying to use the shared stack and my parquet stuff, too, so there's something that's definitely not right.

That sounds scarily like an ABI incompatibility creeping in somewhere, but it's very hard to tell where. This is why I'm scared of installing stuff into the stack using Conda.

I'll play around to see if I can get something which works, but this may be effectively insoluble within the constraints that Conda puts upon us.

I don't suppose you can pip install pyarrow instead of using conda?

Per DM-15494, this is now working in /software/lsstsw/stack. I don't propose to spend further effort updating old stacks unless there's a critical need, so I reckon we can close this issue. Tim Morton [X] (or anybody else) please reopen if you disagree.