[R] Test segfault on roundtripping with metadata

[paleolimbot]

Describe the bug, including details regarding any error messages, version, and platform.

The last PR that touched metadata that I know about was #35039 (not sure if that is the root cause of this, though).

Link: https://github.com/apache/arrow/actions/runs/4744364907/jobs/8425173205?pr=35219#step:7:25535

Output:

Start test: Table with metadata roundtrip

 *** caught segfault ***
address 0xb8, cause 'memory not mapped'
An irrecoverable exception occurred. R is aborting now ...
Segmentation fault (core dumped)

Component(s)

R

[nealrichardson]

Looking at GHA logs, https://github.com/apache/arrow/actions/runs/4679740818/jobs/8290179008#step:7:28473 is the first time I see this on main. So at least #35074 or before.

[nealrichardson]

Also worth noting that the test in question is https://github.com/apache/arrow/blob/main/r/tests/testthat/test-python.R#L114, i.e. roundtrip between R and Python. And given the verbosity of the test output, it is this line https://github.com/apache/arrow/blob/main/r/tests/testthat/test-python.R#L116 that is crashing because we don't get to the assertion on L117, and it is not likely L115 because we have that line elsewhere in the test suite and it is not crashing.

  pytab <- reticulate::r_to_py(tab)

[paleolimbot]

Very possibly related is this one: #34897 . In that example the crash happens on exit (but it also involves a Python--R interaction).

[paleolimbot]

Hmm...reverting #35039 does seem to do it and the timing checks out; however I don't know why and haven't been able to replicate the segfault locally (also, I may have just gotten lucky the four times I re-ran the check that keeps segfaulting). Trying conda nightlies locally might be a good next step....I think that CI job links to a shared libarrow.so between R and Python which is a little unique in terms of real-world setups.

[paleolimbot]

I managed to reproduce the segfault on the PR reverting #35039 which rules out that (phew!).

[nealrichardson]

I'm now consistently seeing segfault on the flight tests (which also use pyarrow), e.g. https://github.com/apache/arrow/actions/runs/4832158272/jobs/8610556710#step:7:28324