proj.x86_64-darwin build failing

[bdj]

Describe the bug

https://hydra.nixos.org/job/nixpkgs/trunk/proj.x86_64-darwin

The build is failing on darwin since the update to 8.1.1 (https://github.com/NixOS/nixpkgs/pull/138479)

Notify maintainers

@dotlambda @vbgl

[dotlambda]

@bdj Can you check whether adding

preCheck = ''
  export TMPDIR="/tmp"
'';

makes at least one of the two failing tests pass?

[Artturin]

this is the tempfile creation logic in proj_context_test.cpp

static std::string createTempDict(std::string &dirname, const char *filename) {
    const char *temp_dir = getenv("TEMP");
    if (!temp_dir) {
        temp_dir = getenv("TMP");
    }
#ifndef WIN32
    if (!temp_dir) {
        temp_dir = "/tmp";
    }
#endif
    if (!temp_dir)
        return std::string();

    dirname = temp_dir;

    std::string tmpFilename;
    tmpFilename = temp_dir;
    tmpFilename += DIR_CHAR;
    tmpFilename += filename;

    return createTmpFile(tmpFilename) ? tmpFilename : std::string();
}

[risicle]

I can't reproduce the failure on macos 10.15

[risicle]

I think we agreed in chat that we should disable these specific tests on darwin because none of us can reproduce this failure outside hydra, so I'm going to prepare a PR to do that but I think we need to keep this issue open until that has gone through and we can check that hydra's output is identical to our own builds. Otherwise we could be missing the fact that it's a genuinely broken output that's being built on hydra rather than just a dodgy test going wrong on hydra.

[veprbl]

I can reproduce failure on Mojave

[r-burns]

Oh, maybe that's it? I can't reproduce on catalina, and I've seen failures for other packages which are due to the builders using mojave.

[toonn]

I've been looking into this and it's because of a hardcoded assumption on the number of allowed open file descriptors on Linux AFAICT. On macOS the limit used to be 256 instead of Linux's 1024, don't know whether that changed.

[veprbl]

Do we have enough information to report the issue upstream? https://github.com/OSGeo/PROJ/issues

[toonn]

I'm looking into this. OPEN_MAX from stdio.h is the canonical source of the maximum number of file descriptors that can be open simultaneously but according to SO it's deprecated on Linux?

[veprbl]

What about getrlimit(RLIMIT_NOFILE)?

[toonn]

That's what I'm doing now but it does give up a little POSIX compatibility so I don't know how upstream will feel about it.

[risicle]

So.. #145371 deals with the CApi.open_plenty_of_contexts failure, but what about GridTest.VerticalShiftGridSet_gtx?

[toonn]

@risicle, I'm not sure about that, it doesn't seem like an actual failure. There's weirdness with the test suite where some other tests are sometimes affected by failures.

[risicle]

:crossed_fingers: