Bug #3824
Updated by Tom Clegg about 10 years ago
Cf. commit:a5819ec1e48fba90658fcf676ffc50c1f216d484 and commit:0eb5711ade0f74e556b0a1c10909dbf0bdecb63f
Some refactoring of that section of crunch-job may be in order here.
Requirements:
* JOB_WORK and TASK_WORK must not be the same directory.
* JOB_WORK and TASK_WORK must exist when a crunch script starts.
* JOB_WORK and TASK_WORK (and other crunch-specific directories) are *not* required to exist in the docker image being used. (A design goal is that non-Arvados-aware docker images can be used to run jobs. Expecting special tmp directories won't fly. But we might have to make some reasonable assumptions like "docker image has a world-writable /tmp".)
Probable approach:
* Run as much stuff as possible inside the docker container. E.g., the "build script" should run inside the container. It should have no trouble creating the required directories.
* Something like commit:a5819ec1e48fba90658fcf676ffc50c1f216d484 but fix whatever failed to propagate log messages, and make sure to create CRUNCH_SRC in the container
** @docker run -a stdin -i@ seems to be necessary to get stdin through to the container. See #4044 notes.
** We probably want to use @docker run --sig-proxy@ too.