Disc Usage during Repository Sync
While performing a repository sync recently I noticed that the WORKING_DIRECTORY was of a much much larger size than I expected.
It is currently assigned 150GB, yet failed to sync 3 repositories (I believe RHEL7 Source, RHEL7 Debug, and possibly RHEL7 Server) due to running out of disk space.
I was of the understanding that the working directory was not intended to hold the entire repository till completion, but only the individual unit's being transferred, on a successful transfer, that would then be ingested and end up as an artefact and stored in the artefact's path.
I believe an additional side affect of this, is that though units have been downloaded cleanly, and I believe could have been converted to an artefact, this has not been done, and a complete re-download is required as evidenced by the fact that on an additional sync the same behaviour occurs.
Maybe this would be the case any way, unless the artefact is converted to content, I'm unsure of those specifics.
Updated by lmjachky over 1 year ago
I made a couple of experiments and noticed that our pipeline is not really creating artifacts from temporary files along the way during the sync process. This is caused by the parameter
minsize (https://github.com/pulp/pulpcore/blob/354383883032277e7a1f7dc7ddf2dc0a5bc40fad/pulpcore/plugin/stages/api.py#L84) that retains the pipeline from saving artifacts unless the queue is filled up (https://github.com/pulp/pulpcore/blob/eda2c890214a76e6f6ffa18cc939d04273b1fa13/pulpcore/plugin/stages/artifact_stages.py#L229).
Furthermore, all the temporary files are being deleted only when the whole sync process ends. Because of the parameter delete in
NamedTemporaryFile (https://github.com/pulp/pulpcore/blob/354383883032277e7a1f7dc7ddf2dc0a5bc40fad/pulpcore/download/base.py#L122), we should handle the removal manually by ourselves. The working directory (temporary files stored under a specific task) is cleared as a whole only when the sync process reasonably finishes.
Also, another problem is that in some cases pulp_rpm do not handle the removal of temporary repodata files at all. When I synced from the http://mirrors.sonic.net/epel/playground/8/Everything/x86_64/os/, the downloaded repodata (3.9GB) were not removed even when the sync process successfuly ended. It seemed to me that I never reached this return statement: https://github.com/pulp/pulp_rpm/blob/f14be05d58f06e794676a5222b443cd9d082f031/pulp_rpm/app/tasks/synchronizing.py#L428, weird. Yet, the sync proceeded without failures.
So, I propose to examine some of the default parameters we changed in https://github.com/pulp/pulpcore/pull/440/files and call
os.unlink right after we create artifacts. I am attaching some of the advised changes in the description and unassigning myself.
Updated by dalley over 1 year ago
- Status changed from NEW to ASSIGNED
- Assignee set to lmjachky
- Sprint set to Sprint 99
@Lubos, This is important work and we want to get to this soon (discussed in the 3 month planning meeting), but nobody has had time to pick it up yet. If you could pick it back up when you return next week that would be excellent
Updated by lmjachky over 1 year ago
The commit https://github.com/pulp/pulpcore/commit/887dafa189639e310b7bd78ab96adbb188aa5a66 partially resolved the problem. Since it was merged, all temporary directories are properly cleared. Therefore, the statement "in some cases pulp_rpm do not handle the removal of temporary repodata files at all" is not true anymore.
However, the disk space is freed only when a task finishes. So, temporary files are still not being removed gradually through the sync, while the task is running. I am investigating more possibilities of resolving this issue.
Added by Lubos Mjachky over 1 year ago
Move downloaded files instead of copying them
Before this commit, the condition that checks whether pulp's temporary files are stored within /var/lib/pulp was never met because it looked for files starting with /var/lib/pulp/media (MEDIA_ROOT). MEDIA_ROOT is however no longer pointing to /var/lib/pulp.