Matplotlib useable on clusters with read-only resp. non-existing home directories on compute nodes. #2695
Conversation
on clusters whith readonly home directories on compute nodes. The standard allow on Linux to specify configuration and cache directories independent from the existence of a home directory. The functions _get_xdg_config_dir() and _get_xdg_cache_dir() respect this now. Function _get_config_or_cache_dir() allow now to use also on Linux an existing old style configuration directory "~/.matplotlib". If this does not exist, the configuration directory will be expected in the location specified by the XDG_CONFIG_HOME environment variable. If this location is undefined or it cannot be found resp. created, a temp. configuration directory will be created in the directory for temp. files.
|
Hi Thomas, yes the functionality o fboth PR is equivalent, and the implementation of get_xdg*_dir() in branch 1.3 even a bit more efficient. I close the PR for the master branch. |
|
Thanks! It looks good to me, but I am not familiar with this particular corner of the library or desktop path standards so I am hesitant to merge it. |
|
Just a bit background information for the potential merger/committer. Here is the link to the Freedesktop standard that I mentioned in the now closed PR. Compare http://standards.freedesktop.org/basedir-spec/basedir-spec-latest.html#variables paragraphes 2 and 8. The other part of the change was motivated by the comment of the function _get_configdir(), especially point 2: 2a. On Linux, if 2b. On other platforms, choose The comment seems to be a bit contradictionary: choose the old location and warn, and the next sentence says basically ignore it if XDG_CONFIG_HOME has been defined. I think one should choose as specified the old config and warn. This would not lead to a loss of configuration settings because a new config is written into a (potentially empty) directory specified by the environment variable. Due to our discussion here I now think that one even should warn that the existence of the old config and the environment variable at the same time is conflicting. Alternatively, if we want to stick with behaviour that a XDG_CONFIG_DIR definition has higher priority than to use an existing old config, I can change the pull request so that it only contains what is needed to make it running on the cluster. |
|
Here's the motivation for the current behavior: Many users have existing configuraiton in The current XDG-compliant behavior that we are moving towards is that if there is an I'm not sure what you're saying is contradictory about the comment. We use matplotlib never writes configuration files, so I'm not sure what you mean when you say "a new config is written into a (potentially empty) directory specified by the environment variable". |
|
Ok, thanks for the explanations. I agree that we should keep that.
The assumption came from the observation that at least the config directory is created during the module import. But right, it was always empty after my tests. I supposed that this was the case because I only tested small things for the purpose of the setup test. I propose that I change my commit in that way, that I only repair the crash due to the non-existing resp. non-readable home directory in the cluster environment and keep at the same time the current behaviour exactly. |
|
As discussed, this pull request is replaed now by an improved variant #2719 -> closing the request here. |
Details are in the comment of the commit.
A corresponding pull request for branch 'master' exixts: #2692