What is the current best practice for issues with

MPLCONFIGDIR

? Should modules be updated to set this, and if so what?

What's this about?

Several modules set this. Best practice is always to use a folder relative to the work directory.

Agreeing with the BP

So should make a modules PR if you discover this warning/error in a module then?

matplotlib’s internal default is already relative to

$HOME

, which nextflow sets to the workdir. Did you find a situation where this doesn't work? https://matplotlib.org/stable/install/environment_variables_faq.html#envvar-MPLCONFIGDIR

Not sure where you have read that but Nextflow (

25.10.4

at least) definitely doesn’t explicitly set

$HOME

to the

workDir

. You can see which env vars it is setting for the

slurm

executor here (all other env is dropped, thus whatever the container sets for

$HOME

) is used:

PATH="$PATH" ${TMP:+SINGULARITYENV_TMP="$TMP"} ${TMPDIR:+SINGULARITYENV_TMPDIR="$TMPDIR"} ${NXF_TASK_WORKDIR:+SINGULARITYENV_NXF_TASK_WORKDIR="$NXF_TASK_WORKDIR"} SINGULARITYENV_NXF_DEBUG="${NXF_DEBUG:=0}"

Now yes,

--no-home

should in theory mean that it’s set to something else but I do not see this in reality, the container correctly opens in the

PWD

but the

HOME

is set to the host home still. Even if I run

singularity exec

and explicitly have

HOME=$PWD

I still see that

$HOME

is my host

$HOME

in the container. Possibly there is some global option or configuration here which is causing this to be overridden? The only way so far I have been able to force it is by setting

APPTAINER_HOME

Also worth noting that it is actually running

apptainer

not

singularity

Oops, that's embarrassing. I am so used to dealing with an older nextflow pipeline that we're currently porting to nf-core, and that pipeline did explicitly set

$HOME

inside containers to the workdir. I realize now that nextflow itself doesn't touch

$HOME

. Sorry about that.

No I think this behaviour seems like a problem with my apptainer tbh but also something other people have experienced given that many modules do set MPLCONFIGDIR explicitly

What problems do you see? Are you getting the

Matplotlib created a temporary config/cache directory at...

message, or something worse?

Basically yeah it tries to create one in a random tmp location. Not always a problem the pattern which is widely used in nf-core to solve this works well to remove annoying warnings and reduce any issues

I can see the reasoning for why apptainer chose their defaults (like the home behavior) given their target use case / audience. Environment isolation just wasn't the main goal. At least they do let you opt in to isolation though.