ch-fromhost has a complicated and messy task, and this is made more difficult by some drawbacks that have both been designed in and accumulated over time:
The script is 500 lines of fairly dense POSIX sh (not even Bash).
It does not support r/o images, e.g. SquashFS which are important for scaling (#286).
Lack of standardization. Vendors are starting to provide OCI hooks that do a similar task, and we can’t use them.
Adds a separate step to the workflow (#1361).
For nVidia, uses an external program (nvidia-container-cli) that must be installed and kept up to date (#940).
This issue is to make this feature easier to maintain. Options I can think of are:
Rewrite the script in Python.
Drop the script in favor of OCI hooks (whether done directly by the ch-run or a Python helper).
Do nothing.
I marked the priority “high” because I think this may be important for supporting AI workloads well.
ch-fromhost
has a complicated and messy task, and this is made more difficult by some drawbacks that have both been designed in and accumulated over time:The script is 500 lines of fairly dense POSIX sh (not even Bash).
It does not support r/o images, e.g. SquashFS which are important for scaling (#286).
Lack of standardization. Vendors are starting to provide OCI hooks that do a similar task, and we can’t use them.
Adds a separate step to the workflow (#1361).
For nVidia, uses an external program (
nvidia-container-cli
) that must be installed and kept up to date (#940).This issue is to make this feature easier to maintain. Options I can think of are:
ch-run
or a Python helper).I marked the priority “high” because I think this may be important for supporting AI workloads well.