runtime-spec
runtime-spec copied to clipboard
opencontainers
question about valid values runtime choose to support
In config.md, spec says Unless support for a valid value is explicitly required, runtimes MAY choose which subset of the valid values it will support.
I think there may will be a problem, if different runtime support differen subset. If so, we can't guarantee the portability.
I think this is against our goal of standardization.
I think the idea was to avoid failing a runtime as non-compliant if the host system didn't support a feature. But with the current wording, you can have compliant runtimes that use that language to not support most values described in the spec (everything except the ‘ro’ mount option on Windows 1 and ‘l3CacheSchema’ on Linux 2?) even when the host system supports the values.
I'd rather require runtimes to support at least the values the spec lists as valid (with runtimes having the option to support additional values as they see fit). I believe the “kernel doesn't support the feature you've asked for case” is covered by 3:
If the runtime cannot apply a property as specified in the configuration, it MUST generate an error and a new container MUST NOT be created.
Dropping the “runtimes MAY choose which subset” line would make using compliant runtimes more reliable, although we'd probably want to either inline our valid-values list (like we do for ‘ro’ in Windows mounts 1) or link to a specific version of an external doc (like we do for Solaris rlimits [4,5]). Linking to a floating external doc (like we do for Linux rlimits [6,7]) will lead to issues as values are added to and removed from the external doc.
There have also been concerns about linking to man pages for lists of valid values, because man pages may be an incomplete or incorrect reflection of what the kernel actually supports. I'd address those cases by either:
a. Submitting upstream patches to fix the man pages, and then linking the spec to the patched version, or b. Inlining the valid values in the spec so we don't need a normative link to the buggy man page.
With both of those being fairly straightforward, I don't think “man pages are unreliable” is a sufficient reason to avoid requiring runtime support for values that the spec lists as valid.
