Inconsistent Marshalling Order in `client/llb` about buildkit HOT 6 CLOSED

I have opened #4706. I'd like to see if it passes CI first before full review, as I've had some trouble running the full test suite locally

from buildkit.

Actually, I think I looked at it wrong. The code you have pointed out only seems to change the marshaling order, so the order of elements in the definition array. The actual input order for each vertex, and therefore the content addressable digest for each operation seems to have correct guaranteed ordering already.

We can still make it consistent, but it should not have any effect for the builds at all. On the daemon side these are all read into a map and keyed by their digest, so individual ordering has no impact.

from buildkit.

This is not intended and can be directly fixed. Iiuc this does not affect cache as the indexing of inputs is consistent. But it may have an impact on parallel requests where some nodes need to do extra work to compute cache keys instead of just being deduplicated in LLB graph.

from buildkit.

Got it, thanks for the prompt response here! Just to clarify, were you planning on fixing this? I do have a fix I was looking at and can open a PR

from buildkit.

If you have a PR already, then open it. If not, then I can carry this as well.

from buildkit.

I believe this can be closed now due to #311.

from buildkit.