Arvados

Currently, if container_image is

• ...a docker hash, repo, or repo:tag, then we look up tag links and use the indicated collection
• ...a PDH, then we look in the indicated collection to see whether it has a single file called (sha256:)?{64-hex-digits}.tar. If so, we use that collection; if not, we try to look for tag links as if it were a docker repo:tag or hash (!).
• ...a UUID, then we look for tag links as if it were a docker repo:tag or hash.

It is possible to have a docker image with a name that happens to look like a PDH or UUID. In such cases, the image selected by the current code depends not only on whether a real collection with the PDH-like string happens to exist, but also the contents of that collection.

It would be more predictable to say

• If container_image looks like a collection UUID (".....-4zz18-..........")
  • ...and that collection is readable, then use that collection.
  • ...otherwise, fail.
• If container_image looks like a PDH ("hexdigest+decsize")
  • ...and that collection is readable, then use that collection.
  • ...otherwise, fail.
• Otherwise, treat it as a docker repo, repo:tag, or hash, and look up tags as usual.