Cache Manifest: duplicate entries in a manifest ? Is a sha an identifier ?

Hello there, I'm currently trying to deeply understand what's in a cache manifest and I have a bunch of questions, feel free to just redirect me to a blog post or something, I couldn't find any that would answer these.

1/ In some builds a cache manifest can have more than one layer with the same digest, but I see that the code treats those as a single, same thing; for example in here :
https://github.com/moby/buildkit/blob/36c5550a84d5ef558c1b9ac733cba128feb37929/cache/remotecache/import.go#L76-L79

Here are some questions:

• Can a layer be identified by its blob (digest) ? Or should it rather be identified by its full path eg: $parent_digests + $digest ?
• what gives a digest ?
• if two digests are identical, can we only assume that the operation was the same ?
• Could duplicate digests cause some issues down the line ?
• Would it be a correct thing to sort of rewrite a manifest and merge the duplicate entries together ? ( it feels like no because then we would end up with an incorrect result and sometimes the result would be an invalid loopey dag )

2/ Same paragraph and same questions but for cache records instead of layers. A cache manifest can have more than one records with a similar digest.

3/ cacheManager.Query

• What will a cacheManager.Query actually lookup ? A digest ? but if the digest can be duplicate, won't that operation be incorrect then ? Or will it rather ask for the full path leading there ?

• I also saw that upon loading a cache manifest, the loaded digests are randomized here :

https://github.com/moby/buildkit/blob/36c5550a84d5ef558c1b9ac733cba128feb37929/cache/remotecache/v1/cachestorage.go#L47-L52

When the cache item has links. But, would it be possible to rename things before saving the manifest ? So as to have each entry be a unique one ?

Thanks :)

Start by reading cache/remotecache/import.go at lines 76-79 and cache/remotecache/v1/cachestorage.go at lines 47-52, then trace cacheManager.Query. The issue provides no test or concrete change to run. Done would require an agreed explanation or documentation covering duplicate layer and cache-record digests, lookup identity, and manifest rewriting.