Allow extended filtering on relationships

[merkys]

This PR addresses #437 by allowing fully-fledged filtering on relationships. Prior to it, filtering was allowed, but only on id and meta.description of relationships. Now all properties of related entries should in principle be allowed to filter on, to arbitrary depth (OPTIONAL).

This introduces a slight backwards-incompatibility by abandoning special treatment for description. Thus queries which previously matched/mismatched, but were perfectly legal, now will return HTTP 400 for all standard entry types except files where description property is defined since v1.2.0. I am not sure how severe this incompatibility is - if we want to avoid it, we will have to mandate the special handling of description by re-routing it to meta.description of a relationship instead of entry's property.

Fixes #437.

[rartino]

I'm confused. Even if we disregard that the meta subfield of the JSON:API relationship only makes sense in JSON:API, what in this PR explains why the meta subfield accesses the meta of the relationship, rather than the meta of the linked object?

[ml-evs]

I wholeheartedly support this, and indeed have been breaking the spec a bit myself to allow filtering on DOIs of references via this exact mechanism. This would be straightforward to enable in optimade-python-tools, as we currently hvae to disable filtering on fields other than ID in joins.

The only issue I see here is touched on by @rartino, previously we would allow filtering on the relationship description (which nobody ever used, and the filter was never implemented afaik), but somehow this is a breaking change that removes this functionality. Since we do not have a top-level description field of an entry, this could be kept in unambiguously, though slightly annoying for implementers, I think it makes sense (and we could never standardize a field called description for each entry in this case).

[ml-evs]

I think we actually need to revise this query example, since "name" is the required field for a references author and lastname is optional. How about a filter on DOI?

Or (in addition to this example) something more standardized, give me all references that correspond to structures of a given formula (or composition range)?

e.g., all literature references for Tantalum: /v1/references?filter=structures.elements HAS "Ta"

[merkys]

Query on all structures coming from the same DOI sounds good. One more suggestion of a bit more complicated query would be to filter for structures of certain anonymous formula related to publications appearing in year 2024.

[merkys]

I have added the said examples in 85c0a94, can you give them a look and mark as resolved if OK?

[ml-evs]

See my main review comment, but I think we have to keep this still

[merkys]

Yes, this has to be discussed. There are two options: retain full backwards compatibility by either encoding the special meaning of description or forbidding properties bearing such name, or scrap the special treatment of description altogether. This PR does the latter.

[merkys]

The only issue I see here is touched on by @rartino, previously we would allow filtering on the relationship description (which nobody ever used, and the filter was never implemented afaik), but somehow this is a breaking change that removes this functionality.

A somewhat drastic move would be to say that this design choice was a bug 😅 In a sense this would not be false, as meta for references object is JSON:API-specific, and we are trying to keep a generalized approach as much as possible.

Since we do not have a top-level description field of an entry, this could be kept in unambiguously, though slightly annoying for implementers, I think it makes sense (and we could never standardize a field called description for each entry in this case).

files entry type has top-level description property since v1.2.0, which complicates things a bit.

[rartino]

I don't think filtering on the human readable description of a relationship is particularly crucial. However, the "role" of a relationship discussed in #523 is crucial to be able to filter on. (E.g. "does this calculation have an output structure that contains sodium").

[merkys]

However, the "role" of a relationship discussed in #523 is crucial to be able to filter on. (E.g. "does this calculation have an output structure that contains sodium").

Right, I agree that #523 needs further discussion. Nevertheless #523 should not block this PR.

[rartino]

Since this PR asks to change how the dot operator works, I think we should try to peek forward enough so we do not paint ourselves into another corner.

Are there good reasons to not let the relationship still be the 'relationship object' (or strictly in OPTIMADE a list of relationship objects) and then let the target of the relationship object explicitly live under, e.g., 'target'?

To clarify, we could adopt a model of the structures relationship in an optimade calculation to work like:

"structures": [
  {
    "description": "Initial structure for relaxation",
    "role": "input",
    "target": <here lives the structure entry>
},
{
    "description": "Final structure after relaxation",
    "role": "output",
    "target": <here lives the structure entry>
}

This would mean that a query for all calculations that have an output structure that contains sodium could be expressed as:

calculations?filter=structures.role:structures.target.elements HAS "output":HAS "Na"

(As I type this up, I realize that we are actually missing the construct with the "inner" HAS for zip list constructs in our grammar, but that seems like an oversight that we should fix, given that we support e.g. the full set of fuzzy string operators in that position.)

Edit: note, this means that @ml-evs's query on DOIs now would be, e.g.:

structures?filter=references.target.doi HAS "10.1103/PhysRev.140.A1133"

[merkys]

Are there good reasons to not let the relationship still be the 'relationship object' (or strictly in OPTIMADE a list of relationship objects) and then let the target of the relationship object explicitly live under, e.g., 'target'?

I am not sure if I am getting this right, so bear with me. Do you mean retaining entry's relationships as they are (v1.2.0) and then including explicit copies of the related entries inside entry's "attributes"? If so, I think explicit inclusion might be redundant as JSON:API's "included" will cater for that already. Moreover, we may encounter arbitrary depth issues, which JSON:API avoids by only providing immediate relationships. Nevertheless such approach would indeed greatly simplify the specification of dot operator in filters.

[ml-evs]

I have had a query related to this and think it would be worth discussing again at an upcoming meeting. For my implementations I am still breaking compliance by allowing filtering on references.doi...