Preserves source of union members and enum values in supergraph

[pcmanus]

While composition allows unions and enums to differ between subgraphs (at least under certain conditions), the supergraph currently was not preserving which subgraph defined which union member/enum value, and in particular the query planner made the (sometimes incorrect) assumption that unions and enums were defined the same in all the subgraphs in which they are defined.

As shown in #2256, this was leading to genuine issues in the case of unions, where the query planner was generating invalid subgraph queries. I am not sure this created similar issues for enum values due to a combination of how the merging rule for enum work and the fact that values of enum are not types, but it is nonetheless a bad idea to run the query planner with incorrect information on the subgraph it generates queries for, and this can get in the way of other toolings that wants to use the supergraph (for instance, this gets in the way of #2250).

To fix this, this patch ensures the supergraph preserve the information regarding which subgraph defines which union member and enum values in the supergraph by adding 2 new dedicated "join spec" directives.

[codesandbox-ci]

This pull request is automatically built and testable in CodeSandbox.

To see build info of the built libraries, click here or the icon next to each commit SHA.

[pcmanus]

I'll note that this PR is currently based on top of the #2279 PR. This is because that later PR was already adding the code to introduce a new 0.3 join spec, which this patch needs too. I can try to separated the aforementioned code so this can be merge first if we really really want, but as I do think we should make an effort in getting those 2 PRs in the same release, exactly because they both need new things from the join spec (and, all other things being equal, it's nice to avoid burning new versions too often), so maybe we just review #2279 first and then get to this.

[clenfest]

Code LGTM, but can you add some tests to make sure that it's correct? Also for extractSubgraphsFromSupergraphs, ideally there would be tests that exercise both 0.3 and 0.2 functionality to make sure both paths work.

[pcmanus]

can you add some tests to make sure that it's correct?

Mind being more specific about what you feel is inadequately tested? The patch adds support for correctly keeping the source of union members and enum values, and the patch currently:

• Adds the test from the description of #2256 which tests that the user-visible consequence of the bug fixed here is solved.
• Un-skip this test which had been added in a previous patch but was failure as a consequence of this bug.
• Adds new tests that more specifically validates that both union members and enum values are correctly re-extracted back from the supergraph.

I did just added a modification to the maint test checking a supergraph output so it includes unions and enum values, but the details of the supergraph are ultimately an implementation detail, so I don't want to over-index testing that.

ideally there would be tests that exercise both 0.3 and 0.2 functionality to make sure both paths work.

Well, it's a bit tricky. The 0.3 version is well exercised by virtue of being the current version, but I agree we'd ideally want test ensuring that we can read 0.2 version properly. Except that while we preserve the reading of older versions, we don't preserve the writing of them, so to test this we need to generate supergraphs with older version and ensure they can still be read correctly on newer versions. It's probably doable, but a bit involved if we want this to not be too manual and unmaintainable. In any case, I'd like to tackle that question separately from this particular issue as it's more general. I've created #2297 with a possible initial, but I'm not entirely convinced yet.