Querying an type for just `nodes { __typename }` returns no nodes when there are some

The bug

Using the local ElasticGraph test schema, run DEBUG_QUERY=1 bundle exec rake boot_locally. Then run this query:

query {
  parts {
    total_edge_count
    nodes {
      __typename

      # ... on ElectricalPart {
      #   voltage
      # }
    }
  }
}

It returns:

{
  "data": {
    "parts": {
      "total_edge_count": 20,
      "nodes": []
    }
  }
}

This is a nonsensical response: it says there are 20 results but none are returned!

If you uncomment the ... on ElectricalPart fragment, it works correctly:

{
  "data": {
    "parts": {
      "total_edge_count": 20,
      "nodes": [
        {
          "__typename": "MechanicalPart"
        },
        {
          "__typename": "MechanicalPart"
        },
        {
          "__typename": "ElectricalPart",
          "voltage": 120
        },
        {
          "__typename": "ElectricalPart",
          "voltage": 120
        },
        {
          "__typename": "ElectricalPart",
          "voltage": 220
        },
        # ...
      ]
    }
  }
}

Debug Queries

The DEBUG_QUERY=1 output shows the problem. For the first query, it's this:

[
  {
    "index": "electrical_parts,mechanical_parts"
  },
  {
    "size": 0,
    "track_total_hits": true,
    "_source": false
  }
]

For the 2nd query, it's this:

[
  {
    "index": "electrical_parts,mechanical_parts"
  },
  {
    "size": 51,
    "sort": [
      {
        "created_at": {
          "order": "desc",
          "missing": "_last"
        }
      },
      {
        "id": {
          "order": "asc",
          "missing": "_first"
        }
      }
    ],
    "track_total_hits": true,
    "_source": {
      "includes": [
        "voltage"
      ]
    }
  }
]

Notice the difference in size. I believe it has to do with the fact that no "real" fields are requested (just __typename) which causes ElasticGraph to not ask for any fields on the queried documents (no _source: {includes: [...]}) which in turn makes it ask for size: 0 since it doesn't think it needs any actual documents.

A case that works!

Notably, the query works if you use edges { node { ... } }:

query {
  parts {
    total_edge_count
    edges {
      node {
        __typename
      }
    }
  }
}

Returns:

{
  "data": {
    "parts": {
      "total_edge_count": 20,
      "edges": [
        {
          "node": {
            "__typename": "MechanicalPart"
          }
        },
        {
          "node": {
            "__typename": "MechanicalPart"
          }
        },
        # ...
      ]
    }
  }
}

The logged query shows why it works:

[
  {
    "index": "electrical_parts,mechanical_parts"
  },
  {
    "size": 51,
    "sort": [
      {
        "created_at": {
          "order": "desc",
          "missing": "_last"
        }
      },
      {
        "id": {
          "order": "asc",
          "missing": "_first"
        }
      }
    ],
    "track_total_hits": true,
    "_source": {
      "includes": [
        "__typename"
      ]
    }
  }
]

I thought that nodes { ... } and edges { node { ... } } work the same so the fact that they are different here is concerning. We should evaluate if there are any other ways they differ.

Also buggy on concrete types

It's a problem on concrete types, not just abstract types.

query {
  addresses {
    total_edge_count
    nodes {
      __typename
    }
  }
}

{
  "data": {
    "addresses": {
      "total_edge_count": 10,
      "nodes": []
    }
  }
}

[
  {
    "index": "addresses"
  },
  {
    "size": 0,
    "track_total_hits": true,
    "_source": false
  }
]

Interestingly, edges { node { __typename } } doesn't fix it here:

query {
  addresses {
    total_edge_count
    edges {
      node {
        __typename
      }
    }
  }
}

{
  "data": {
    "addresses": {
      "total_edge_count": 10,
      "edges": []
    }
  }
}

[
  {
    "index": "addresses"
  },
  {
    "size": 0,
    "track_total_hits": true,
    "_source": false
  }
]