execute: Correctly report missing root type error

[IvanGoncharov]

cc: @yaacovCR

[IvanGoncharov]

@yaacovCR Probably it should be just errors without data, what do you think?

[yaacovCR]

The spec is unclear, sometimes it says that if execution fails because of missing information, you should just errors, sometimes it says that if execution starts, you should get some data, and it is internally contradictory on whether request execution is considered execution for these purposes, definitionally it calls it execution, but the parts of execution seem validation like, and so in the subsections it says to return only errors and that “execution” is considered to start only with operation execution @benjie has a bit in the glossary about this maybe he has thoughts

[yaacovCR]

Because of that ambiguity in requests, where it is called execution even though later not considered execution, I am inclined to say the same is possibly true in the execution operations section, so we can return just errors

[yaacovCR]

But the spec should be clarified, I think, so kind of doesnt matter what we choose, except if we can anticipate how we would like this to go, in terms of usefulness for clients, I don’t even know what the right answer is, in general, documents should be validated, so this case shouldn’t even come up, so I would say should be similar to any missing field where document not validated where I assume null is returned so data should be null here

[yaacovCR]

Although currently the subscribe code I think just returns errors, that’s the better option there because data for subscribe belongs only in payloads. So maybe consistency with subscribe for execute means data should be skipped for all the operations

[benjie]

From Response/Response Format/Data:

If an error was raised during the execution that prevented a valid response, the data entry in the response should be null.

From Executing Requests:

3. If operation is a query operation:
   • Return ExecuteQuery(operation, schema, coercedVariableValues, initialValue).

From ExecuteQuery():

2. Assert: queryType is an Object type.

---

Assuming that this error is triggered by the assert above, then it's part of the execution process and thus the data should be null. Interesting that this is a should not a must, actually. Anyway; since we're the reference implementation we should follow the should.

[yaacovCR]

Just so I understand that is because the assertion is made in the operation execution algorithm section, and not the request execution section? Or is either section considered execution? Variable coercion section says:

If a request error is encountered during input coercion of variable values, then the operation fails without execution.

That is ambiguous itself seems to mean operation execution, as opposed to request execution.

Anyways, we should align subscription behavior to whatever we do for queries

[benjie]

I think you’re right; there is ambiguity there. I’ll have a go at adding some clarity.

[benjie]

It turns out that this is addressed a couple paragraphs lower down:

https://spec.graphql.org/draft/#sec-Errors.Request-errors

Namely: ExecuteRequest() can raise "request errors" which are deemed to occur "before execution begins". So "execution" refers to neither "operation execution" nor "request execution" but to the phase of request/operation execution that occurs after the last possible point at which a request error could be raised. Note this could be during the CreateSourceEventStream() algorithm.

[benjie]

Here's my attempt to clarify this with a non-normative note: graphql/graphql-spec#894