Stitching composition result is not predictable

[n1ru4l]

The composition of stitching projects is not predictable.

In stitching the composition result is affected by the order of the service SDLs passed to the stitching implementation.
Within Hive, we always load the service SDLs ordered by their creation date. By doing so the order of the services is unpredictable, as the service's position within the list can change based on when it was last published. If there are conflicting type definitions within multiple services this can lead to unexpected composed schemas.

Example:

# service A
type Notification {
  message: String
}

# service B
type Notification {
  message: String!
}

If B is behind A, in the composed schema Notification.message will be of type String!.
If A is behind B, in the composed schema Notification.message will be of type String.

graphql-hive/packages/services/schema/src/orchestrators.ts

Lines 61 to 63 in c6568d1

	const { allStitchingDirectivesTypeDefs, stitchingDirectivesValidator } = stitchingDirectives();
	const parsedStitchingDirectives = parse(allStitchingDirectivesTypeDefs);
	const stitchingDirectivesNames = extractDirectiveNames(parsedStitchingDirectives);

Since we detect schema changes based on diffing composed schemas the service order change can lead to unexpected schema changes being reported, which are unrelated to the schema being published, but related to the changed order of the services that was triggered by a publish.

This confuses users of Hive and makes the predictability and usage of the composed schema (and services CDN endpoint) unsafe.