ApolloTracingExtension regression populating resolvers field.

[coady]

Describe the Bug

import asyncio
import strawberry
from strawberry.extensions import tracing

@strawberry.type
class Query:
    @strawberry.field
    def node(self) -> str:
        return ''

schema = strawberry.Schema(Query, extensions=[tracing.ApolloTracingExtension])

for i in range(2):
    result = asyncio.run(schema.execute('{ node }'))
    assert result.extensions['tracing']['execution']['resolvers'], i  # second time fails

outputs

    assert result.extensions['tracing']['execution']['resolvers'], i  # second time fails
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
AssertionError: 1

Only the first request populates resolvers as expected. This worked in previous version.

System Information

• Strawberry version (if applicable): >=0.240

Additional Context

Upvote & Fund

• We're using Polar.sh so you can upvote and help fund this issue.
• We receive the funding once the issue is completed & confirmed by you.
• Thank you in advance for helping prioritize & fund our backlog.

[coady]

The regression occurred in 0dcf23d, which changed how extensions are created. It appears a new extension instance is created on each request, resolve is called on the old instance, then get_results is called on the new instance. So this should be affecting other schema extensions as well; any which maintain state in resolve.

[erikwrede]

Hey @coady,
thanks for reporting this and investigating where the regression occured. Looking deeper into the issue, I see the following conflict we introduced:

The Middleware Manager is cached, at the same time it caches all resolver functions, including the extension resolvers wrapping them. Resolver functions are bound to a specific extension instance, and recently we started creating a new extension instance for each request. That means that for each request after the first, the resolver function we are calling is still bound to the first-ever schema extension instance.

My suggestion: let's create a custom MiddlewareManager subclass that injects the current instance of each extension into the resolvers. Alternatively, we can just disable caching for now and take the performance penalty. Tagging @bellini666 and @DoctorJohn for a second opinion. The relevant code is graphql_core.middleware @ L46 and strawberry.schema._get_middleware_manager