Strawberry DataLoader at module level: cache leaks across requests (stale data) and 'RuntimeError: Event loop is closed' on another event loop; create loaders per request in get_context
Create DataLoaders inside the per-request context, not as globals. With strawberry.asgi.GraphQL, subclass it and override get_context:
class App(GraphQL):
async def get_context(self, request, response):
client = NessieClient()
return {"request": request, "response": response,
"customer_loader": DataLoader(load_fn=lambda ids: batch_customers(client, ids)),
"merchant_loader": DataLoader(load_fn=lambda ids: batch_merchants(client, ids))}and in resolvers await info.context["customer_loader"].load(self.customer_id). Pass extensions=[QueryDepthLimiter(...)] to strawberry.Schema, not to the GraphQL app.
What goes wrong with a module-level loader = DataLoader(load_fn=...) (verified with strawberry-graphql 0.243.0): the loader binds the event loop lazily via get_event_loop() on first use and keeps its cache for the life of the process. Under a single uvicorn loop, request 2 got request 1's cached value without calling the batch function: stale balances when the loader wraps a banking API such as Nessie v2, plus unbounded memory. In scripts or tests that call asyncio.run() per query, a cache miss on the second loop fails with RuntimeError: Event loop is closed (and 'coroutine dispatch_batch was never awaited'). A hand-rolled loader that grabs the loop or creates Futures at import time typically just hangs until the client times out.
Verify batching with a request counter in the HTTP client: in one measurement, accounts{customer} for 2 accounts went from 4 REST calls to 3 (one batched customer fetch). The saving grows with fan-out, and merchants repeated across transactions dedupe within a request.