The Fix
Upgrade to version 0.13.8 or later.
Based on closed Kludex/starlette issue #1022 · PR/commit linked
Production note: This tends to surface only under concurrency. Reproduce with load tests and watch for lock contention/cancellation paths.
@@ -1,3 +1,9 @@
@@ -1,3 +1,9 @@
+## 0.13.8
+
+* Revert `Queue(maxsize=1)` fix for `BaseHTTPMiddleware` middleware classes and streaming responses.
from fastapi import FastAPI, Request, Response
app = FastAPI()
@app.middleware('http')
async def foo(request: Request, call_next):
resp = await call_next(request)
if resp.status_code == 404:
return Response('Oh no!')
return resp
@app.get('/meow')
def meow():
return 'Meow!'
Re-run the minimal reproduction on your broken version, then apply the fix and re-run.
Option A — Upgrade to fixed release\nUpgrade to version 0.13.8 or later.\nWhen NOT to use: This fix should not be used if the application relies on the previous queue behavior for streaming responses.\n\n
Why This Fix Works in Production
- Trigger: ERROR: Task was destroyed but it is pending!
- Mechanism: The response from call_next was discarded, causing pending tasks to accumulate
- Why the fix works: Reverted the Queue maxsize fix in BaseHTTPMiddleware to restore previous behavior and prevent pending tasks accumulation. (first fixed release: 0.13.8).
- If left unfixed, failures can be intermittent under concurrency (hard to reproduce; shows up as sporadic 5xx/timeouts).
Why This Breaks in Prod
- Shows up under Python 3.8 in real deployments (not just unit tests).
- The response from call_next was discarded, causing pending tasks to accumulate
- Surfaces as: ERROR: Task was destroyed but it is pending!
Proof / Evidence
- GitHub issue: #1022
- Fix PR: https://github.com/encode/starlette/pull/1026
- First fixed release: 0.13.8
- Reproduced locally: No (not executed)
- Last verified: 2026-02-09
- Confidence: 0.75
- Did this fix it?: Yes (upstream fix exists)
- Own content ratio: 0.69
Discussion
High-signal excerpts from the issue thread (symptoms, repros, edge-cases).
“I have opened a separate PR on reverting the maxsize fix here for discussion: https://github.com/encode/starlette/pull/1028 It's good to know that the fix in version 0.13.7…”
“We have reverted the fix from 0.13.7 that lead to this problem. I am going to close this issue as a result, because the behavior…”
“Thanks for reporting this issue and thanks for putting together a straightforward example”
“@erewok, thanks for looking into this”
Failure Signature (Search String)
- ERROR: Task was destroyed but it is pending!
Error Message
Stack trace
Error Message
-------------
ERROR: Task was destroyed but it is pending!
Minimal Reproduction
from fastapi import FastAPI, Request, Response
app = FastAPI()
@app.middleware('http')
async def foo(request: Request, call_next):
resp = await call_next(request)
if resp.status_code == 404:
return Response('Oh no!')
return resp
@app.get('/meow')
def meow():
return 'Meow!'
Environment
- Python: 3.8
Why It Broke
The response from call_next was discarded, causing pending tasks to accumulate
Fix Options (Details)
Option A — Upgrade to fixed release Safe default (recommended)
Upgrade to version 0.13.8 or later.
Use when you can deploy the upstream fix. It is usually lower-risk than long-lived workarounds.
Fix reference: https://github.com/encode/starlette/pull/1026
First fixed release: 0.13.8
Last verified: 2026-02-09. Validate in your environment.
When NOT to Use This Fix
- This fix should not be used if the application relies on the previous queue behavior for streaming responses.
Verify Fix
Re-run the minimal reproduction on your broken version, then apply the fix and re-run.
Did This Fix Work in Your Case?
Quick signal helps us prioritize which fixes to verify and improve.
Prevention
- Add a stress test that runs high-concurrency workloads and fails on thread dumps / blocked locks.
- Enable watchdog dumps in prod (faulthandler, thread dump endpoint) to capture deadlocks quickly.
Version Compatibility Table
| Version | Status |
|---|---|
| 0.13.8 | Fixed |
Related Issues
No related fixes found.
Sources
We don’t republish the full GitHub discussion text. Use the links above for context.