Investigate First

Cause

It's a race on shared state between threads. current_user is a class attribute, so every thread in a gunicorn worker process reads and writes the same slot. With --threads 4, two requests can interleave like this:

Alice is shown Bob's cart. Two details make it worse:

1. The bad entry sticks. The wrong cart is stored in _cart_cache under Alice's id. Her later requests to that worker hit the cache and keep showing Bob's cart until the process restarts. So one short race turns into a lasting leak for that user on that worker.
2. The key and the value come from different sources. The cache key uses the thread-local request.user.id, but the value is loaded from the shared CartView.current_user. Nothing checks that the two refer to the same user.

Why only production, and why rarely

• Staging (--threads 1): each worker process handles one request at a time, so nothing else can overwrite current_user between the write and the read. The bug can't happen there. Separate gunicorn workers are separate processes, so they don't share the class attribute.

• Production (--threads 4): requests share the attribute. It only goes wrong when all three of these line up:

  • a cache miss (first visit to that worker, or after a restart)
  • a second request writes current_user in the short gap before load_cart reads it
  • the GIL switches threads at exactly that point

  That combination fits "once in a few hundred requests."

How to confirm it before changing anything

These checks are cheap. I haven't run any of them; this diagnosis comes from reading the code.

1. Log the mismatch. Temporarily log a warning in production or in a load test whenever CartView.current_user.id != request.user.id just before load_cart. Any hit proves this mechanism.
2. Reproduce in staging. Run staging with --threads 4 and send concurrent requests from many different users to the cart page. A fresh worker gives the most cache misses. Then check which user's items each response contains. Running the same load with --threads 1 should show zero leaks.
3. Force it in a test. Make load_cart sleep briefly or wait on a threading.Barrier, then call CartView().get() from two threads with different users. This reproduces it every time.
4. Check that it persists. Once a user has seen a wrong cart, their repeat requests to the same worker should keep showing that same wrong cart. Restarting the worker should clear it. If that's what happens, the cache is what turns the race into a lasting leak.

If check 1 or 2 fires, the cause is confirmed. Note that any fix will also need to clear the poisoned entries in _cart_cache, either by restarting the workers or by discarding the cache. Holding off on the fix as you asked.