Diagnosing "It's Returning the Wrong Thing": A Second Troubleshooting Walkthrough

Checking the application's own log for this morning's price update confirms it completed successfully, updating every affected product:

The source of truth is correct. Whatever's serving stale prices, it isn't because the update itself failed or ran incompletely.

This is the tempting answer, so it gets checked properly rather than assumed. A quick request with response headers visible shows no cache layer is even in front of this path:

No X-Cache, no Age header, nothing indicating a cache sits in front of this request. The symptom looked like caching, but there's no cache configured here to actually be the cause.

Checking the access log's own recorded Host field for both an "old price" and a "new price" request confirms both landed on the same, correct virtual host, serving from the same document root:

Identical host, identical config on both requests. Whatever's different between these two, it isn't which virtual host answered.

This site runs behind an Nginx reverse proxy, load-balancing across two Apache backends — a detail worth checking directly, using an Nginx log format that includes $upstream_addr, the actual backend server that handled each request:

Requests alternate between two backend servers, and only one of them is actually serving new prices. The rollout of this morning's deployment reached backend-2 but never completed on backend-1 — which explains the exact symptom precisely: which price a user sees depends entirely on which backend the load balancer happens to route them to on that request.