Not in the application logs. Not in the worker logs. In the audit log of a sidecar proxy—a small, overlooked Envoy instance running on a node that had been scheduled for retirement six months ago. The entry read:
There was no stack trace. No reference number. No helpful “Did you mean...?” suggestion. Just six words and a three-letter code that felt less like a system message and more like a taunt.
Another theory, darker and more romantic, was that “rgh” stood for “Run-time Garbage Heap”—an internal nickname for a now-decommissioned orchestration layer that scheduled batch jobs using a custom scheduler written in a language whose name management had tried to forget. That scheduler had a feature: when it lost track of a job, it didn’t just fail. It assigned an impossible execution ID—one that existed in the liminal space between “submitted” and “never started.” invalid execution id rgh
The rgh part, however, was a mystery. In most systems, error codes follow a logic: E1001 for auth failures, 4xx for client errors. But rgh was not a code. It was a whisper.
So the system did the only logical thing a machine can do when faced with an orphaned miracle: it marked the execution ID as invalid. Not wrong. Just... disconnected. A floating point in a network graph that no longer contained its origin. Not in the application logs
Four rows updated.
But execution IDs are not immortal. They expire. They get garbage-collected. They are wiped from Redis caches during a midnight failover. And when a client—innocent and oblivious—presents that ID again, asking, “What happened to my job?” the system does not apologize. It does not explain. It simply says: invalid . The entry read: There was no stack trace
At 3:47 AM, they found it.