Silent failure — when failure looks like success (the general case)

No exception, no log line. There was a day with five of these.

What it said	What was true
CI green	Nothing actually ran
Search: 0 results	A 2KB bot-block page
Progress file: ok:225	It re-read the husk it had produced itself
PUT 200 SUCCESS	Value unchanged. Six weeks of no-ops
Function: "none found"	It was treating a tuple like a string
Cron exited 0	The guard caught itself; the job never ran once
registered: true, reason: null	Zero images
Doc says "all deleted"	68 left

Principle: look at the raw data before you rule

• 0 results, "not found", success flags are observations, not conclusions. Look at how the observation was produced first.
• Read the original fields, not the summary or status string.
• Don't check whether it says "done"; count how many came out.
• Check response size and type. len < threshold or an unexpected type means failure.

When only a traceback is left and even "new: 0" never prints, dead and empty look the same

A notification batch died on a KeyError — bracket access (rec['url']) on one field of one record of external data. The log held a traceback and nothing else; the "new: 0" line a healthy run prints was missing too. "No orders today" and "the batch is dead" produce the same log.

• Log one line at batch start and one at batch end. Even "new: 0" has to appear, or death and absence are indistinguishable.
• Treat every field of external data as optional. Use .get(). Bracket access kills the whole run over a single record missing one key.

---

---