Independent developer. I build commerce systems, native apps and games, and I run them in production.

A health check that passed on its own error

if "logged in" in out.lower():   # "Not logged in" matches too
    return OK

A token watchdog reported OK while the login had expired, because the failure message contains the success string. A real deploy failed the same morning.

Rule Test every probe against the failure it is meant to catch.

HTTP 200, resource name returned, nothing changed

client.copy_from(op.update_mask,
    protobuf_helpers.field_mask(None, updated._pb))

A widely copied Google Ads update pattern completes without an exception and leaves the field untouched. Only a read-back shows it.

Rule A write is not done until the value has been read back.

Text that ended at the emoji

A platform saved replies and reported success, but dropped the emoji and everything after it, signature included. A classic 3-byte UTF-8 limit, with no error anywhere.

Rule Compare the whole stored text, not the status.

One changed character in ten modules

Content moved through an agent into a browser editor arrived with a single corrupted character. Length matched, the markup stayed valid, and one of them reached production.

Rule Hash in chunks before writing. Length proves nothing.