zeekayandHanzo Dev 13404430a5 a frame that arrived is a trace, not news ...

The broker named three APIs as the chatty ones and announced every other
frame at Info. That is a guess about which calls a client repeats, and it
was wrong about the one that matters. Measured in production: Metadata was
824 of 1200 log lines, 69% of everything the process said. Eight minutes of
that rotates a pod's boot errors out of reach, which is how the last outage
came to need forensics instead of a read.

Every received frame is Debug now, so there is no list to keep correct as
APIs are added. Info stays for what an operator did not already know.

Two races were in the way of proving it, and `go test -race ./...` had never
passed here — seven at HEAD, before any of this.

Serve published the listener while Addr and Shutdown read it, with nothing
ordering the three. The bound port is now settled before the listener is
published, so a caller that can see one can see the other.

The level itself was a plain global read by every logging goroutine, and the
sink was chosen by a log.SetOutput on every single line — a process-global
written concurrently from every connection, which also left no way to read
back what was written. The level is atomic and the sink is chosen once.

TRACE was missing from the severity map and scored zero by absence; saying
it is cheaper than rediscovering why the order comes out right.

Co-authored-by: Hanzo Dev <[email protected]>

2026-08-13 19:14:27 -07:00

..

logging.go

a frame that arrived is a trace, not news

2026-08-13 19:14:27 -07:00