Fourteen panels, opened three times in six months, replaced by one message a day in the channel people already read.
yesterday, checkout:
requests 41,208 (+4% w/w)
errors 18 (0.04%)
p95 310ms (+12ms w/w)
slowest query 1.2s orders.customer_id lookup
the dashboard still exists, for when one of these numbers
is wrong. it is a diagnostic tool, not a monitoring one.
A dashboard requires somebody to decide to look, which happens after a problem is already known. A daily figure in a channel gets read passively and builds an expectation, so the week the p95 moved by 80 ms three people noticed. Keeping the dashboard for diagnosis is the right split — it was never bad, it was just the wrong medium for noticing.