ipx logged a line for every request: 2,306 a day, over half its log by
size, repeating Caddy's access line and the request's span. It stayed
because ipx had no request metric and the dashboard's API row was built
on the line. /metrics now has ipx_http_request_duration_seconds, a
histogram by method, route and status (5 ms to 10 s), the event stream
left out as the panels left it out. The route is the pattern, "" for an
unrouted path, and an unusual method is "other", so a scanner cannot make
a row a request. A request that went well is logged at debug, which the
admin page's log view keeps; a failed one keeps its warning in Loki.
The dashboard's API row reads the histogram (the slowest-request series
becomes the 99th percentile, a bucket's bound; the slowest themselves are
in Tempo). prom() takes instant queries and table() a unit for it.
The prod-check skill, reviewed whole: check 3 reads the histogram; the
directory has metrics, which it said it had not; "all gauges" was half
counters; the healthcheck removed in #218 was still named three times;
Caddy's /api/events aborts (600 a week, a page closing its stream) are
filtered as by design; the monitoring section checks Tempo's block poll,
which failed on 2026-10-25 on one empty block and left every search with
only the last 15 minutes; and the lines to expect from each container
are counted over a day, now that iPX can be quiet for an hour.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rays
c788112190