The in-memory metrics system covered one gunicorn process at a time, was lost on restart and was never used. Remove /admin/metrics-dashboard, /admin/metrics, starpunk/monitoring, the HTTP/query timing, memory monitor, feed counters, METRICS_* settings and psutil. Keep /health, /admin/health (without its metrics block), pool stats and the feed cache. ADR-072. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2.0 KiB
ADR-072: Remove the metrics dashboard
Status
Accepted (v2.12.0, 2026-11-03). Supersedes ADR-053 and ADR-060.
Context
v1.1.1 and v1.1.2 added built-in performance monitoring (ADR-053): HTTP
request timing middleware, a wrapper on every database connection that
timed queries, a memory-sampling thread, feed-generation counters, an
in-memory buffer of the results, and an HTML dashboard at
/admin/metrics-dashboard with a JSON twin at /admin/metrics (ADR-060
covered adapting the data for the dashboard template).
In practice it told us little:
- The buffer lives in memory in each of the four gunicorn processes, so the dashboard showed whichever process served the request, and everything was lost on restart or deploy.
- Nothing in it was ever used to find or fix a problem. Logs (with
correlation IDs),
/healthand/admin/healthwere. - It cost something on every request and every query, kept a thread per
process, and needed
psutil, about 1,300 lines of code and three test files.
Decision
Remove the dashboard and everything that only fed it: the starpunk/monitoring
package, the HTTP metrics middleware (including its X-Request-ID
response header), MonitoredConnection, the memory monitor, the
feed-generation tracking calls, /admin/metrics-dashboard,
/admin/metrics, the METRICS_* settings and the psutil dependency.
Keep:
/healthand/health?detailed=true;/admin/health, without itsmetricssection (database, pool statistics, filesystem and disk checks stay);- connection pool statistics (
get_pool_stats) and the feed cache, which do work of their own; - structured logs with correlation IDs (ADR-054).
Consequences
- Less code and one fewer dependency; no timing work on each request or query; one fewer thread per process.
- Slow database queries are no longer logged. If that is ever needed, a query timer can be added at the pool level, or an external tool used.
METRICS_*environment variables are ignored if still set.- Bookmarks to
/admin/metrics-dashboardnow get a 404.