Repository navigation
Observability: log line only, or also a Prometheus metric on trigger? #61
Description
Activity
- added a parent issue
on Sep 1, 2026 Phone-friendly options (reply with a letter; unblocked by your #57 answer):
- A: log only (current).
slog.Warnon trigger. Zero new surface. - B (recommended): log + Prometheus counter
gomodel_stream_repetition_triggers_total{provider,model}. Alertable; matches how every other gateway labels loop aborts (vLLM usesfinished_reasonlabels). - C: log + counter + per-choice histogram of repeat counts. Richest; most code for marginal value.
Grill note: the counter is what makes the guard tunable later — without it you cannot see which providers actually loop.
- A: log only (current).
B
Resolved: B. Log line + Prometheus counter
gomodel_stream_repetition_triggers_total{provider,model}on every trigger, alongside the existingslog.Warn. Counter makes the guard tunable later (shows which providers actually loop). Implementation note: reuse the gateway's existing metrics registry pattern; no histogram for v1.
Status
Options A/B/C posted in a follow-up comment. Waiting on user reply.
Status
Options A/B/C posted in a follow-up comment. Waiting on user reply.
Question
Question
The prototype logs slog.Warn on trigger. Should we also expose a counter (e.g. gomodel_stream_repetition_triggers_total{provider,model}) so operators can alert on looping models?