Aggregate AI latency buckets before reporting service percentiles

Specify which waiting interval the report describes

Two replicas serve permitted AI drafting attempts. Their dashboards each display p95. Averaging those two numbers looks convenient, but it discards the distributions that a combined service question needs. A familiar percentile label does not make its values additive.

Select one narrow descriptive profile: already available classic Prometheus histograms, nonnegative durations in seconds, identical bucket boundaries and measurement meaning, and disjoint observations from one finalized common window. This is an interpretation worksheet, not a recommendation to instrument a new system with classic histograms.

The clock guide governs an individual interval. Here the declared interval is adapter acceptance through its terminal local observation, excluding earlier queue time. Provider computation, user waiting and successful business completion remain different quantities. Actual instrumentation and window coverage are unimplemented prerequisites, not established facts.

Read the metric type before combining its numbers

The maintained Prometheus metric-types reference describes classic histogram buckets as cumulative counters with inclusive upper bounds. Its positive-infinity bucket equals the observation count. A summary instead exposes calculated quantiles alongside count and sum. The same body cautions that classic histograms with different boundaries are not generally aggregatable.

The histogram practice guide explains that averaging replicated summaries’ precomputed quantiles does not yield the overall percentile. Compatible histogram counts permit aggregation before quantile estimation; histogram-derived quantiles remain estimates. The guide now favors native histograms overall. This worksheet only interprets an existing, explicitly selected classic representation and makes no library or migration recommendation.

Both actual bodies were read on 1 November 2026. No pinned documentation release, installed Prometheus version, deployed collector or instrumentation library was established. No linked specification, query example or collector implementation supplies additional evidence here.

Combine one invented cumulative-count sheet

Use three teaching boundaries: 0.1 seconds, 0.5 seconds and positive infinity. They are invented reporting choices, not a production latency target. Every number below is assigned for arithmetic teaching; none came from a request, scrape, histogram object or experiment.

Add corresponding rows across the two disjoint replicas. The teaching ledger therefore assigns 95 of 100 observations at or below 0.5 seconds. That is 95% within this chosen boundary, not an observed service result.

Do not add 90, 95 and 100 into 285 attempts. These rows overlap: the smaller-boundary observations are already included in the larger rows. Nor can this sheet reveal the exact individual duration occupying a percentile position. Keep the useful threshold count without manufacturing missing raw measurements.

Make the window and denominator an admission decision

The displayed table represents invented counts for a completed window, not arbitrary raw counter snapshots. An actual system needs a separately supported way to obtain complete window evidence while accounting for restarts, missing series and collection gaps. This article provides no subtraction, rate expression, query or reset-handling procedure.

Before a real merger, bind the interval definition, units, label population, bucket layout, source instances and window to one protected reporting plan. Establish that observations are neither duplicated across replicas nor omitted from a required source. A matching metric name or attractive graph cannot supply those associations.

Retain the expected observation inventory and known exclusions. For example, a metric containing only successful responses cannot silently become latency for every attempted task. An unresolved timeout still needs an operational disposition even if it contributes no finished duration to this selected dataset. State the measured population instead of replacing missing evidence with fast completion.

The conditional positive accepts only the bounded descriptive report after all actual prerequisites hold. It changes no retry, capacity, data-release or authorization decision. If the selected denominator is zero, preserve an unavailable fraction rather than announcing perfect performance.

Walk eight inert reporting cards

Keep a reporting result within its evidence

Actual capture boundaries, complete compatible series, supported counter-to-window processing, nonduplicated attempt associations and current permission to inspect telemetry remain unknown here. Unknown prerequisites block a real positive. Aggregate reporting also needs appropriate label minimization and access controls; do not attach private prompts, credentials or customer content to this teaching record.

No metric, query, histogram, summary, observation stream, collector, benchmark or model runtime was constructed or executed for these cases. The authored tables are paper arithmetic. This worksheet implements no monitoring service and establishes no production latency, statistical confidence interval, future distribution, availability commitment or performance improvement.

Boundary is prelaunch with planned request-path checks. Its existing synthetic policy evidence concerns different behavior. A latency report cannot prove that an answer was correct, a provider stopped work, an action completed or a user currently has permission to perform it.