Latency

Four buckets rather than a stored percentile, why that is the right shape, and what p95 means here.

Latency is stored as four disjoint counters, not as a percentile.

Bucket	Range
ms_under_100	Under 100ms
ms_under_500	100–500ms
ms_under_2000	500ms–2s
ms_over_2000	Over 2s

Disjoint, not cumulative: a 300ms call is counted in 100–500ms only.

Why buckets

Percentiles cannot be summed. Storing a p95 per day and averaging it across a month gives a number that is not the p95 of anything.

Four disjoint counters can be summed, and they aggregate correctly across any date range and any set of tools. A histogram of four buckets is enough to answer the question anyone actually has — is this tool slow? — without pretending to a precision the storage does not have.

What p95 means here

The p95 latency card walks the cumulative counts to find the bucket holding the 95th-percentile call, and reports that bucket’s upper edge.

So a p95 of 500ms means “the 95th-slowest call is somewhere in the 100–500ms band” — an upper bound, not the value itself. Interpolating inside the bucket would invent precision the four counters do not have.

The top bucket is open-ended, so a p95 that lands there has no numeric edge and reads as >2s.

With no calls at all, the card reads a dash rather than zero — an absent measurement, not a fast one.

The distribution chart

The same four counters as a chart, with shares that sum to 1. Rendered on the overview server-wide and on each tool page scoped to that tool, from the same component and the same shape.

Reading it

The bucket that matters most is >2s. Two seconds is roughly where an agent’s user starts noticing it is waiting, and the slow tool insight fires when more than 10% of a tool’s calls land there.

A long tail of slow calls with a fast p50 is usually one of:

• an unbounded query that gets slow with large inputs
• a cold cache or a cold connection pool
• an upstream API you do not control, on its own bad days

The distribution tells the three apart better than any single number: the first spreads across buckets as inputs grow, the second is a thin spike, the third is bursty over the calls per day chart.

Latency and cost are the two where rising is bad

The delta on the p95 card colours a rise as bad, while the arrow follows the actual change. That verdict is decided API-side rather than in the interface, so the REST API and the MCP tools reach the same conclusion without re-deriving it.

Related

• Response size — the other per-call cost of an answer
• Slow tool insight