Skip to main content
Use metrics.v1.MetricsService to answer questions such as “how many invocations failed in the last hour?” and “how long do cold boots take by day?” The web UI’s dashboards use the same three read-only RPCs. Their messages are defined in protos/metrics/v1/metrics.proto. QueryTimeseries and QueryAggregate read the invocations table in PostgreSQL, so an invocation counts once its record is written. GetQueue reads the queue on NATS instead. Every query is scoped to the caller’s tenant.

RPCs

Metrics

Duration means skip invocations that never reached the phase. An errored call with no acquire does not pull the acquire mean toward zero.

Dimensions

Since every query is scoped to the caller’s tenant, grouping by tenant returns one group.

QueryTimeseries

Every series has the same ascending buckets over the window, zero-filled, so points at the same index refer to the same bucket. Each point’s values follow the order of metrics. When a filter covers the group_by dimension, the response has one series per requested value, in request order, even when a value has no invocations.
A shortened response looks like this:

QueryAggregate

Takes the same fields except granularity and tz_offset_minutes. The response has one AggregateRow per group for the whole window, with key and values in the order of metrics.

GetQueue

Takes function_ids, or every function of the tenant when empty. Returns queued, invocations waiting to start (in the queue, or taken by a host that is building the function or waiting on its max_concurrency), and running, from what live hosts last reported. If the queue cannot be read, the call is unavailable.

Errors

A missing since_ms, an until_ms not after it, a window over 90 days, no metrics, an unsupported metric or a missing granularity are each invalid_argument. A failed query against PostgreSQL is also reported as invalid_argument.