Metrics reference

NativeLink v1.6.5 declares 30 OpenTelemetry instruments. This page is generated from the declarations in nativelink-util/src/metrics.rs and from every .add() and .record() call site across the workspace, so the Emitted column is not a claim, it is the result of looking.

That column is the reason this page is generated rather than written. Declaring an instrument costs nothing and emits nothing; the instrument only produces a series when some code path calls it. A catalogue written by hand cannot tell those two states apart, and the gap between them is where dashboards go quietly wrong.

Every instrument at a glance

How an instrument becomes a Prometheus series

NativeLink speaks OTLP only; there is no Prometheus scrape endpoint in the binary. The names in the middle column above are what the shipped OpenTelemetry collector produces after two transformations:

1. The Prometheus exporter in otel-collector-config.yaml is configured with namespace: nativelink, which prefixes every series with nativelink_.
2. The exporter rewrites the OTLP name into Prometheus form: dots become underscores, and monotonic counters gain a _total suffix. Histograms fan out into _bucket, _sum and _count series.

cache.operations (a counter) is therefore scraped as nativelink_cache_operations_total, and cache.operation.duration (a histogram) as nativelink_cache_operation_duration_bucket and friends. Change the collector's namespace and every name on this page changes with it.

Where a name in the Prometheus series column is unmarked, it is not a prediction: it is the name that actually appears in the shipped rules, dashboards and alerts under deployment-examples/metrics/ and kubernetes/. A name marked (unreferenced) is derived from the two rules above, because no shipped artifact mentions that metric at all: nothing has ever queried it, so nothing has confirmed the exporter's spelling of it.

Declared but never emitted

2 instruments are built by the meter and stored on the metrics struct, but no code anywhere in the workspace ever calls .add() or .record() on them. They will never appear in your Prometheus output, at any configuration, on any release this page was generated from.

Cache metrics

Global cache metrics instruments.

Declared as CACHE_METRICS in nativelink-util/src/metrics.rs:419, backed by the CacheMetrics struct.

Every cache instrument is emitted from one place: the cache_metrics store wrapper. A store that is not wrapped emits nothing, no matter how the collector is configured.

cache.operation.duration

Histogram of cache operation durations in milliseconds

Histogram buckets: 19 explicit boundaries, from 0.001 to 5000.0 ms. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

cache.operations

Counter of cache operations by type and result

cache.io

Counter of bytes read/written during cache operations

cache.size

Current total size of all cached data in bytes

cache.entries

Current number of entries in cache

cache.item.size

Histogram of individual cache entry sizes in bytes

Execution metrics

Global remote execution metrics instruments.

Declared as EXECUTION_METRICS in nativelink-util/src/metrics.rs:506, backed by the ExecutionMetrics struct.

Execution instruments are emitted from the scheduler process as actions change state, plus the per-action CPU time and peak memory a worker reports back when it finishes. Worker processes emit no execution metrics of their own.

execution.stage.duration

Histogram of stage durations in seconds

Histogram buckets: 15 explicit boundaries, from 0.001 to 3600.0 s. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

execution.total.duration

Histogram of total execution durations in seconds

Histogram buckets: 13 explicit boundaries, from 0.01 to 7200.0 s. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

execution.queue.time

Histogram of queue wait times in seconds

Histogram buckets: 12 explicit boundaries, from 0.001 to 600.0 s. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

execution.active.count

Current number of actions in each stage

execution.completed.count

Total number of completed executions

execution.stage.transitions

Number of stage transitions

execution.cpu.time

CPU time an action used, as sampled by the worker.

Histogram buckets: 12 explicit boundaries, from 0.1 to 7200.0 s. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

execution.peak.memory

Peak memory an action used, as sampled by the worker.

Histogram buckets: 8 explicit boundaries, from 1_048_576.0 to 68_719_476_736.0 By. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

execution.output.size

Histogram of output sizes in bytes

Histogram buckets: 8 explicit boundaries, from 1_024.0 to 10_737_418_240.0 By. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

execution.retry.count

Counter for execution retries

Worker fleet metrics

Global worker fleet metrics instruments.

Declared as WORKER_METRICS in nativelink-util/src/metrics.rs:817, backed by the WorkerMetrics struct.

Worker fleet instruments are emitted from the scheduler process as workers connect, heartbeat, pause, drain and leave. They describe the pool as this scheduler instance sees it.

worker.connected.count

Workers currently connected.

worker.connections

Workers that have joined, cumulative.

worker.disconnections

Workers that have left, cumulative, by reason.

worker.keepalives

Keepalives received, cumulative.

worker.state.count

Connected workers currently paused or draining.

gRPC server metrics

Global gRPC serving metrics.

One duration histogram covers rate, errors and latency: the count gives rate, the status attribute separates errors, and the buckets give latency.

Declared as RPC_METRICS in nativelink-util/src/metrics.rs:921, backed by the RpcMetrics struct.

Emitted by the OtlpLayer tower middleware on every gRPC server NativeLink starts, so a CAS, a scheduler and a worker API endpoint each report their own.

rpc.server.duration

Duration of inbound gRPC calls, by service, method and status.

Histogram buckets: 15 explicit boundaries, from 0.001 to 300.0 s. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

Scheduler metrics

Global scheduler metrics.

Declared as SCHEDULER_METRICS in nativelink-util/src/metrics.rs:974, backed by the SchedulerOtlpMetrics struct.

Emitted from the scheduler process once per matching pass, whether or not the pass assigned anything.

scheduler.matching.duration

How long a matching pass takes. The saturation signal: this climbing while the queue is non-empty means matching is the bottleneck.

Histogram buckets: 13 explicit boundaries, from 0.0001 to 10.0 s. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

scheduler.matching.passes

Matching passes, by result.

Store tier metrics

Global tiered-store metrics.

Declared as STORE_TIER_METRICS in nativelink-util/src/metrics.rs:1023, backed by the StoreTierMetrics struct.

Emitted from the fast_slow store wrapper only. A deployment without a fast_slow store emits none of these.

store.tier.operations

Reads served per tier, by result.

store.tier.io

Bytes moved per tier and direction.

Health check metrics

Global health-check metrics.

Declared as HEALTH_METRICS in nativelink-util/src/metrics.rs:1074, backed by the HealthMetrics struct.

Emitted each time the health check endpoint runs, one observation per registered health indicator.

health.checks

Health check results, by namespace and status.

Connection pool metrics

Global connection-pool metrics.

Declared as CONNECTION_METRICS in nativelink-util/src/metrics.rs:1105, backed by the ConnectionMetrics struct.

Emitted by the gRPC connection manager and the Redis store as they hand out and re-establish connections. A deployment that uses neither emits none of these.

connection.pool.available

Free slots at the moment a connection was taken.

Histogram buckets: 11 explicit boundaries, from 0.0 to 512.0 {connection}. Anything past the last boundary lands in the +Inf bucket, so a quantile computed above it is a lower bound, not a measurement.

connection.pool.acquisitions

Acquisitions, by pool and result.

connection.reconnects

Reconnects, by pool.

Attributes

Every attribute key NativeLink attaches to a metric, and the closed set of values it can take. The Prometheus label is the key with dots replaced by underscores.

Cache attributes

Execution attributes

Worker attributes

gRPC attributes

Scheduler attributes

Store tier attributes

Health attributes

Connection attributes

cache.operation.name values

cache.operation.result values

execution.stage values

execution.result values

Reading the source

If anything here disagrees with the binary, the source wins:

• nativelink-util/src/metrics.rs: every declaration
• nativelink-util/src/telemetry.rs: the OTLP exporter and its interval
• nativelink-store/src/cache_metrics_store.rs: every cache call site
• nativelink-scheduler/src/simple_scheduler_state_manager.rs: every execution call site

Where to go next

NextWire up the telemetry pipeline

The collector, the exporter interval and the queries worth alerting on: the operational half of this reference.

SidewaysConfiguration reference

Every configuration field, generated the same way from the same source of truth.