Skip to main content
CoreWeave Sandboxes publish per-organization metrics that describe every active sandbox: how many are running, what resources they requested, and how much CPU and memory they’re consuming. Use these metrics to track fleet activity, attribute usage, and build your own dashboards and alerts.
CoreWeave sandboxes are in public preview. For access, contact your CoreWeave account team, CoreWeave Support, or email support@coreweave.com.

Prerequisites

The metrics are available in the Explore section of the Grafana instance that CoreWeave hosts. To access Grafana, you must be logged in to the CoreWeave Cloud Console and be a member of the admin, metrics, or write groups. You can also query the same metrics programmatically through the metrics API at observe.coreweave.com with an API access token.
Queries are automatically scoped to your organization. Selectors on cluster or cluster_org are rewritten to your organization, so a query that references another organization returns your own data rather than an empty result.

Common labels

All cwsandbox_* metrics carry the following labels: The label contract is identical whether your sandboxes run on CoreWeave-managed shared capacity or on your own CoreWeave Kubernetes Service (CKS) clusters with a managed runner. To break usage down by placement, group by the runner_id label on cwsandbox_sandbox_info.

Lifecycle metrics

These metrics come from the sandbox control plane and update about every 30 seconds. Per-sandbox series exist while the sandbox is active and disappear shortly after it stops.

Metric cwsandbox_sandboxes

The number of active sandboxes in each lifecycle state. Units are in sandboxes. The following labels are available for filtering and grouping:

Metric cwsandbox_sandbox_info

A constant 1 series that carries the identity of each active sandbox. Use it to join identity labels onto the usage metrics. The following labels are available for filtering and grouping:

Metric cwsandbox_sandbox_age_seconds

The age of each active sandbox since creation. Units are in seconds. Labeled by sandbox_id.

Metric cwsandbox_sandbox_requested_cpu_millicores

The CPU each sandbox requested at admission, not measured usage. Units are in millicores. Labeled by sandbox_id.

Metric cwsandbox_sandbox_requested_memory_bytes

The memory each sandbox requested at admission, not measured usage. Units are in bytes. Labeled by sandbox_id.

Metric cwsandbox_sandbox_requested_gpus

The GPUs each sandbox requested at admission. Units are in GPUs. Labeled by sandbox_id and gpu_type.

Usage metrics

These metrics are measured on the node running each sandbox and update about every 30 seconds.

Metric cwsandbox_sandbox_cpu_usage_seconds_total

A counter of the CPU time each sandbox has consumed. Units are in CPU seconds. Labeled by sandbox_id. Use rate() or increase() to compute consumption over a window.

Metric cwsandbox_sandbox_memory_working_set_bytes

The current working-set memory of each sandbox. Units are in bytes. Labeled by sandbox_id.
Usage values reflect the sandbox workload container. For sandboxes running on microVM isolation, the values come from inside the guest and exclude virtualization overhead and platform sidecar processes. The CPU counter resets if the workload container restarts, which rate() and increase() handle automatically. A short gap in a series means collection was briefly interrupted. Stale values are never reported.

PromQL examples

Running sandboxes over time

CPU cores in use per sandbox

CPU usage enriched with profile and runner

Join the usage counter with cwsandbox_sandbox_info on sandbox_id to attach identity labels:

Total memory in use across the fleet

CPU efficiency, used versus requested

Requested GPUs by type

Longest-running sandboxes

Last modified on August 18, 2026