Watch every remote build in real time.
A live view over your Buildbarn RBE cluster — cache hit rate, worker saturation, queue depth, execution latency — read straight from the Prometheus it already exposes. No new database, no dashboard to hand-build, and the same numbers are agent-callable tools.
Your cluster already emits everything. Read it.
Every Buildbarn pod exposes Prometheus metrics at :9980/metrics, and you're already scraping them. Standing up a bespoke dashboard to see them is busywork.
builds points at your Prometheus (BUILDS_PROMETHEUS_URL) and runs the PromQL — instant and range queries over a rolling one-hour window — that turns raw Buildbarn counters into the four views you actually watch during a build: cache hit rate, CAS/AC throughput, worker saturation by platform, and scheduler queue depth with execution-latency percentiles. It holds no state of its own. And because every panel is backed by a typed query, the same numbers are exposed as MCP tools — so an agent can check the cache hit rate or the worker count exactly the way the console does.
It doesn't scrape a new thing. It reads the Prometheus your Buildbarn cluster already exposes.
The whole cluster, at a glance.
The Overview panel is a KPI strip: one tile per signal, each with its hour-ago delta (the last sample vs the first in the window). Values below are illustrative — the panel renders whatever your cluster reports.
| signal | value | vs 1h ago |
|---|---|---|
| cache hit rate | 87% | ▲ 4 pts |
| active workers | 3 | steady |
| queue backlog | 0 | drained |
| median exec latency | 340ms | ▼ 60ms |
| artifacts served (1h) | 12,481 | — |
| bytes served (1h) | 1.2 GiB | — |
Every read and write, by store.
The Cache panel breaks throughput out by store — Content-Addressed Storage and the Action Cache — and tracks the action-cache hit rate that decides whether a build is fast or cold. Series and rates below are illustrative.
| series | value | prometheus source |
|---|---|---|
| AC hit rate (GetActionResult) | 87% | grpc_server_handled_total |
| CAS read | 420 ops/s | blob_access_operations_blob_size_bytes_count |
| CAS write | 96 ops/s | blob_access_operations_blob_size_bytes_count |
| AC read | 210 ops/s | blob_access_operations_blob_size_bytes_count |
| AC write | 44 ops/s | blob_access_operations_blob_size_bytes_count |
Scrape → query → shape → render.
Four stages, no state in the middle.
Scrape
Every Buildbarn pod (scheduler, workers, storage) exposes Prometheus at :9980/metrics. Your Prometheus already scrapes it — builds adds nothing to the collection path.
Query
On each panel load, builds runs PromQL instant and range queries against BUILDS_PROMETHEUS_URL — rates over [5m], a rolling 1h window at a 60s step (60 points). No new time-series database.
Shape
Results are flattened into typed builds.v1 DTOs and Vega samples — the same shape whether the caller is a panel or an MCP tool.
Render
Vega-Lite meridian panels draw the charts in the console; the identical numbers answer the overview, cache_hit_rate, list_workers, cache_stats, and list_invocations tools for agents.
Every panel names its metric.
These are the real Buildbarn series each view is computed from — verified against a live cluster's :9980/metrics, because the earlier metric names most dashboards copy don't exist in current builds.
| panel | prometheus metric | window |
|---|---|---|
| cache hit rate | grpc_server_handled_total{grpc_method=GetActionResult} | rate [5m] |
| cache throughput | ..._blob_access_operations_blob_size_bytes_count | rate [5m] |
| workers by platform | ..._build_queue_workers_created_total − ..._removed_total | sum by(platform) |
| queue backlog | ..._tasks_scheduled_total − ..._tasks_completed_..._count | instant |
| exec latency p50/p90/p99 | ..._tasks_executing_duration_seconds_bucket | histogram_quantile |
Not another dashboard to maintain.
The alternative is standing up and babysitting a dashboards deployment, guessing the metric names, and leaving agents out entirely.
No Prometheus reachable? Every panel degrades to an empty state — never an error, never a fabricated number.
source · github.com/fastverk/plugin-builds · tools: overview, cache_hit_rate, list_workers, cache_stats, list_invocations
Prove it's safe to merge.
builds is one of 14 plugins in the fastverk console — hosted, or in your own cloud.