Live metrics

A deliberately small operational panel. e2b-compat exposes /metrics in Prometheus text format; tools/metrics-scraper.sh polls it and appends one JSON line per sample to data/metrics/YYYY-MM-DD.jsonl. This page reads whatever’s sitting in that directory at build time and plots it. No DB, no dashboard server — just JSONL and a build step.

Alive state

sandboxes + kernels currently in the registry

No samples under data/metrics/. Run mise run metrics:watch while the gateway is busy, then rebuild.

Sandbox lifecycle

create / kill / pause / resume — rate per second

No samples under data/metrics/. Run mise run metrics:watch while the gateway is busy, then rebuild.

Execution path

/execute requests + errors — rate per second

No samples under data/metrics/. Run mise run metrics:watch while the gateway is busy, then rebuild.

Per-sandbox CPU

One line per live sandbox, each labeled with the first 12 hex chars of its id. Source is coppice_sandbox_cpu_percent from rctl(8); 100 = one full core. See per-sandbox metrics for the scraper/label shape.

coppice_sandbox_cpu_percent — per-sandbox

No samples under data/metrics/. Run mise run metrics:watch while the gateway is busy, then rebuild.

How to feed this page

# On your laptop, with honor reachable and the gateway running:
mise run metrics:scrape      # one-shot append to data/metrics/TODAY.jsonl
mise run metrics:watch       # 30 s loop; Ctrl-C to stop

# Run some load so the page has something to show:
mise run demo:notebook       # or run example/07-pool-fanout.sh

# Rebuild the site to see the new samples:
pnpm build

Git-ignored by default. If you want samples to land on the deployed site, drop the data/metrics/ line from .gitignore and commit them — every sample is ~300 bytes.

What’s tracked

metric	type	source
`coppice_uptime_seconds`	gauge	process clock
`coppice_sandboxes_alive`	gauge	in-memory registry size
`coppice_kernels_alive`	gauge	ipykernel PID table size
`coppice_sandboxes_created_total`	counter	POST /sandboxes on success
`coppice_sandboxes_killed_total`	counter	DELETE /sandboxes/:id on success
`coppice_sandboxes_paused_total`	counter	POST /sandboxes/:id/pause
`coppice_sandboxes_resumed_total`	counter	POST /sandboxes/:id/resume
`coppice_execute_requests_total`	counter	POST /execute, counted at the end
`coppice_execute_errors_total`	counter	/execute that emitted an ExecEnv::Error frame
`coppice_kernel_spawns_total`	counter	ipykernel_launcher spawn succeeded
`coppice_kernel_exits_total`	counter	SIGTERM or crash reaped
`coppice_sandbox_create_ns_sum`	counter	ns spent in create, pair with `_created_total`
`coppice_execute_ns_sum`	counter	ns spent in /execute, pair with `_requests_total`

No histograms yet — if the running-mean from ns_sum / requests_total stops being precise enough we’ll swap in the metrics crate, but the present shape is enough to show whether the system is doing work.