pscale_exporter
A Go exporter for Dell PowerScale (OneFS). It authenticates to the OneFS platform
API, collects broad cluster, node, protocol, quota, snapshot and capacity metrics, and
exposes them via both a Prometheus /metrics endpoint and an optional OTLP
metric push — both fed from one shared snapshot. The powerscale_ metric prefix matches
dell/csm-metrics-powerscale for
dashboard compatibility.
Features
- Dual export: Prometheus pull (
/metrics) and an optional OTLP metric push, fed
from one shared snapshot.
- Broad OneFS coverage: cluster, nodes, protocols (NFS/SMB export & share
counts), quotas, snapshots, and capacity — combining typed OneFS resources with the
raw statistics API (curated stat keys plus the protocol summary).
- Multi-cluster: one process monitors many OneFS clusters; every metric carries a
cluster label.
- Operational: one shared gopowerscale session per cluster, graceful per-cluster
degradation (an unreachable cluster is marked down without taking the exporter down),
hot config reload (SIGHUP + file watch), snapshot-based
/health JSON plus always-200
/livez//readyz probes, and optional OTLP tracing.
- CSM-compatible naming: the
powerscale_ metric prefix matches Dell's
csm-metrics-powerscale so existing dashboards work.
Quick start
make cli
export PSCALE1_PASSWORD='your-monitor-password'
./bin/pscale_exporter --config config.yaml
# metrics: http://localhost:9444/metrics health: http://localhost:9444/health
Or install with Homebrew: brew install fjacquet/tap/pscale_exporter.
Or with Docker: docker pull ghcr.io/fjacquet/pscale_exporter:latest.
Local test stack
A docker compose stack brings up the exporter alongside Prometheus, an OpenTelemetry
collector, and Grafana for a complete end-to-end test:
PSCALE1_PASSWORD='your-monitor-password' docker compose up -d --build
Point config.yaml at a real cluster and set PSCALE1_PASSWORD for a true end-to-end
run; with the default example cluster the stack still validates wiring (powerscale_up
goes to 0 and the PowerScaleClusterDown alert fires). To exercise the OTLP push path,
set opentelemetry.metrics.enabled: true with endpoint: "otel-collector:4317" in
config.yaml. A pull-based variant (published GHCR image, no local build) is in
docker-compose.ghcr.yml.
Documentation
Full docs at https://fjacquet.github.io/pscale_exporter/:
Development
make tools # install golangci-lint, cyclonedx-gomod, govulncheck (pinned)
make sure # fmt + vet + test + build + golangci-lint
make ci # the gate CI runs (adds go test -race + govulncheck)
Troubleshooting on a remote site
--debug logs every OneFS request (full URL, size, duration), per-cluster parse
summaries (node/quota/sensor counts, which curated stat keys came back missing), and a
payload snippet whenever a response doesn't parse. For a schema surprise we can't
reproduce, ask the operator to run one cycle with response dumping:
pscale_exporter --config config.yaml --debug --once --dump-dir /tmp/pscale-dump
# then ship /tmp/pscale-dump/<cluster>/ back — one verbatim JSON file per endpoint,
# usable directly as test fixtures. Payloads contain no credentials.
To validate a live cluster end to end, --trace logs every OneFS API response body
(method, URL, status; headers — and thus the session cookie — are never logged), and
--once --debug prints every collected sample sorted in exposition style:
pscale_exporter --config config.yaml --once --debug --trace > validate.out
grep -v '^{' validate.out > samples.txt # every sample (diff against docs/metrics.md)
grep 'API trace' validate.out # raw OneFS payloads for anything missing
Notes
- A read-only OneFS account (a role with read access to
ISI_PRIV_STATISTICS,
ISI_PRIV_QUOTA, and ISI_PRIV_DEVICES — the latter is required by the
cluster-nodes inventory) is sufficient for collection.
- IOPS and bandwidth are already per-second gauges — aggregate with
sum/avg in
PromQL, never rate().
- Metric names are unit-explicit:
_bytes, _bytes_per_second,
_operations_per_second, _seconds, _percent.
- The
powerscale_ prefix matches Dell's csm-metrics-powerscale for dashboard reuse.
Security
/metrics, /health, /livez and /readyz have no authentication or TLS and bind
0.0.0.0 by default, and the metrics expose cluster, node, quota path and share names.
Run the exporter on a trusted network, set server.host to a specific interface, or put it
behind a reverse proxy. TLS verification of the OneFS API is on by default
(insecureSkipVerify is an opt-in). Use a read-only OneFS account. To report a
vulnerability, see SECURITY.md.
Stability and contributing
From 1.0, metric names, labels and units, CLI flags, config keys and the default port are
stable across 1.x; metrics not yet seen on a live cluster are marked Provisional in the
metrics reference. Whether a metric
appears depends on your OneFS version. The full policy is
ADR-0007. Contributions are welcome:
see CONTRIBUTING.md.
License
Apache License 2.0 — see LICENSE.
