Technical Overview

How Crest predicts drive failures.

A walkthrough of the telemetry pipeline, the failure model, and the alert surface.

/dev/sda /dev/nvme0 Crest Agent TLS 1.3 egress Risk Model Cox PH per model PagerDuty Prometheus
1

Telemetry ingestion.

The Crest agent reads SMART data via a smartmontools-compatible interface, NVMe admin commands, and vendor health logs. It normalises attribute values across drive models and sends attribute deltas to the Crest backend over TLS every 5 minutes.

The agent runs as a systemd service or Docker container. Single binary, no kernel modules, no packet inspection. It reads from /dev/sd* and NVMe namespaces using standard health log commands.

Minimum overhead: polling interval 5 minutes, memory footprint under 32MB.

crest status - storage-01
$ crest status --host storage-01   # Agent v1.4.2 connected. 5 drives monitored.   DRIVE ID MODEL HEALTH STATUS NV-04E2 Samsung 980 Pro 2TB 74 AT RISK <-- 7-12 days ST-B91C Seagate IronWolf 6TB 81 WATCH <-- 14-21 days WD-7A2F WD Blue 4TB 97 HEALTHY WD-C3A1 WD Red Plus 8TB 94 HEALTHY KX-55D7 Kioxia CD6-R 3.84TB 88 HEALTHY   # Last telemetry push: 3 min ago. Next: 2 min.
2

Failure modeling.

The prediction model uses survival analysis - specifically Cox proportional hazards regression - trained per drive model family. Treating drive failure as a time-to-event problem means the model outputs a probability curve, not a threshold comparison.

Each drive model has distinct failure signatures. A Samsung 980 Pro degrades differently from a Seagate IronWolf. Training per-model means the risk score accounts for what "normal degradation" looks like for that specific drive, not for drives in general.

Risk thresholds: 70% probability = Watch state. 90% probability = Critical. Both include median time-to-failure estimate with confidence interval.

0 70% 90% 100% Watch Critical Time (days) Failure probability NV-04E2
3

Surface the alert where your runbook lives.

When a drive crosses the risk threshold, Crest fires the configured alert channel. Each alert includes the drive identifier, host, predicted failure window, confidence level, and recommended action.

"Swap NV-04E2 on storage-02 before 2026-07-18 (7-12 day window, 91% confidence)" is an actionable runbook item. "Critical disk event detected" is not.

All integrations
PagerDuty incident payload
"summary": "Crest: Drive NV-04E2 at critical risk",
"severity": "critical",
"custom_details": {
  "drive_id": "NV-04E2",
  "host": "storage-02",
  "model": "Samsung 980 Pro 2TB",
  "health_score": 74,
  "risk_level": "critical",
  "predicted_window": "7-12 days",
  "confidence": "91%",
  "action": "Replace before 2026-07-18"
}

Install in under 10 minutes.

1
Download the agent
Binary for Linux x86_64 or ARM64, or pull the container image. No kernel module, no special privileges beyond block device read access.
2
Run on each storage host
The agent autodiscovers drives at /dev/sd* and NVMe namespaces. Reads health telemetry on first run.
3
Paste your API key
Copy the API key from the Crest dashboard. Agent authenticates and starts pushing telemetry.
4
Fleet appears within 60 seconds
Initial health scores are computed within the first few polling cycles. Risk modeling begins after baseline telemetry accumulates.