▥ Metrics Fieldbook

Small systems / careful observations

Three nodes.
One quiet lab.

A compact notebook for infrastructure health, model routing and the work between deployments.

Illustrative lab: all hosts and measurements on this site are fictional, synthetic demo data — not live telemetry.

Node inventory

Reference snapshot · not live

cedar-01

Operational

Inference gateway

Anthropic · Claude

Uptime
47 days, 08 hours
30-day availability
99.98%
CPU load18%
Memory used42%

Routes short interactive requests. A compact worker pool keeps queue latency predictable; long-running jobs go elsewhere.

basalt-02

Operational

Batch processing

DeepSeek

Uptime
23 days, 16 hours
30-day availability
99.91%
CPU load34%
Memory used61%

Processes overnight classification and summarization batches. Jobs checkpoint between chunks rather than holding one large request open.

cirrus-03

Operational

Evaluation sandbox

GLM

Uptime
12 days, 03 hours
30-day availability
99.84%
CPU load12%
Memory used36%

Runs multilingual evaluation suites against pinned prompts. Results are reviewed before any routing policy changes.

Recent operator notes

Maintenance / October 3, 2026

Leave room for the recovery path

Capacity planning has to include retries, log rotation and the next deployment, not just the happy-path request.