Changelog

The full fleet management dashboard ships in this release. Realtime device telemetry via Server-Sent Events, 6 dashboard pages (Overview, Devices, Deployments, Events, Drift, Settings), JWT authentication, and the complete REST API backing it.

• Fleet Overview — live device count, online/offline/drift status, deployment health at a glance
• Device management — per-device telemetry, model version, hardware class, last-seen, latency history
• Deployment page — staged rollout controls, canary targeting by hardware class, health gate configuration
• Drift monitor — KL divergence timeline per device, alert thresholds, rollback controls
• Events log — full audit trail, filterable by device, event type, severity
• SSE live updates — all dashboard pages refresh without polling; 30s heartbeat from each device

Statistical drift detection now runs on the device itself, not in the cloud. Kullback-Leibler divergence computed via a sliding circular buffer with integer arithmetic optimised for ARM Cortex-A class CPUs. Under 2ms per check on Jetson Nano.

• Configurable warn (0.4) and alert (0.7) KL divergence thresholds
• Monitors input distribution, feature layer activations, and output confidence independently
• Upstream alert fires within 4 minutes of threshold crossing in production tests
• Baseline distribution captured automatically at model activation time
• Sliding window of 200 inferences; recomputes every 100 — 60μs average overhead at 30fps

Stage-based canary rollouts with per-hardware-class targeting. Define stages by device class (jetson_orin, jetson_nano, raspberry_pi5) not just percentage. Health gates block progression if accuracy drops more than a configurable delta.

• Hardware fingerprint captured at agent registration (CPU arch, RAM, accelerators, OS)
• Multi-variant model registration — one deployment version, multiple format files (TensorRT/TFLite/ONNX)
• Health gate on accuracy delta, latency delta, error rate — configurable per stage
• Automatic rollback on health gate failure — no manual intervention required
• mlops deploy v4.0 --stage 1:hw_class=jetson_orin,count=1 --health-gate accuracy_delta=-0.03

Complete rewrite of the sync protocol from push-based to pull-based. The agent polls the control plane for desired state, reconciles idempotently, and buffers telemetry locally during outages. Binary delta compression reduces model update transfer size by 95.7% on typical version-to-version updates.

• Pull-based sync loop — agent owns the deployment, not the CI/CD pipeline
• Local SQLite buffer — 500MB cap, chronological flush on reconnect
• bsdiff delta compression — 42.3MB model → 1.8MB delta on v3.1→v3.2 update
• SHA-256 checksum verification before activation — corrupt transfer = safe fallback, no silent failure
• Tested on 6-hour outage on 2G connection — zero data loss, full reconciliation on reconnect

Systematic binary size reduction to make the agent viable on constrained devices. 84.3% reduction from initial build to production binary.

• -ldflags="-s -w" — strip debug symbols and DWARF info (−12MB)
• musl libc cross-compilation — genuinely static binary, no glibc dependency (−8MB)
• UPX --best --lzma compression — 50ms startup overhead, acceptable for long-running agent (−14MB)
• Build tags for optional features — Prometheus, gRPC streaming excluded from minimal build (−4MB)
• gorilla/websocket → nhooyr.io/websocket, logrus → zerolog (−1.6MB)