Explorer and exporter were scanning the full Parquet lake every 1–5s
(~2000 files, 200MB+), driving ~2.2 CPU cores. Limit metrics to the
last N flights, cache DuckDB views and the partition tree, slow live
polls to 2s, and keep drones alive after seal to stop restart churn.
Add node-exporter, scan-duration metrics, and a Swarm Platform Grafana
dashboard for node CPU/memory and scan health.
Use the real hive column name (`drone`) in the prototype's live query
and pin the Grafana datasource uid to `swarm-prom` so provisioned
dashboards resolve their Prometheus source in k3d.
Ansible provisions the k3d cluster and Flux controllers; Terraform
modules deploy the simulated fleet and ground warehouse. Compose and
k3d share var/t1 (live lake) and var/t3 (warehouse). The prototype
gains a live mode fed by the explorer read-only SQL API.
The field was a fixed square, leaving wide monitors with dead side
margins. Field width now follows the window aspect (capped at 2.2:1,
quantized to avoid rebuilds on minor resizes); routes, obstacles and
transit spawns stretch with it, and the grid keeps 100 m cells.
Mobile obstacles used to bounce until they wedged into far corners, and
a single perimeter route let the mesh stretch into segments that stopped
exchanging data. Even drones now fly the perimeter while odd drones fly
an X through the center, keeping bridge links alive, and transit objects
enter at one edge, cross the area, and respawn at a fresh edge.
Top-down SVG view: patrol flight with separation and obstacle avoidance,
mesh links appearing within radio range, pose-broadcast flashes, bulk
sync transfers with live rate and cumulative up/down counters, channel
hopping and battery per drone. Pure client-side simulation (Vite/React/TS).
eSlider