HOW IT WORKS: THE MILITARY TELEMETRY PIPELINE

1.1 Dual-Feed Meteorological Sourcing

Modern military decision-making demands both high-resolution forward projections and legally authoritative, certified airfield observations. The system queries two primary data providers simultaneously:

• Open-Meteo Numerical Weather Prediction (NWP): Queries surface temperature, apparent temperature, relative humidity, precipitation, weather code (WMO), surface pressure, wind vectors, wind gusts, cloud cover layers (low/mid/high), cloud base, solar direct radiation, and 7-day numerical forecast projections.
• NOAA Aviation Weather Center (AWC) Live METARs: Official METAR observations from co-located military and joint-use airfields (e.g., KPOB at Fort Liberty, KHLR at Fort Cavazos, KFBG at Fort Moore, EDDN for USAG Bavaria). Provides certified ceiling, runway visibility, flight categories, altimeter settings, and present weather codes.
• Order of Battle (OOB) Enrichment: Every observation is joined with installations_oob.json to attach the host command (e.g., FORSCOM, USARPAC, USAREUR-AF), tenant combat formations, primary weapon platforms (Apache, Abrams, HIMARS), and mission profile.

1.2 Batch Query Optimization & Network Fallback

# All 43 installation coordinates queried in 1 single HTTP GET
params = {
    "latitude": "41.3918,35.1391,31.1352,...",
    "longitude": "-73.9563,-79.0060,-97.7758,...",
    "current": "temperature_2m,dew_point_2m,relative_humidity_2m,"
               "wind_speed_10m,wind_gusts_10m,visibility,cloud_cover",
    "hourly": "temperature_2m,dew_point_2m,cloud_base,direct_radiation",
    "forecast_days": "7",
    "timezone": "auto"
}
# Typical execution: ~1.2s across all 43 installations worldwide

Network Resilience: If the primary batched request fails due to temporary connection drops on the Pi, the engine automatically falls back to sequential per-base queries with exponential backoff (retrying at 2s, 4s, and 8s).

4.1 What is Predicate Pushdown?

In traditional database querying, all rows from disk are loaded into RAM before applying a WHERE filter. On a Raspberry Pi, this wastes CPU cycles and exhausts memory.

With Predicate Pushdown, DuckDB pushes the filter conditions directly into the storage scan operator:

┌─────────────────────────────────────┐
│             READ_PARQUET            │
│  Function: READ_PARQUET             │
│  Projections: base_id, temp_c, ...  │
│  Filters: base_id = 'usma_west_point│  <-- PUSHED DOWN!
│  Partition Filter: year='2026'      │  <-- SKIPS OTHER YEARS!
└─────────────────────────────────────┘

4.2 Historical REST API Endpoints