Architecture Case Studies & Evidence Framework
Deep-dive technical retrospectives analyzing real platform engineering challenges. Structured around the canonical decision framework: Problem → Decision → Implementation → Trade-offs → Evidence → Result.
GeoCrop: Crop Classification from Satellite Rasters on Self-Hosted Infrastructure
Processing multi-band Sentinel-2 satellite tiles (B02, B03, B04, B08, 10m resolution) across agricultural regions traditionally causes out-of-memory (OOM) kills on cloud VMs, while commercial Earth Observation APIs impose unsustainable per-query licensing costs.
Architected a streaming windowed raster inference pipeline using rasterio block windows, computing exactly 51 immutable spectral/phenological features (FEATURE_ORDER_V1), and executing quantized XGBoost/LightGBM models on CPU inside strict 1GB RAM limits.
FastAPI gateway orchestrates asynchronous jobs via Redis Queue (RQ). Workers stream Sentinel-2 COGs from Digital Earth Africa STAC catalog, perform windowed ndarray feature extraction, run model inference, and export Cloud Optimized GeoTIFFs directly to in-cluster MinIO S3.
Windowed raster processing increases total inference duration by ~14% compared to full-scene in-memory arrays, but guarantees zero OOM container crashes and allows multi-tenant concurrency on modest 6-core nodes.
Over 180 field-scale inference jobs completed. Zero OOM terminations recorded in Prometheus cAdvisor telemetry. TiTiler dynamic map tiler serves sub-50ms web tiles directly from MinIO COG outputs.
Zero external cloud API costs. Full ownership of the pipeline end-to-end — from raw Sentinel-2 ingestion to classified output — at 10m spatial resolution.