Posts tagged with "production"

How a 20,000-example few-shot similarity pipeline exposed silent full-table scans in pgvector — and what it took to fix them. Covers missing vector_size, operator class mismatches, HNSW migrations with Alembic, iterative scans, quantization, and production observability.

How to fix XGBoost out-of-memory errors when training multi-class models with thousands of classes on GPUs: gradient-pair matrix, three strategies (CPU, One-vs-Rest, Hierarchical Softmax), and when to use each.