Municipal architecture standards require different node traits for permit OCR enrichment versus wide resident-service joins. What selection principle should engineers follow?

Municipal architecture standards that require different node traits for permit OCR enrichment versus wide resident-service joins should follow the principle of choosing CPU-optimized workers for CPU-bound OCR enrichment and memory-optimized workers for wide joins. OCR/PDF/image-style enrichment is typically CPU-intensive per task, while wide joins spill and shuffle memory-heavy intermediates—so node family selection should track the bottleneck, not a one-size municipal default.

Always picking the cheapest burstable node regardless of OCR or join profile optimizes sticker price over performance and often increases runtime cost through spills and timeouts. Using GPU ML nodes for every CSV parse including tiny lookup tables misapplies accelerators and inflates spend for workloads that do not need GPUs. Forcing single-executor clusters for both OCR and wide joins “to simplify billing” destroys parallelism required for civic batch scale and does not address CPU vs memory trait differences.

Exam caveat: map workload profile → node type (CPU vs memory vs GPU) before picking SKUs. GPUs appear often as distractors for non-ML SQL/ETL. Spot/cheap nodes can be valid for fault-tolerant batch, but not when the stem asks for trait-matching principles across OCR vs wide joins.

Operational check: profile permit OCR CPU utilization versus resident-service join spill metrics, assign separate job-cluster policies (CPU-optimized vs memory-optimized), and validate runtime and cost against mixed “cheapest node” baselines before standardizing architecture.