Vector workload sizing

Size compute, memory, and storage together

Vector workloads can stress CPU, memory, storage capacity, IOPS, and throughput at the same time. Size for the heaviest operation, which may be a large vector index build rather than steady query traffic.

Give index builds enough memory

maintenance_work_mem controls memory for maintenance operations such as CREATE INDEX and VACUUM. Its default is calculated when the server is provisioned and does not automatically follow later compute scaling.

SET maintenance_work_mem = '4GB';
SET max_parallel_maintenance_workers = 4;
CREATE INDEX ...;

The parameter is dynamic, so tune it carefully for the build session within the memory the server can safely provide.

Match the compute tier to sustained load

Burstable relies on CPU credits and can throttle under sustained vector workloads. General Purpose or Memory Optimized is a better fit when load is continuous.

Protect storage headroom

Storage can grow but not shrink. Premium SSD v2 lets you tune capacity, IOPS, and throughput separately, while the compute size can still cap usable IOPS. Keep free space available so ingestion and maintenance do not push the server into read-only protection.