Configure compute and storage for vector workloads

What this covers

How to choose and change the compute SKU, memory and storage of an Azure Database for PostgreSQL flexible server so a vector workload has enough RAM for its indexes, enough CPU for distance calculations and enough IOPS for the rest. It also covers the storage behaviours the exam likes to test: storage can only grow, autogrow, the read-only threshold, and the fact that both the VM and the disk cap IOPS.

Exam weight: Develop AI solutions by using Azure data management services (30–35%)

Skills measured

• "Configure compute, memory, and storage resources to support vector workloads"

Key concepts

Size from the vector index outward

1. Vector data ≈ dims × 4 bytes × rows (halfvec: × 2 bytes). 1,536 dims × 2M rows ≈ 12 GB.
2. Index ≈ +50–100% for HNSW (1.5–2× vector data in total), less for IVFFlat. DiskANN with product quantization keeps a compressed copy in memory.
3. RAM target: the hot index plus working set should fit in memory with headroom. The module suggests RAM at least 1.5× the vector index for 1–10M vectors.
4. vCores: concurrency × per-query CPU. Sustained CPU above about 70% at peak means scale up.

Compute SKUs

vCores can be scaled up or down. The change needs a short restart, so connections drop and apps must retry. You can scale up temporarily for a big index build or backfill and scale down after.

VM limits on IOPS

Every SKU has a maximum IOPS and throughput, for example D4ds_v5 about 6,400 IOPS and 145 MiB/s, E8ds_v5 about 12,800 IOPS. A fast disk can't exceed the VM's cap, so a small VM on a large disk is still IO-bound. Check Disk IOPS consumed percentage and Storage IO percent.

Storage options

Premium SSD v2 is usually the better price/performance choice for IO-heavy vector work, because you can buy IOPS without buying capacity.

Storage rules the exam tests

• Storage only scales up, never down. Over-provisioning is permanent until you migrate.
• At 95% used, or less than 5 GiB free, the server switches to read-only. Writes fail; backups and WAL archiving can suffer.
• Storage autogrow (Premium SSD) grows the disk before it fills.
• Alert on storage_percent, and include index size and WAL in the plan. A new HNSW index on a large table can add gigabytes at once.

Memory-related parameters

shared_buffers is set per SKU and changes when you resize. maintenance_work_mem should be raised for index builds. work_mem should be set per session. See Tune PostgreSQL for pgvector.

How it works

Hands-on

Azure portal

1. Server, Compute + storage.
2. Select Compute tier (Memory Optimized for large vector indexes) and Compute size.
3. Choose Storage type (Premium SSD v2 for tunable IOPS), set size, IOPS and throughput.
4. For Premium SSD, enable Storage Auto-growth.
5. Save, then create an alert on Storage percent > 80.

Azure CLI

# Scale compute (short restart)
az postgres flexible-server update -g <resource-group> -n <server-name> \
  --tier MemoryOptimized --sku-name Standard_E8ds_v5

# Grow storage (cannot shrink later) and turn on autogrow (Premium SSD)
az postgres flexible-server update -g <resource-group> -n <server-name> \
  --storage-size 512 --storage-auto-grow Enabled

# Premium SSD v2: tune IOPS and throughput independently
az postgres flexible-server update -g <resource-group> -n <server-name> \
  --iops 12000 --throughput 500

az postgres flexible-server show -g <resource-group> -n <server-name> \
  --query "{sku:sku.name, tier:sku.tier, storage:storage}"

-- Measure before sizing
SELECT pg_size_pretty(pg_total_relation_size('document_chunks')) AS total,
       pg_size_pretty(pg_indexes_size('document_chunks'))        AS indexes;
SHOW shared_buffers;

Python

def recommend_sku(rows: int, dims: int, half: bool = False, hnsw_factor: float = 2.0) -> dict:
    """Rough sizing helper: RAM >= 1.5x (vectors + HNSW)."""
    vec_gb = rows * dims * (2 if half else 4) / 1024**3
    working_gb = vec_gb * hnsw_factor
    ram_needed = working_gb * 1.5
    # Memory Optimized v5: 8 GiB per vCore
    vcores = next(v for v in (2, 4, 8, 16, 20, 32, 48, 64, 96) if v * 8 >= ram_needed)
    return {"vector_gb": round(vec_gb, 1), "ram_target_gb": round(ram_needed, 1),
            "suggested": f"Standard_E{vcores}ds_v5"}

print(recommend_sku(2_000_000, 1536))

Exam traps

Quick review

• Vector GB = dims × 4 × rows; HNSW roughly doubles the total.
• General Purpose 4 GiB per vCore; Memory Optimized about 6.75–9.5 GiB per vCore.
• Compute scales up and down with a short restart; storage only scales up.
• Premium SSD v2: up to 80,000 IOPS and 1,200 MB/s, tuned separately from size.
• VM SKUs cap IOPS regardless of disk.
• Read-only at 95% storage or under 5 GiB free; enable autogrow and alerts.

Practice

Practice: compute, memory and storage questions

Official references

• AI-200 study guide
• Compute options
• Storage options
• Unit: Scale for high-volume workloads