ベクトルデータベースのコスト最適化

Vector workloads are dominated by memory. Indexes like HNSW keep vectors in RAM, so cost scales with vector count times dimensions times precision.

Approximate memory = vectors x dimensions x bytes_per_value, plus index overhead. Cutting any factor cuts cost.

Scalar quantization stores each value as 1 byte (int8) instead of 4 (float32), cutting memory ~4x with minimal recall loss.

Product quantization (PQ) compresses vectors into compact codes, achieving much higher compression at some accuracy cost. Great for huge datasets.

Binary quantization stores each dimension as a single bit, giving up to 32x compression. Use with a reranking step to recover accuracy.

Many models support Matryoshka embeddings or PCA, letting you truncate dimensions (e.g. 1536 to 768) with little quality loss and big savings.

Keep hot vectors in RAM and cold vectors on disk or object storage. Offload inactive tenants or old data to cheaper tiers.

Each replica multiplies memory cost. Use replicas for availability and read throughput only as needed; autoscale on real traffic.

HNSW is fast but RAM-heavy. IVFFlat or disk-based indexes (DiskANN) trade some speed for far lower memory cost on large corpora.

Track cost per million vectors and queries per dollar. Benchmark recall before and after each optimization so quality stays acceptable.

A practical sequence:

• Reduce dimensions if the model allows
• Apply scalar or binary quantization
• Tier cold data to disk
• Add reranking to recover recall

Test your cost knowledge.

You learned to cut vector DB costs with quantization, dimensionality reduction, tiered storage, and smart index and replica choices, while measuring recall to protect quality.