# Vector search on object storage: Performance at scale without the infrastructure tax

Most vector databases require three systems: raw data in a lake, metadata in a warehouse, embeddings in a vector index. Three copies of state. Three things to keep in sync.

LanceDB stores everything in one table on object storage. Compute nodes are stateless. No RAM constraints at scale.

## Why teams switch

### Compute-storage separation  
Data lives on object storage at $0.02/GB/month. Compute scales with query load, not data size. 10 TB of data, one small query node during off-peak. No paying for idle capacity.

### One table. Actual data  
Embeddings, metadata, and raw blobs in the same table. Not links to S3. Blobs. Vector search, full-text search, and SQL filtering compose into a single query. No round trips.

### Write a new column without rewriting the table  
Adding a column doesn't rewrite existing data. Zero copy. New embedding model? Access control column? Test two models side by side? Column-level operations, not table-level rewrites.

### IVF-based indexing  
Inserts go to the appropriate partition without touching others. Deletes handled by lightweight bitmaps. MVCC for concurrent reads and writes. Index rebuilds happen asynchronously.

## Comparison

| Feature    | Traditional Vector Database                         | LanceDB                                       |
|------------|---------------------------------------------------|----------------------------------------------|
| **Cost**   | RAM-bound. Replication for availability means 2-3x storage. Thousands/month at 100M vectors. | Object storage at $0.02/GB/month. Only hot index in memory. Stateless compute. |
| **Scale**  | What has to stay hot? Index structures, caches, metadata, allocator overhead all in RAM. | Large index persisted to disk. Lance format optimized for random access. Small hot set in memory. |
| **Search**  | Embeddings and metadata in vector DB. Raw docs/images in S3. Second call for originals. | Vector, full-text, and SQL in one query against one table. Raw blobs stored inline. |
| **Data model**  | Three systems: lake for raw data, warehouse for metadata, vector DB for embeddings. | One table. Embedding is a column. Metadata in other columns. Raw binary in another column. |
| **Schema changes** | New column means rewriting every row. Full rewrite tax every time your app evolves. | Column-level operations. Existing columns untouched. Zero copy. |
| **Best for** | Sub-millisecond p99 at any cost. Small static datasets. Zero infrastructure decisions. | Best cost-to-performance at scale. Teams that want to understand and control their infrastructure.

## The Power of the Lance Format

### Vector Search  
- Fast scans and random access from the same table — no tradeoff  
- Zero-copy access for high throughput without serialization overhead

### Multi-Modal  
- Raw data, embeddings, and metadata in one table — not pointers to blob storage  
- No separate metadata store to keep in sync

## Enterprise-Grade Requirements

### Security  
Granular RBAC, SSO integration, and VPC deployment options.

### Governance  
Data versioning and time-travel capabilities for auditability.

### Support  
Dedicated technical account management and guaranteed SLAs.

### Talk to Engineering  
Or try LanceDB OSS — same code, scales to Cloud.