벡터 DB 저장 아키텍처

When we talk about vector databases, we're really talking about storing and searching those special number lists called embeddings (or vectors). Just like your regular files need a hard drive, vectors need a place to live.

But not all storage is created equal! The way a vector database stores its data deeply impacts how fast it can find similar vectors and how much data it can handle.

You might wonder why we can't just use a normal database to store vectors. The challenge is that vector databases need to do something very specific and very fast: similarity search.

• High-Dimensional Data: Vectors are long lists of numbers (hundreds or thousands!). Storing them efficiently is key.
• Fast Comparisons: Finding "similar" vectors means comparing many of these long lists quickly. This requires specialized indexing and retrieval strategies, which depend heavily on the underlying storage.

The fastest way to access data is to keep it in your computer's Random Access Memory (RAM). In-memory vector databases do exactly this.

• How it works: All vector data and their indexes are loaded directly into RAM when the database starts.
• Pros: Unmatched speed for queries, instant access.
• Cons: Limited by available RAM, data is lost if the system restarts (unless explicitly saved to disk), more expensive per gigabyte than disk storage.

In-memory storage is perfect for situations where speed is paramount and data persistence isn't the primary concern, or where the dataset is small enough to fit comfortably in RAM.

• Small Datasets: When your collection of vectors is manageable (e.g., thousands or a few million).
• Temporary Caching: Storing frequently accessed vectors as a "hot cache" to speed up responses.
• Rapid Prototyping: Quick experiments where setting up persistent storage is overkill.

For larger datasets that need to survive restarts, vector databases use disk-based storage, typically on solid-state drives (SSDs) or traditional hard disk drives (HDDs).

• How it works: Vectors and their indexes are written and read from disk, just like regular files.
• Pros: Data persistence (it stays even after a power off!), can handle very large datasets, generally cheaper per gigabyte than RAM.
• Cons: Slower query speeds compared to in-memory, as reading from disk takes more time.

Most production-ready RAG applications rely on disk-based storage as their primary vector store. This ensures data integrity and the ability to scale to vast amounts of information.

• Large-Scale RAG: Storing billions of document chunks for comprehensive knowledge bases.
• Primary Data Store: The main, durable repository for all your vector embeddings.
• Cost-Effective: A practical choice when you need to store a lot of data without breaking the bank.

Many advanced vector databases use a hybrid approach, intelligently combining in-memory and disk-based storage. Think of it like your computer's operating system using RAM for active programs and disk for everything else.

This strategy aims to get the best of both worlds: fast access for frequently used data and persistence for the entire dataset.

What happens when your vector dataset is so huge it can't fit on a single machine, or when you need super high availability? That's where distributed storage comes in.

• How it works: Data is split into smaller pieces (shards) and spread across many different servers, often in a cluster.
• Pros: Massive scalability (can grow almost infinitely), high fault tolerance (if one server fails, others can take over), high availability.
• Cons: Increased complexity in setup and management, network latency can impact performance.

Distributed vector databases are the backbone of large-scale AI applications that handle enormous amounts of data and require uninterrupted service.

• Petabyte-Scale Data: When you have truly massive collections of vectors.
• High Availability: For mission-critical applications where downtime is unacceptable.
• Global Reach: Distributing data geographically for faster access in different regions.

You're building a RAG system for a small internal company knowledge base (10,000 documents) where quick responses are important, but the data must be persistent. Which storage architecture is generally the most practical choice for the primary vector store in this scenario?

We've explored the different ways vector databases store their data, each with unique trade-offs:

• In-Memory: Fastest, but volatile and capacity-limited.
• Disk-Based: Persistent, scalable to large datasets, good balance for most needs.
• Hybrid: Combines in-memory for speed with disk for persistence.
• Distributed: For massive scale and high availability across many machines.

Choosing the right architecture depends on your specific needs for speed, persistence, and scalability. Next, we'll dive into the algorithms that make similarity search fast!