벡터 DB 영속성과 확장성

In this lesson, we'll explore two crucial concepts for any production-ready vector database: Persistence and Scalability.

Imagine building a powerful search engine. You wouldn't want to lose all your indexed data every time the system restarts, right? That's where persistence comes in!

Persistence means your data survives even if the application or server shuts down. It's saved to a durable storage like a disk, not just kept in temporary memory.

• Why it matters: Prevents data loss.
• Without it: All your carefully generated vector embeddings would vanish on restart.
• Goal: Ensure data durability and reliability.

A common way to achieve persistence is by saving the vector index and associated data directly to files on a disk. This is often seen in local or embedded vector databases.

• How it works: Data is written to specific file formats (e.g., binary files, HDF5).
• Pros: Simple to implement for smaller datasets.
• Cons: Can be slower for very large datasets, manual management required.

Let's see a simple example of a persistent vector store using ChromaDB. It saves your collections to a local directory, so your data is safe across sessions.

Some vector databases use traditional databases (like PostgreSQL or SQLite) as their underlying persistence layer. The vector data might be stored in a special column type (e.g., pgvector extension for PostgreSQL).

• Benefits: Leverages existing database features like transactions, backups, and replication.
• Example: pgvector allows PostgreSQL to store and query vector embeddings efficiently.

Scalability refers to a system's ability to handle a growing amount of work—more data, more users, more queries—without a significant drop in performance.

• Why it matters: Your application's success means more data and users.
• Goal: Maintain fast search speeds and reliability as your system grows.

Vertical scaling (or 'scaling up') means adding more resources (CPU, RAM, faster storage) to a single server. Think of it as making one machine super powerful.

• Pros: Often simpler to manage initially.
• Cons: There's a limit to how powerful a single machine can be. It also creates a single point of failure.

Horizontal scaling (or 'scaling out') means distributing your data and workload across multiple servers. This is how large-scale cloud services operate.

• Sharding: Splitting your entire dataset into smaller, independent pieces (shards) and storing each shard on a different server.
• Replication: Creating copies of your data/index across multiple servers for fault tolerance and to handle more read requests.

Choosing a scaling strategy involves trade-offs:

• Cost: More servers generally mean higher costs.
• Complexity: Distributed systems are inherently more complex to design, build, and maintain.
• Performance: Scaling can improve performance but also introduce network latency.
• Consistency: Ensuring all copies of data are identical can be challenging in distributed systems.

You've learned about the critical roles of persistence and scalability in vector databases. Persistence ensures your valuable embeddings are never lost, while scalability allows your system to grow and handle increasing demands.

Understanding these concepts helps you choose and design robust vector database solutions for production applications.