Milvus
Distributed vector database built for billion-scale — the choice once one node stops being enough.
Milvus separates storage and compute and scales horizontally, which is what you need past a few hundred million vectors. The operational cost is real: plan for Kubernetes, not a docker-compose file.
Best for: Billion-scale, horizontally scaled vector search
Deploy: Self-hostable
What to do about it
- Check licence terms
- Verify data handling
- Estimate operational cost
What it is
Milvus is a vector database that scales horizontally. It separates storage and compute, which allows it to handle large amounts of data.
This separation is key to its design. It is built for large-scale deployments, past a few hundred million vectors.
Where it fits in a deployment stack
Vector databases like Milvus solve the problem of efficient similarity search in high-dimensional spaces. They normally sit next to data ingestion pipelines and machine learning models in a production system.
They take over the work of indexing and querying large datasets, allowing for fast and accurate retrieval of similar items. This is a critical component in many applications, such as image and video search, recommender systems, and natural language processing.
These databases are often used in conjunction with other data storage systems, such as relational databases and NoSQL databases, to provide a comprehensive data management solution.
How to evaluate it
To evaluate Milvus, you need to consider its performance, scalability, and ease of use. This involves testing its capabilities with your specific use case and data.
- Measure p95 latency of the retrieval step under 50 concurrent sessions
- Feed it a dataset of 1 million vectors and measure the indexing time
- Inject a node failure and verify the system's ability to recover
- Inspect the query results for accuracy and relevance
- Compare the performance of different indexing algorithms
- Verify the consistency of query results across multiple nodes
What to verify before you commit
Before committing to Milvus, you need to verify its suitability for your specific use case and environment. This involves checking its licence, data handling, operational cost, failure behaviour, and potential lock-in.
- Check the licence terms for any restrictions on usage or distribution
- Verify how it handles data encryption and access control
- Estimate the operational cost of running it on your infrastructure
- Test its behaviour in case of node failures or network partitions
- Evaluate the potential lock-in to a specific cloud or infrastructure provider
- Check for any dependencies on specific libraries or frameworks
Where this fits in the FDE path
- FDE career path — the levels, the titles that map to them, and what changes at each step
- Interview guide — what the loop tests beyond coding
- Pay data — 223 postings — midpoint median $231,000, employer-published ranges only
- 922 live FDE roles — pulled daily from employers’ own job boards
RAG in Production Playbook
Chunking strategies, reranking setups, evaluation harnesses and the seven failure modes that kill every RAG demo in week three.