Other day i was trying out Azure SQL for storing vector DB. After a POC i was able to get results on sub 1 second to get the results with 2 core serverless instance. It begs me to reconsider does dedicated Vector db really worth it ?
Here is the article on how to store and query vector data https://devblogs.microsoft.com/azure-sql/vector-similarity-s...
Ask yourself the same question about an OLAP system. Once you’re at a large enough scale, it makes sense to use tools custom designed for your access patterns.
I work on Azure SQL, but it’s pretty clear that this is a cool demo. It may scale for some usages, but I wouldn’t build a production system on it.
Hyperscale is awesome. My point is more for vector search. Once you’re at a large enough scale, the simple algorithms for similarity search break down. You’d need something that implements HNSW and designed their database indexes around it.
The article says you had 500ms latency while searching through only 25k records. Presumably at 1 query at a time? That's probably 2 orders of magnitude slower than a vector database.
Comments
Other day i was trying out Azure SQL for storing vector DB. After a POC i was able to get results on sub 1 second to get the results with 2 core serverless instance. It begs me to reconsider does dedicated Vector db really worth it ? Here is the article on how to store and query vector data https://devblogs.microsoft.com/azure-sql/vector-similarity-s...
Ask yourself the same question about an OLAP system. Once you’re at a large enough scale, it makes sense to use tools custom designed for your access patterns.
I work on Azure SQL, but it’s pretty clear that this is a cool demo. It may scale for some usages, but I wouldn’t build a production system on it.
Well tried it with ~ 34 million records , which is good enough sample for our usecase
At what number of queries per second?
You mean something like the Hyperscale service tier? This is what we are building our new products on top of.
Hyperscale is awesome. My point is more for vector search. Once you’re at a large enough scale, the simple algorithms for similarity search break down. You’d need something that implements HNSW and designed their database indexes around it.
The article says you had 500ms latency while searching through only 25k records. Presumably at 1 query at a time? That's probably 2 orders of magnitude slower than a vector database.
Is that good or bad?
~ 34 million embeddings , Seems this is a good enough sample for our use case.