How to Assign More Memory to Weaviate Database Process via NUMA

General
,
1

Hi community,

I am Decheng. I have successfully deployed Verba with a local Weaviate database on my machine. Since Weaviate runs in memory, I want to assign more memory to the Weaviate process using NUMA.

I found that llamaindex can manage vector memory, as shown in this example:

llamaindex Vector Memory Example

In this example, when vector_store is set to None, VectorMemory creates a default in-memory vector store. This means all data is stored in memory, providing fast access but using a lot of memory and not persisting data after the program ends.

I want to achieve similar memory management for Weaviate. Does this approach make sense, and can I use a similar method to assign more memory to the Weaviate process via NUMA?

Thank you!

2

Hi!

By default, using hnsw index, Weaviate will store your vectors in memory.

With 1.25 we introduced dynamic index. If configured your collection starts with a flat index, and when hitting a threshold, it will convert automatically to hnsw, leveraging in memory index.

Other ways of saving on memory is using PQ or, in case you have a multi tenant scenario, loading and off loading tenants on demand.

Let me know if this helps.

Thanks!