Estimate memory

[Karcherry]

Is there any equation to roughly estimate memory usage based on different data scales and index configurations? It is indeed helpful for building index to prepare sufficient resources and reduce OOM cases.

[cutecutecat]

After some research, we found that the memory usage during the K-means phase of internal builds (vchordrq index) often fluctuates significantly (±50%) due to the feature of the system memory allocator. It is difficult to estimate the memory usage. Therefore, we only recommend using the internal builds (vchordrq index) for small data (fewer than 5 million rows, 768 dim) and assuming memory usage will not exceed 32 GB.

For large-scale data, you can try these methods if memory usage is a concern:

• External build (vchordrq index): Move the memory-consumed K-means out from PostgreSQL.
• vchordg index (since 0.5.0 / preview): It does not consume extra memory beyond shared_buffers, providing a build experience similar to pgvector. Say goodbye to OOM.

[Karcherry]

@cutecutecat That sounds great. I know maintenance_work_mem is used to control the memory when building HNSW in pgvector. So shared_buffers is now used to control memory in the new index type? I'd like to give vchordg a try.

[winstxnhdw]

Hey, I am also curious about this. Is VectorChord recommended for use in low-resource systems? e.g. < 1GB RAM? RedisStack is consistently able to provide great performance while using < 30MB with a few hundred documents.

[VoVAllen]

@winstxnhdw Yes, note it's just a postgres instance, so low resource should work well. With few hundred documents, usually it doesn't need a index, and Vectorchord vchordrq can still help here.