Best Practices for Multitenancy, Multiple Vectors, and Columnar Storage

General Discussions - English
1

Hi,

I’m currently designing the architecture for a new project using Manticore Search and I would love to get your advice on a few key points to ensure I’m following the best practices.

Here are my questions:

1. Multi-tenancy Strategy I need to implement a multi-tenant system. I’m considering two main approaches:

  • Approach A: Use a single, large table with a tenant_id attribute. All queries would then need a WHERE tenant_id = ... filter.
  • Approach B: Create a separate table for each tenant (e.g., docs_tenant_1, docs_tenant_2, etc.).

My question is: what is the generally recommended approach for performance, scalability, and ease of management? If I choose Approach B, is it possible and efficient to perform a single query that searches across multiple (or all) tenant tables at once?

2. Multiple Vector Columns in a Single Table My documents have distinct semantic features that I would like to capture in separate vector fields. Is it possible for a single table to have more than one vector column?

If this schema is valid, can a single KNN query perform a search against both vector columns simultaneously? For example, finding a document that is a close match for a query vector on vector_1 and/or another query vector on vector_2 within the same search operation, or would this require two separate searches?

3. Columnar Storage for Vectors at Scale I am expecting to have around 15 million documents in the main table. Given this scale, would you recommend defining the vector fields with engine='columnar' to optimize for RAM usage? What are the performance trade-offs (if any) for KNN search speed when using columnar storage for vectors compared to the default row-wise storage?

thanks.

2
  1. If you mostly need to search within a single tenant (which is usually the case in multi-tenant setups), I’d go with separate tables. You can still create a distributed table that combines all tenants if you need to search across them. Just make sure that all IDs are unique in that case.
  2. Yes. Here’s an example:
mysql> drop table if exists test; create table test ( title text, v1 float_vector knn_type='hnsw' knn_dims='4' hnsw_similarity='l2' quantization='1bit', v2 float_vector knn_type='hnsw' knn_dims='4' hnsw_similarity='l2' ); insert into test values ( 1, 'yellow bag', (0.653448,0.192478,0.017971,0.339821), (-0.148894,0.748278,0.091892,-0.095406) ), ( 2, 'white bag', (-0.148894,0.748278,0.091892,-0.095406), (0.653448                                                                                                                                                                                                       nn_
--------------
drop table if exists test
--------------

Query OK, 0 rows affected (0.01 sec)

--------------
create table test ( title text, v1 float_vector knn_type='hnsw' knn_dims='4' hnsw_similarity='l2' quantization='1bit', v2 float_vector knn_type='hnsw' knn_dims='4' hnsw_similarity='l2' )
--------------

Query OK, 0 rows affected (0.01 sec)

--------------
insert into test values ( 1, 'yellow bag', (0.653448,0.192478,0.017971,0.339821), (-0.148894,0.748278,0.091892,-0.095406) ), ( 2, 'white bag', (-0.148894,0.748278,0.091892,-0.095406), (0.653448,0.192478,0.017971,0.339821) )
--------------

Query OK, 2 rows affected (0.00 sec)

--------------
select * from test
--------------

+------+------------+---------------------------------------+---------------------------------------+
| id   | title      | v1                                    | v2                                    |
+------+------------+---------------------------------------+---------------------------------------+
|    1 | yellow bag | 0.653448,0.192478,0.017971,0.339821   | -0.148894,0.748278,0.091892,-0.095406 |
|    2 | white bag  | -0.148894,0.748278,0.091892,-0.095406 | 0.653448,0.192478,0.017971,0.339821   |
+------+------------+---------------------------------------+---------------------------------------+
2 rows in set (0.00 sec)
--- 2 out of 2 results in 0ms ---

--------------
select id, knn_dist() from test where knn ( v1, 5, (0.286569,-0.031816,0.066684,0.032926), { ef=2000, oversampling=3.0, rescore=1 } )
--------------

+------+------------+
| id   | knn_dist() |
+------+------------+
|    1 | 0.28146550 |
|    2 | 0.81527930 |
+------+------------+
2 rows in set (0.00 sec)
--- 2 out of 2 results in 0ms ---

--------------
select id, knn_dist() from test where knn ( v2, 5, (0.286569,-0.031816,0.066684,0.032926), { ef=2000, oversampling=3.0, rescore=1 } )
--------------

+------+------------+
| id   | knn_dist() |
+------+------------+
|    2 | 0.28146550 |
|    1 | 0.81527930 |
+------+------------+
2 rows in set (0.00 sec)
--- 2 out of 2 results in 0ms ---
  1. Definitely use quantization for RAM optimization (released in version 13.2.3 just a few days ago), along with oversampling and rescoring (to keep the accuracy). Columnar storage can also help improve performance.
1 Like
3

Thank you for your advice, I’m still running some tests and so far with excellent results.

I read in the issues that embedding has been implemented in the latest version, but I can’t find any documentation on how to use it. Is it possible to have some examples?

thanks.

4

It’s not fully set up yet: it can automatically create embeddings for documents, but not for queries yet. We’re still working on that part.

5

understand, thank you. For now, it would be useful for me to know how automatic embedding works.

I would need to know if I can configure any model / endpoint or only the local open-source ones?

6

You can check out this PR with tests covering the new functionality (WIP) - Tests: auto-embeddings in columnar by donhardman · Pull Request #3472 · manticoresoftware/manticoresearch · GitHub

You should be able to use any embedding model from Hugging Face which is compatible with GitHub - huggingface/candle: Minimalist ML framework for Rust which Manticore uses to convert texts to embeddings. There’s also an integration with OpenAI API.

1 Like