Multi-tenancy with Milvus

[sgaseretto]

I want to know if it will be a good practice to use collections and partitions for tenants and users. At first glance it seems that it will work, but I want to know what could be the drawbacks of using Milvus that way. One example I could think about is a SaaS for documents search, where a collection is assign to an Organization (tenant) and inside that collection we assign partitions to users inside that organization.

[shiyu22]

There is a feature issue about multi-tenancy, you can follow it #254.

[sgaseretto]

Great @shiyu22 ! But imagine you have a backend that will handle the access and queries to the Milvus DB. You can achieve a sort of filtering with collections and partitions (at least that is what I understand). So at first glance I imagine you can assingn a collection to a tenant and then partitions to its users, and build indices for those collections and partitions. My question is if this can be considered as a good practice or doing this will degrade Milvus performance?

[gujun720]

Hi, sgaseretto

One of the challenges of embeddings similarity matching is memory consumption. If you want better performance, then you have to allocate adequate memory to the collections.

If there is not enough memory, Milvus will swap the least recently used data out of the memory. At this moment, there is no option to pin data in memory. (If you think it's important, please raise a feature request ticket. It will be prioritized if more community users would upvote it.)

So if you want to assign partitions to different users, you should be aware that the QoS could not be guaranteed if there is not enough memory reserved for all the users.

Or you might consider sharding the data into different Milvus nodes. We are continuously improving Mishards.