Multitenancy and localizing data on specific shard

[boltnev]

Hello.

We would like to try Qdrant in our project for search with scalar filtering and want to understand how well it will fit our needs.

Our data consists of several hundred of thousands independent groups: some groups contain millions of points, while others contain tens of thousands. We always searching within specific group.

We would like to understand whether it is worth using multitenancy to improve search performance.

If we use an index with is_tenant, are all tenant data stored in one shard, or is the data of one tenant distributed across multiple shards?

Will we gain performance benefits over multitenancy if we manually specify the used shards during upsert and search operations?

[timvisee]

If we use an index with is_tenant, are all tenant data stored in one shard, or is the data of one tenant distributed across multiple shards?

When using an is_tenant: true payload index, all data is still spread across all shards. That is assuming you're using the (default) automatic sharding strategy. If you're using user-defined sharding your data might land in any of the shards associated with the given shard key. is_tenant: true does not try to isolate values into separate shards. Instead it tries to group points (on segment level) having the same tenant value close together on disk or in memory. That makes it a lot more disk and CPU cache friendly.

Long story short: if using is_tenant: true, your data is still spread across all shards.

Will we gain performance benefits over multitenancy if we manually specify the used shards during upsert and search operations?

If you're talking about using user-defined sharding as well, then I'm not entirely sure. If you have just a few tenants which you can isolate into shard keys, then I'd say yes, because it'd limit the search scope.

However, you mention having thousands of independent groups with wildly different sizes. Shards (each replica) has a bit of overhead, in which case it might become quite costly if you need thousands of them. I expect the cost to outweigh the potential performance savings.

In your case I highly recommend to just use payload based multitenancy with is_tenant: true configured on the associated payload index.

It's probably best to do an experiment to show if such configuration meets your performance expectations and requirements.

[boltnev]

Thank you for clarification