MongoDB Vector Search Architecture

The mongot Process

Each mongot process has its own persistent volume that is not shared with the database or other search nodes. Storage is used to maintain vector indexes that are built from the data continuously sourced from the database. The index definitions (metadata) are stored in the database itself.

The mongot performs the following actions:

• Manages the vector index.

  The mongot is responsible for updating the index definitions in the database.

• Sources the data from the database.

  The mongot nodes establish permanent connections to the database in order to update indexes from the database in real time.

• Processes vector search queries.

  When mongod receives a $vectorSearch query, it directs the query to one of the mongot nodes. The mongot that receives the query processes the query, aggregates the data, and returns the results to mongod, which forwards the results to the user.

The mongot components are tightly coupled with a single MongoDB replica set and cannot be shared across multiple databases or replica sets. Each replica set deployment has its own dedicated search nodes.

Networking

Network connectivity between mongot and mongod goes in both directions:

• mongot establishes connections to the replica set to source the data used to build indexes and run queries.

• mongod connects to mongot to forward search-related operations such as index management and querying the data.

The mongod acts as the proxy for all search queries. You never interact directly with the mongot.

Operator-Managed Deployment

When both the mongot and mongod processes are deployed inside the Kubernetes cluster, the Kubernetes Operator performs configuration for both processes automatically. Specifically, the Kubernetes Operator:

1. Finds the MongoDB CR referenced by MongoDBSearch using spec.source.mongodbResourceRef, or by a naming convention by looking for the MongoDB CR with the same name as MongoDBSearch.

2. Generates mongot configuration in a YAML file and saves it to a config map named <MongoDBSearch.metadata.name>-search-config.

3. Deploys the Vector Search stateful set named <MongoDBSearch.metadata.name>-search with storage and resource requirements configured according to spec.persistence and spec.resourceRequirements in the CR.

4. Updates configuration of every mongod process by adding the necessary setParameter options, including the hostnames and port numbers of the mongot hosts.

External MongoDB Deployment

When the MongoDB replica set is outside of Kubernetes, you deploy mongot using the Kubernetes Operator and perform some steps manually. The Kubernetes Operator handles configuration of the search pods, but you must reconfigure your MongoDB nodes and the networking.

Security

If the MongoDB server is inside the Kubernetes cluster, the Kubernetes Operator automatically sets up keyfile authentication for Vector Search. If the MongoDB server is external, you must create a Kubernetes Secret containing the replica set's keyfile credential and reference it in the MongoDBSearch CR.

Limitations

You can't deploy Vector Search on the following architectures:

• IBM Power (ppc64le)

• IBM Z (s390x)