Feast¶

Feast is an open source feature store framework that manages and serves machine learning features for training and inference. The ScyllaDB integration provides a Feast online store for low-latency, distributed, real-time feature serving, with built-in support for vector search. It materializes feature values into a self-hosted ScyllaDB cluster or ScyllaDB Cloud.

ScyllaDB is used as an online store only. Feast still requires a separate offline store for historical feature retrieval. The ScyllaDB integration does not implement an offline store.

Installation¶

Install Feast with the scylladb extra, which pulls in scylla-driver automatically:

pip install feast[scylladb]

Configuration¶

Configure ScyllaDB as the online store in your feature_store.yaml. Set online_store.type to scylladb and provide the connection details for your cluster.

Self-hosted ScyllaDB:

project: scylla_feature_repo
registry: data/registry.db
provider: local
online_store:
  type: scylladb
  hosts:
    - 172.17.0.2
  keyspace: feast
  username: scylla
  password: password

ScyllaDB Cloud:

When connecting to ScyllaDB Cloud, set local_dc to the datacenter name shown on the Connect tab of your cluster in the ScyllaDB Cloud Console.

project: scylla_feature_repo
registry: data/registry.db
provider: local
online_store:
  type: scylladb
  hosts:
    - node-0.aws_us_east_1.xxxxxxxx.clusters.scylla.cloud
    - node-1.aws_us_east_1.xxxxxxxx.clusters.scylla.cloud
    - node-2.aws_us_east_1.xxxxxxxx.clusters.scylla.cloud
  keyspace: feast
  username: scylla
  password: xxxxxx
  local_dc: AWS_US_EAST_1

Configuration Options¶

Parameter	Type	Default	Description
`hosts`	list[str]	(required)	Contact-point host addresses.
`port`	int	`9042`	CQL port.
`keyspace`	str	`feast_keyspace`	Target ScyllaDB keyspace.
`username`	str	`None`	Auth username.
`password`	str	`None`	Auth password.
`local_dc`	str	`None`	Local datacenter name for DC-aware load balancing.
`request_timeout`	float	`None`	Driver request timeout in seconds.
`read_concurrency`	int	`100`	`concurrency` argument passed to the driver’s `execute_concurrent_with_args` for reads. Controls how many CQL statements are in-flight at once.
`write_concurrency`	int	`100`	`concurrency` argument passed to the driver’s `execute_concurrent_with_args` for writes. Controls how many CQL statements are in-flight at once.
`vector_similarity_function`	str	`COSINE`	Default similarity function for vector indexes. Supported: `COSINE`, `DOT_PRODUCT`, `EUCLIDEAN`. Can be overridden per-feature via the `similarity_function` Field tag.

Vector Search¶

ScyllaDB Cloud supports approximate nearest-neighbour (ANN) vector search. To enable it for a feature view, tag the embedding Field with vector_index=true and specify the number of dimensions:

from feast import FeatureView, Field
from feast.types import Array, Float32, String

documents_fv = FeatureView(
    name="documents",
    entities=[item],
    schema=[
        Field(name="text", dtype=String),
        Field(
            name="embedding",
            dtype=Array(Float32),
            tags={
                "vector_index": "true",
                "dimensions": "768",
                "similarity_function": "COSINE",  # COSINE | DOT_PRODUCT | EUCLIDEAN
            },
        ),
    ],
    online=True,
    source=push_source,
)

When feast apply runs, it automatically creates the necessary tables and ANN index for any feature view with vector-tagged fields.

To query the top-k most similar documents:

result = store.retrieve_online_documents_v2(
    features=["documents:text", "documents:embedding"],
    query=[0.1, 0.2, ...],   # your query embedding
    top_k=10,
    distance_metric="COSINE",
)

The distance_metric accepts the following values:

COSINE: cosine similarity.
DOT_PRODUCT: dot-product (inner-product) similarity.
EUCLIDEAN: Euclidean (L2) distance.

Additional Resources¶

Was this page helpful?