Databricks AI Summit 2023 Block Session

• Block

Block walked through real examples from its apps, Square (payments), Cash App, Tidal, and TBD.

Federation

The first building block is federation. You hide the complexity behind standard interfaces so different apps and databases can talk to each other. Here Databricks sits in the middle as the gateway, much like Unity Catalog.

Data streaming and compute

Every app sends all of its data to Kafka (run by Confluent), and Databricks does the compute. Databricks also holds the historical data, which feeds the ML platform.

Composability

Composability means building out of small components that work together. You break a big problem into functional pieces and solve them one at a time. Their examples were Marketing ML for personalized business experiences and Risk ML for fraud detection.

Lakehouse layers

The Lakehouse is a stack of layers.

1. Federated storage, connected to the compute layer (Databricks jobs and clusters).
2. A feature platform that turns business problems into values you can act on.
3. A model training layer, with MLflow for features and training, so data is stored and processed one standard way.
4. Model serving.

How Block fills those layers

Payment aggregation lives in the compute and federated storage layers, and features like connected users feed the fraud patterns. Risk training sits on the MLflow layer, and marketing sits on the model training layer. Several feature providers plug into the system. The LLM layer, with LangChain enriching context from the data, sits on model serving.

Open source models

The Lakehouse supports open source models, so you can swap between CPU, GPU, and models like MPT.

Applications

They talked about form filling, code generation, and fine-tuning.