Vector databases such as ElasticSearch and Pinecone offer fast ingestion and querying on vector embeddings with ANNs. However, they typically do not decouple compute and storage, making them hard to integrate in production data stacks. Because data storage in these databases is expensive and not easily accessible, data teams typically maintain ETL pipelines to offload historical embedding data to blob stores. When that data needs to be queried, they get loaded back into the vector database in another ETL process. This is reminiscent of loading data from OLTP database to cloud storage, then loading said data into an OLAP warehouse for offline analytics.

Recently, “lakehouse” offerings allow direct OLAP querying on cloud storage, removing the need for the second ETL step. The same could be done for embedding data. While embedding storage in blob stores cannot satisfy the high TPS requirements in online settings, we argue it’s sufficient for offline analytics use cases like slicing and dicing data based on embedding clusters. Instead of loading the embedding data back into the vector database for offline analytics, we propose direct processing on embeddings stored in Parquet files in Delta Lake. You will see that offline embedding workloads typically touch a large portion of the stored embeddings without the need for random access.

As a result, the workload is entirely bound by network throughput instead of latency, making it quite suitable for blob storage backends. On a test one billion vector dataset, ETL into cloud storage takes around one hour on a dedicated GPU instance, while batched nearest neighbor search can be done in under one minute with four CPU instances. We believe future “lakehouses” will ship with native support for these embedding workloads.