Samuel Partee describes Redis as a composable database whose modules add document storage, full-text search, vector search, and other capabilities to its original caching layer.
2
Redis can act as an online feature store and vector database, which lets teams add ML retrieval and recommendation features without introducing another service into an existing Redis stack.
3
Partee connects his Redis work to high-performance computing, distributed training, scalable inference, and a climate-modeling project that streamed features through Redis and ran a PyTorch model online.
Summary
Samuel Partee explains how Redis is expanding from a web cache into a platform for ML serving. He describes Redis modules for JSON, search, and vector search, then connects them to online feature stores and recommendation systems. The practical argument is that teams already using Redis can add embeddings and similarity search without adding another database or service. Partee also discusses the challenge of explaining infrastructure value to product managers, so he builds complete demos that make faster retrieval visible in an application. The conversation then moves into Partee's high-performance computing background, including distributed training, hyperparameter optimization, scalable inference, and work with climate models. In that project, Redis carried numerical features between a distributed Fortran model and a PyTorch model. He is candid that dedicated vector products may offer different features, while arguing that Redis has advantages when it is already part of a company's stack.
Redis is adding layers above its original caching use case
Partee describes Redis as a three-layer system. The first is the open source caching foundation, traditionally placed in front of a durable database to speed up retrieval. The second is the module ecosystem. RedisJSON adds document-database behavior, RediSearch adds full-text search, and vector search adds similarity queries. The third is a change in how the platform is used, including online feature stores. Redis Stack bundles the base database with popular modules in one container. Partee says developers responded strongly because it is easy to deploy and can support different kinds of applications.
Redis entered ML data use cases because customers were already using it there
The move toward AI and ML came from observing how companies were using Redis. Partee points to DoorDash's real-time route optimization and recommendations, Uber's Michelangelo work, and other large companies asking whether Redis Enterprise could support feature-store workloads. He says Redis offered single-digit millisecond retrievals, which made it useful for ML pipelines that needed fast feature access. Redis then started building more support around those existing patterns instead of treating AI as an unrelated product direction.
Complete applications make infrastructure benefits easier to understand
Partee says it is difficult to communicate the value of a database through latency claims alone. His vector search demo turns unstructured data into embeddings, stores them in Redis, and lets users find similar images or text. A full-stack application makes the behavior visible to people who may not understand the infrastructure underneath. He describes a React and TypeScript frontend with a FastAPI backend running in a small Docker-based deployment. The point of the demo is to let engineers and nontechnical stakeholders see what the database enables.
Redis vector search can add similarity queries without changing an existing stack
Partee explains that vector databases store embeddings, which are lists of numbers produced by ML or deep learning models. Redis Search can index those vectors inside Redis and support similarity or distance comparisons. This is useful when a company already uses Redis as a cache, since storing vectors there can add search and recommendation behavior without a new dependency. He gives the example of a retailer that wanted product pages to show similar items using both text and image attributes. Weighted vector search can combine those signals to produce more varied recommendations.
The advantage over a dedicated vector service can be operational simplicity
When Mihail Eric asks about Pinecone and Weaviate, Partee frames the difference around existing platform use. Redis may already be deployed, paid for, and approved inside a company's infrastructure. Adding vector search can therefore avoid another service, dependency, or purchasing process. Partee acknowledges that other products may have features Redis does not, while saying that vector search is a major area of focus. He also mentions Redis Enterprise capabilities such as geographic distribution and flash storage for large embedding tables.
Partee's background in high-performance computing shaped his interest in Redis
Partee traces his interest in speed to early work with Cray supercomputers and parallel programming. He worked with Chapel, distributed versions of NumPy, distributed feature selection, hyperparameter optimization, and distributed training. That work led to scalable inference and the SmartSim framework, which exposed him to Redis. He enjoys optimizing collective communications and MPI even when the improvement is only a few percent. The conversation also compares Kubernetes with HPC schedulers such as Slurm, with Partee describing a possible combination of Slurm's scheduling strengths and Kubernetes' APIs and deployment flexibility.
Redis supported an online ML loop inside a climate model
Partee describes a paper created with the National Center for Atmospheric Research, the University of Victoria, and HPE collaborators. The project used Redis as an online feature store alongside a large, distributed Fortran climate model. Numerical features were streamed into Redis, where a PyTorch model predicted eddy kinetic energy. That prediction was streamed back into the model. Partee says the added prediction improved the model's skill by supplying information that was missing from its lower-resolution version. The setup spread a Redis instance across 128 nodes and used about eight terabytes of memory.
Feature platforms can hide some of the complexity of distributed systems
In the closing lightning round, Partee says he is bullish on feature platforms. He connects them to the same simplification that makes Hugging Face's model encoding workflow easy to use. Feature platforms can reduce the amount of distributed-system detail teams must handle, especially when several databases support different purposes. His answer also fits the broader discussion: the infrastructure still matters, but useful abstractions can make it easier for teams to build and deploy ML applications.
"I think they're really important, and the larger abstractions that make things easier for people make a lot of these distributed systems that we used to have to think for about four days much simpler."Samuel Partee46:36
Who should watch
You already run Redis and want to assess whether it can support embeddings, similarity search, or online feature retrieval without adding another database.
You are building ML infrastructure and need a clear comparison between an existing data platform and a dedicated vector service.
Your work involves distributed training, scalable inference, HPC scheduling, or connecting ML models to scientific simulations.
