# Feature stores, and second thoughts

75 sessions · follows the tags feature-stores, feature-engineering
Page: https://mlopstalks.com/threads/feature-stores

Feature stores promised to fill a gap in the ML platform. People who built and ran them explain where they helped, and why they sometimes reconsidered.

## 2020

6 sessions.

- [Mid-Scale Production Feature Engineering](https://mlopstalks.com/talks/mid-scale-production-feature-engineering) (Dr. Venkata Pingali, Scribble Data). Pushed back: Dr. Venkata Pingali rejects the idea that feature engineering is a simple linear process in which data is merely fed into a model. [31:38](https://www.youtube.com/watch?v=1CcYuVVwOGg&t=1898s)
- [Running a Fintech on Machine Learning](https://mlopstalks.com/talks/running-a-fintech-on-machine-learning) (Caique Lima & Cristiano Breuel, Nubank). Pushed back: The team favored building its feature store internally rather than buying an existing product because it wanted tighter integration and customization, and found no sufficiently mature market solution. [46:45](https://www.youtube.com/watch?v=ry_P5D_d7XA&t=2805s)
- [Feature Stores: An Essential Part of the ML Stack to Build Great Data](https://mlopstalks.com/talks/feature-stores-an-essential-part-of-the-ml-stack-to-build-great-data) (Kevin Stumpf, Tecton). Pushed back: Kevin Stumpf rejects the idea that Tecton should be an end-to-end machine-learning platform, arguing that a focused feature platform can integrate with best-in-class tools for other parts of the workflow. [53:00](https://www.youtube.com/watch?v=IjO8VUCIZxc&t=3180s)
- [A Conversation Around Feature Stores](https://mlopstalks.com/talks/a-conversation-around-feature-stores) (Venkata Pingali, Scribble Data). Pushed back: Venkata Pingali said the future design space would contain many different feature-store architectures rather than one standard implementation. [45:42](https://www.youtube.com/watch?v=ajCd6k_OZO0&t=2742s)

2 more from 2020 on this thread: https://mlopstalks.com/threads/feature-stores/2020

## 2021

13 sessions.

- [Real-time Feature Pipelines, A Personal History](https://mlopstalks.com/talks/real-time-feature-pipelines-a-personal-history) (Hendrik Brackmann, Tide). Pushed back: Hendrik Brackmann says the usefulness of storing dynamic model outputs as features depends on reuse, training design, and system boundaries rather than having one universal answer. [45:02](https://www.youtube.com/watch?v=M1p1uJNbHUg&t=2702s)
- [Machine Learning Feature Store Panel Discussion](https://mlopstalks.com/talks/machine-learning-feature-store-panel-discussion) (Daniel Galinkin, iFood & Matias Dominguez, Rappi & Simarpal Khaira, Intuit). Pushed back: Whether a company needs to buy or build a feature store depends on its use cases and maturity rather than having a single universal answer. [9:29](https://www.youtube.com/watch?v=-TGp2qKz8tA&t=569s)
- [Machine Learning Design Patterns for MLOps](https://mlopstalks.com/talks/machine-learning-design-patterns-for-mlops) (Valliappa Lakshmanan, Google Cloud). Claim: A feature store is useful when multiple models need to reuse the same calculated features, but many use cases only need a library. [13:04](https://www.youtube.com/watch?v=vH7UFZZdja8&t=784s)
- [Building ML Blocks with Kubeflow Orchestration with Feature Store](https://mlopstalks.com/talks/building-ml-blocks-with-kubeflow-orchestration-with-feature-store) (Aniruddha Choudhury, Publicis Sapient). Pushed back: Aniruddha Choudhury says Feast is not a workflow scheduler, data warehouse, pipeline orchestrator, feature engineering tool, or model-serving product. [16:10](https://www.youtube.com/watch?v=UJLQgxr_Za0&t=970s)

9 more from 2021 on this thread: https://mlopstalks.com/threads/feature-stores/2021

## 2022

22 sessions.

- [Feature Stores at Shopify and Skyscanner](https://mlopstalks.com/talks/feature-stores-at-shopify-and-skyscanner) (Matt Delacour, Shopify & Mike Moran, Skyscanner). Pushed back: Matt Delacour says feature engineering is outside the feature store project, while the question concerned whether feature services contain feature-engineering code. [25:26](https://www.youtube.com/watch?v=v42YVgPGKro&t=1526s)
- [Building a Movie Recommendation System on Tecton with Snowflake](https://mlopstalks.com/talks/building-a-movie-recommendation-system-on-tecton-with-snowflake) (David Hershey, Tecton). Pushed back: David Hershey says Tecton does not currently provide built-in automatic drift detection, while external tools are commonly used to profile feature data. [47:52](https://www.youtube.com/watch?v=eqOjdldInjA&t=2872s)
- [Feathr: LinkedIn's High-performance Feature Store](https://mlopstalks.com/talks/feathr-linkedins-high-performance-feature-store) (David Stein, LinkedIn). Pushed back: Real-time features should not be treated as universally necessary because many signals and prediction targets change slowly. [40:04](https://www.youtube.com/watch?v=wHbyT4w_uOc&t=2404s)
- [Recommender System: Why They Update Models 100 Times a Day](https://mlopstalks.com/talks/recommender-system-why-they-update-models-100-times-a-day) (Gleb Abroskin, FunCorp). Pushed back: Gleb Abroskin rejected the description of FunCorp's system as a unified feature store with declarative transformations and one API for offline and online stores. [15:00](https://www.youtube.com/watch?v=mgZTm1ZG04k&t=900s)

18 more from 2022 on this thread: https://mlopstalks.com/threads/feature-stores/2022

## 2023

7 sessions.

- [Tecton 0.6: Notebook-driven Development](https://mlopstalks.com/talks/tecton-0-6-notebook-driven-development) (Jason Dunne, Tecton). Claim: Notebook-driven development can help data engineering, data science, and ML engineering teams improve feature engineering workflows and shorten iteration loops. [4:34](https://www.youtube.com/watch?v=PC5fbgTt4hQ&t=274s)
- [MLOps Build or Buy, Startup vs. Enterprise?](https://mlopstalks.com/talks/mlops-build-or-buy-startup-vs-enterprise) (Aaron Maurer & Katrina Ni, Slack). Claim: Slack can often add a new recommender quickly by adjusting the features, feature weights, and candidate-fetching method for a use case. [16:10](https://www.youtube.com/watch?v=IC2uilYf1sc&t=970s)
- [The Future of Feature Stores and Platforms](https://mlopstalks.com/talks/the-future-of-feature-stores-and-platforms) (Mike Del Balso, Tecton & Josh Wills, Angel Investor). Pushed back: Mike Del Balso said feature templates and reusable feature sets are appealing, but customers' differing data and requirements make simple copy-and-paste solutions unlikely to solve most problems. [1:06:26](https://www.youtube.com/watch?v=p5F7v-w4EN0&t=3986s)
- [Product Strategy for LLM Features When LLMs Aren't Your Product](https://mlopstalks.com/talks/product-strategy-for-llm-features-when-llms-arent-your-product) (Harini Kannan). Pushed back: LLM-based feature engineering should not be viewed only as a prototyping or brainstorming technique; it can be suitable for production in some cases. [11:30](https://www.youtube.com/watch?v=1FUV9VGVCGw&t=690s)

3 more from 2023 on this thread: https://mlopstalks.com/threads/feature-stores/2023

## 2024

16 sessions.

- [AI Innovations: The Power of Feature Platforms](https://mlopstalks.com/talks/ai-innovations-the-power-of-feature-platforms) (Mahesh Murag, Tecton & Jose Navarro, Cleo & Nikhil Garg, Fennel). Pushed back: Nikhil Garg says a feature store and a feature platform are different: a feature store mainly provides storage and serving, while a feature platform also includes computation and other end-to-end capabilities. [44:51](https://www.youtube.com/watch?v=xI_uQF072GE&t=2691s)
- [RecSys at Spotify](https://mlopstalks.com/talks/recsys-at-spotify) (Sanket Gupta, Spotify). Pushed back: A vector database is not always a separate downstream destination for embeddings; in some online use cases it and the feature store serve overlapping roles. [38:08](https://www.youtube.com/watch?v=byH-ARJA4gk&t=2288s)
- [BigQuery Feature Store](https://mlopstalks.com/talks/bigquery-feature-store) (Nicolas Mauti, Malt). Pushed back: Malt's BigQuery feature-store approach is not appropriate for every use case because live feature computation, very fresh data, or data too large for memory may require other systems. [44:39](https://www.youtube.com/watch?v=NtDKbGyRHXQ&t=2679s)
- [Global Feature Store: Optimizing Locally and Scaling Globally at Delivery Hero](https://mlopstalks.com/talks/global-feature-store-optimizing-locally-and-scaling-globally-at-delivery-hero) (Gottam Sai Bharath & Cole Bailey, Delivery Hero). Pushed back: The participants disagreed over the definitions of a feature and a feature store before aligning on shared definitions. [20:34](https://www.youtube.com/watch?v=c9L8vAvCUyA&t=1234s)

12 more from 2024 on this thread: https://mlopstalks.com/threads/feature-stores/2024

## 2025

6 sessions.

- [Scaling your data and AI from 0-100 with open source](https://mlopstalks.com/talks/scaling-your-data-and-ai-from-0-100-with-open-source) (Maarten Breddels, Pycafe & Pranav Aurora, Mooncake & Simba Khadder, Featureform). Pushed back: The speakers disputed the assumption that production feature pipelines should be rewritten and maintained by a separate data or ML team after data scientists hand over notebooks. [52:54](https://www.youtube.com/watch?v=Ja3t2NtDQfo&t=3174s)
- [Hard Learned Lessons from Over a Decade in AI](https://mlopstalks.com/talks/hard-learned-lessons-from-over-a-decade-in-ai) (Mike Del Balso, Tecton). Claim: Tecton's feature store centralized and automated the data pipelines needed to create training sets and provide real-time model features. [5:07](https://www.youtube.com/watch?v=tgGjFmrmxE0&t=307s)
- [Real-time Feature Generation at Lyft](https://mlopstalks.com/talks/real-time-feature-generation-at-lyft) (Rakesh Kumar, Lyft). Pushed back: A conventional feature store is not sufficient for Lyft's hierarchical, geospatially aggregated features. [23:57](https://www.youtube.com/watch?v=saAezbQqarc&t=1437s)
- [Context Engineering 2.0](https://mlopstalks.com/talks/context-engineering-2-0) (Simba Khadder, Redis). Pushed back: Featureform's acquisition by Redis does not signal the death of the feature store. [0:25](https://www.youtube.com/watch?v=J7-TZCOpZVI&t=25s)

2 more from 2025 on this thread: https://mlopstalks.com/threads/feature-stores/2025

## 2026

5 sessions.

- [Graph Neural Networks Just Solved Enterprise AI?](https://mlopstalks.com/talks/graph-neural-networks-just-solved-enterprise-ai) (Jure Leskovec, Stanford University and Kumo.AI). Pushed back: Relational deep learning is fundamentally different from AutoML that generates many joins, aggregates, features, and model trials. [17:34](https://www.youtube.com/watch?v=Tq613N0dyI0&t=1054s)
- [The Semantic Layer and AI Agents](https://mlopstalks.com/talks/the-semantic-layer-and-ai-agents) (David Jayatillake, Cube.dev). Pushed back: Whether feature stores and semantic layers should be treated as essentially the same type of system. [46:09](https://www.youtube.com/watch?v=c-i3yfaoh6k&t=2769s)
- [Real-time features, AI search, Agentic similarities](https://mlopstalks.com/talks/real-time-features-ai-search-agentic-similarities) (Varant Zanoyan & Nikhil Simha Raprolu, Zipline AI). Pushed back: Varant Zanoyan argued that Feather's documentation described the right problems but its implementation had not completed important compute features such as streaming and windowed aggregation. [12:01](https://www.youtube.com/watch?v=kmcYPvLynKo&t=721s)
- [OpenXData Conference](https://mlopstalks.com/talks/openxdata-conference) (Will Angel). Claim: An online feature store provides consistent feature values for training and inference together with discovery, lineage, access control, and versioning. [2:02:14](https://www.youtube.com/watch?v=YInxWv_4N7w&t=7334s)

1 more from 2026 on this thread: https://mlopstalks.com/threads/feature-stores/2026