Agent Platform Feature Store vs Metarank
Agent Platform Feature Store
cloud.google.comMetarank
metarank.aiSide by side
| Figure | Agent Platform Feature Store | Metarank |
|---|---|---|
| Price from | — | — |
| Free plan | — | — |
| Free trial | — | — |
| Open source | — | — |
| Web | Yes | — |
| Windows | — | Yes |
| macOS | — | Yes |
| Linux | — | Yes |
| API | Yes | Yes |
| Self-hosted | — | Yes |
| Plans published | 1 | 0 |
| Spec score · Feature Store Software | 7.4 (#8) | 7.4 (#10) |
Plans
Agent Platform Feature Store
- New Agent Platform Feature Store (pay-as-you-go)Price not published
In their own words
Agent Platform Feature Store
Agent Platform Feature Store is a managed cloud service for storing and serving machine-learning features from BigQuery. BigQuery tables and views act as the offline store, while online serving returns current feature values for low-latency predictions. The service supports online and offline stores, point-in-time joins, and feature monitoring. Registering sources and columns in the Feature Registry is optional; registration enables column selection, source combination, and statistics or drift monitoring. Monitoring can run manually or on a schedule to detect anomalies. Bigtable online serving handles large volumes and near-real-time continuous sync from BigQuery but does not support embeddings. Optimized online serving supports scheduled sync, is deprecated, and is scheduled to sunset on February 17, 2027. Google recommends Vector Search for embeddings. IAM controls access to feature groups, online-store instances, and feature views. Resources must share a region or multi-region with their BigQuery source. The pay-as-you-go plan is usage-based; listed node rates include $0.08 per hour for data processing, $0.30 for optimized serving, and $0.94 for Bigtable serving.
Metarank
Metarank is an open-source ranking service for personalized semantic search and recommendations. It uses customer activity, including clicks and purchases, to adjust ranking and personalize results in real time. Capabilities include recommendations, LambdaMART reranking, automatic feature generation and retraining, and serving multiple models for A/B testing. It can use large language models in bi-encoder and cross-encoder modes for semantic search and reranking. Events can arrive through Apache Kafka, Apache Pulsar, AWS Kinesis Streams, a REST API, or files; files are supported for import, not inference. Persistence options include local memory, a local directory, Redis, and S3. It runs as a Docker container or JAR on Linux, Windows, and macOS, and provides APIs for feedback ingestion, training, ranking, recommendations, LLM inference, and Prometheus metrics. Production guidance recommends Redis persistence, while multiple API instances require an external load balancer because clustering is not provided out of the box.