RonDB
Product · API / developer platform · Data & analytics by Hopsworks
A high-performance, real-time database with SQL capabilities for AI workloads.
Category: Database
- APIs and SDKs
- REST API, SQL, Key-Value API, Python client
- Peak Throughput
- 100M key lookups/sec
- Batch Size Limit
- 4096 rows
- P99 Latency (11 features)
- 7.5ms
- Architecture Type
- In-memory with on-disk support
- Query Interfaces
- SQL, REST API, Key-Value API
- Sub-millisecond latency for real-time AI serving
- 100M key lookups per second in REST API benchmarks
- Pushdown LEFT-JOINs and aggregations to the kernel
- Batch processing for up to 4096 rows per request
- Automated TTL for data cleanup and management
Overview
RonDB is a high-performance, real-time key-value store with SQL capabilities designed specifically for real-time AI workloads. The database is engineered for low-latency workloads using an in-memory architecture that also supports optional on-disk tables. RonDB distinguishes itself by supporting pushdown operations—including projections, predicates, LEFT-JOINs, and aggregations—directly in the database kernel to minimize data transfer and latency. In peer-reviewed benchmarks, it has demonstrated the capacity to handle 100 million key lookups per second via its REST API. It is available as part of the Hopsworks platform for businesses requiring enterprise-grade ML infrastructure.
Database Performance
- In-Memory Architecture: RonDB is an in-memory database engineered for low-latency workloads, with optional support for on-disk tables.
- Pushdown Operations: The database supports pushdown projections, predicates, LEFT-JOINs, and aggregations to the database kernel to reduce latency.
- REST API Server: RonDB 24.10 includes a REST API Server that can boost performance by 25% to 900% for real-time queries.
- Automated TTL Management: The system automates data cleanup through Time-To-Live (TTL) settings for feature data.
Real-Time AI Capabilities
- Online Feature Serving: RonDB serves as the online store for the Hopsworks Feature Store, providing sub-millisecond latency for real-time serving.
- Large Batch Processing: The database handles batch primary key lookups for up to 4096 rows in a single operation.
- Flexible Query Support: RonDB supports multi-part primary keys and general SQL for custom real-time AI use cases.
- Rondis Integration: The Rondis integration simplifies the process of connecting applications to real-time data.
Use cases
- E-commerce platforms power TikTok-style recommendation systems by performing batch primary key lookups for ranking items.
- Cybersecurity teams evaluate network packets for intrusion by augmenting weak signals with contextual data from RonDB.
- Banks perform real-time transaction scoring by joining live signals with historical customer data stored in the database.
Specifications
- APIs and SDKs
- REST API, SQL, Key-Value API, Python client
- Platforms
- Web, API, Kubernetes (AWS EKS, Google GKE, Azure AKS, OVHCloud)
- Limits
- Batch processing up to 4096 rows; demonstrated 100M key lookups/sec throughput
- Deployment
- Cloud (AWS, Azure, GCP), On-premises, Air-gapped, Managed SaaS
- Integrations
- Hopsworks Feature Store, Redis, AWS DynamoDB (comparison), GCP BigTable (comparison)
- Peak Throughput
- 100M key lookups/sec
- Batch Size Limit
- 4096 rows
- P99 Latency (11 features)
- 7.5ms
- Architecture Type
- In-memory with on-disk support
- Query Interfaces
- SQL, REST API, Key-Value API
- Performance Boost (REST API)
- 25% to 900%
- Aggregation Support
- Pushdown to database kernel
- Primary Key Support
- Multi-part primary keys
How it works
- Deployment and Setup: RonDB is typically deployed as the online storage layer within the Hopsworks platform, either as a managed SaaS or on-premises.
- Data Ingestion: Feature pipelines write data to the online store, where RonDB stores the latest feature values per primary key.
- Real-Time Retrieval: Real-time AI systems issue queries via REST API or SQL to retrieve historical and contextual features using entity identifiers.
- Feature Vector Construction: The database executes pushdown operations and returns feature vectors to the AI model within the required service level objective (SLO).
FAQ
How does RonDB differ from other databases used for feature stores?
RonDB is an in-memory database engineered for low-latency workloads, whereas many other feature stores use DynamoDB or Redis. It supports SQL capabilities and pushdown operations like LEFT-JOINs that are not typically available in simple key-value stores.
What is the maximum batch size for lookups?
RonDB 24.10 supports batch primary key lookups for up to 4096 rows in a single operation, which is useful for the ranking step in AI recommender systems.
Does RonDB support on-disk storage?
Yes, while RonDB is primarily an in-memory database for performance, it also supports on-disk tables if your workload requires them.
What kind of throughput can RonDB handle?
In a REST API benchmark, RonDB demonstrated the ability to handle 100 million key lookups per second.
Added by AntSpace from public sources (hopsworks.ai).