AWS re:Invent 2025 - Powering Real-Time Applications with a Modernized Cache (API210)

Powering Real-Time Applications with a Modernized Cache

Reddis: An In-Memory Data Store

• Reddis is an in-memory data structure store used as a cache, session store, message broker, database, and more
• Reddis has over 2.25 million Docker pulls per day and has been voted the most loved non-relational database by developers

Addressing Common Reddis Challenges

Hot Keys

• Hot keys occur when keys are accessed or written to very frequently
• Reddis addresses this with client-side caching, keeping a copy of hot keys in the client library to avoid network round trips

Secondary Indexing

• Users were using sets and sorted sets in Reddis for secondary indexing use cases
• This also led to hot key issues, with the key space becoming imbalanced
• Reddis introduced the Reddis Query Engine to enable schema-based indexing of fields in hashes and JSON, allowing complex queries without hot keys

Unlocking New Use Cases

Session Store and Feature Store

• The Reddis Query Engine enables new use cases like querying the number of users with a specific item in their shopping basket
• It also powers Reddis' vector search capabilities, enabling fast vector similarity search
• This powers use cases like a "semantic caching" solution that caches responses from large language models

Modernizing the Cache

Cache Aside Pattern

• The cache aside pattern involves caching data from a relational database in Reddis
• This can lead to stale data and "thundering herd" issues when the cache expires
• Customers are moving towards a "refresh ahead" or "prefetch" cache pattern instead

Reddis Data Integration (RDI)

• RDI is a low-code solution for synchronizing data from any data source (e.g. Oracle) to the appropriate Reddis data structures
• This allows customers to move all their data into Reddis, not just a subset, and keep it continuously synchronized

Reddis Flex: Autotiering for Scale and Cost

• Reddis Flex extends the key space with slower storage (SSD, NVMe) to enable databases up to 50TB
• This reduces costs by up to 75% compared to an in-memory only approach
• Key use cases include feature stores, where RDI can keep the feature data always in sync

Native JSON Support and Vector Compression

• Reddis added native support for JSON data structures, enabling atomic updates and partial retrieval of nested documents
• This also enables indexing JSON fields with the Reddis Query Engine, including vector search capabilities
• Reddis introduced vector compression for homogeneous arrays, reducing memory usage by up to 92% compared to storing vectors in JSON

Reddis 8.4: The Fastest Reddis Ever

• Reddis 8.4 includes significant performance improvements, up to 40% faster in scaling and 90% lower latency for common operations
• It also includes optimizations for TLS connections, delivering 2x the performance of ElastiCache on the same hardware
• These improvements make Reddis 8.4 the fastest Reddis release to date

Real-World Reddis Applications at IFood

• IFood, a major food delivery platform in Latin America, uses Reddis in several critical applications:
  • GenAI: A platform that centralizes access to AI models, using Reddis for rate limiting
  • RAG: A solution that adds context to language models by storing and retrieving relevant content from Reddis
  • Feature Store: Storing low-latency features in Reddis to power recommendations and other ML models

Key Takeaways

• Reddis has evolved to address common challenges like hot keys and secondary indexing, unlocking new use cases
• Customers are moving towards a "refresh ahead" cache pattern, with Reddis Data Integration (RDI) enabling seamless synchronization
• Reddis Flex provides autotiering to enable larger, more cost-effective Reddis deployments
• Reddis 8.4 is the fastest Reddis release yet, with significant performance improvements
• Real-world examples like IFood demonstrate Reddis' ability to power mission-critical applications at scale