From Salvatore Sanfilippo’s 2009 Cache to the Real-Time AI Data Layer
Salvatore Sanfilippo's Redis evolved from a 2009 startup cache into a critical AI data layer, where Redis LangCache delivers 15x faster responses for cache hits and reduces LLM inference costs by 73%.
The origin and the licensing shift
Valkey is the best option for teams wanting to avoid the new Redis licensing risks. Salvatore Sanfilippo launched Redis in 2009. It became the standard for in-memory data infrastructure, handling caching, session management, and rate limiting. In August 2020, Redis Labs closed $100 million in Series F financing at a valuation exceeding $1 billion. TCV and Bain Capital Ventures led this round, with participation from existing investors like Francisco Partners, Goldman Sachs Growth, Viola Ventures, and Dell Technologies Capital. The industry landscape changed in March 2024 when Redis shifted from a BSD license to a dual SSPL/RSALv2 license. This change restricted cloud providers from providing Redis as a managed service. The Linux Foundation launched Valkey, a fully open-source fork of Redis 7.2.4, which received backing from AWS, Google, Oracle, and Ericsson. AWS ElastiCache and MemoryDB already migrated to Valkey to avoid licensing ambiguity. If you use Redis 7.4.x or Redis Community Edition 8.x strictly within your organization, you likely comply with RSALv2. This compliance holds as long as you do not expose the functionality or modified versions to third parties as a service or distribute them to provide functionality to third parties.
Real-time context for AI pipelines
Redis Iris provides the most efficient way to handle the high-latency demands of AI inference. LLM applications manage several latency metrics: first-token time (TTFT), per-token time (TPOT), total response time, and agent run time. A model reads a prompt in one pass to produce the first token, then streams the rest one at a time. In a tested RAG serving pipeline, retrieval accounted for 41% of end-to-end latency. Redis Iris addresses this through Redis Search for vector search and Redis LangCache for semantic caching. Semantic caching stores model responses keyed by the vector embedding of the query. When a new query arrives, the cache compares its embedding to stored ones and returns the cached response if they are similar enough. In benchmarks, Redis LangCache reported responses up to 15x faster for cache hits and 73% lower LLM inference costs. Redis acquired Featureform in October 2025 to integrate its capabilities. Redis manages machine learning workloads through Feature Form. This system provides unified batch and streaming data pipelines. It also provides workspaces so separate machine learning teams can work in a single Redis instance. The system includes fine-grained job control to give teams visibility into data changes. The platform provides atomic directed acyclic graph updates to make change history easier to view.
| Feature | Detail |
|---|---|
| Redis LangCache Speedup | 15x faster for cache hits |
| Redis LangCache Cost | 73% lower LLM inference costs |
| RAG Retrieval Latency | 41% of end-to-end latency |
| Valkey Base Version | Redis 7.2.4 |
| Feature Form capability | Unified batch and streaming pipelines |
The architectural limits and alternatives
Dragonfly works best for engineers who need to break through the single-threaded bottleneck of traditional in-memory stores. The single-threaded architecture of Redis means it can become a bottleneck on modern multi-core servers. When an engineer evaluates whether to replace Redis, they often look at Dragonfly, which uses a shared-nothing architecture to partition the keyspace across CPU cores and delivers up to 25x higher throughput with 25% to 30% lower memory usage. KeyDB, which Snap Inc. has maintained since its 2022 acquisition, provides multi-threading and subkey expires. Garnet, a project from Microsoft Research, achieves sub-millisecond latency at p99.9 using a tiered storage architecture. For those who need to store terabytes of data, Apache Kvrocks uses disk-based storage with RocksDB to reduce infrastructure costs by 5 to 10 times. Momento provides a serverless caching platform where you pay per-GB of data transfer. Pelikan, a framework from Twitter/X, allows users to compose a cache server from protocol, storage, and node modules. Apache Ignite also provides a distributed computing platform that combines an in-memory storage engine and a data grid.
If you need to eliminate the cache layer entirely, Tacnode’s Context Lake maintains incremental materialized views from change data capture. This approach avoids the latency and consistency issues found in traditional caching.
The architectural gap between a cache and a source of truth remains. Caches like Redis prioritize speed over consistency. AI agents that read a cached velocity counter may approve a fraudulent transaction if they read a stale value. This is a shared-state coherence problem that a faster cache cannot solve. Even a cache that is 30x faster only serves stale data 30x faster. Why do engineering teams still rely on caches for irreversible decision-making?
Technewsdaily
Senior tech writer covering AI, gadgets and cybersecurity. Breaking down the news that matters, every day.
