The economics of caching in 2026: Valkey vs Redis and Dragonfly

Infrastructure shifts and licensing

Snap Inc. reduced caching infrastructure costs by 60 percent in the first half of 2026 when it migrated 70 percent of Redis clusters to ElastiCache Valkey, which lowered annual spending from $2.1 million to $840,000. This shift follows the February 2026 end-of-life for Redis 7.2. Because Redis 7.2 no longer receives patches, infrastructure teams must move to Valkey or adopt the tri-licensed Redis 8 series. Redis 8.6 includes idempotent streams production and least recently modified eviction policies. Valkey 9.0 offers atomic slot migration and supports multiple logical databases in cluster mode.

Valkey is cheaper.

The licensing split drives these decisions. In March 2024, Redis Inc. announced that all future Redis OSS releases would ship under the dual SSPL and RSALv2 licenses. The Linux Foundation responded by launching Valkey to provide a community-driven alternative. Valkey uses the BSD 3-Clause license, which allows for unrestricted commercial use. Managed Valkey on AWS ElastiCache runs 20 to 33 percent cheaper than Redis equivalents.

Performance and memory metrics

Valkey 8.1 provides higher throughput and lower memory usage than Redis. Percona Toolkit benchmarks from 2025 show Valkey 8.0 achieving 1.2 million operations per second on r6g.large nodes, compared to 1.11 million for Redis OSS. This represents an 8 percent throughput advantage. Valkey 8.1 also reduces P99 latency by 22 percent compared to Redis OSS. Momento benchmarks from June 2026 show Valkey 8.1.1 reaching 999,800 SET requests per second, which beats the 729,400 SET requests per second achieved by Redis 8.0.

Dragonfly handles heavy workloads with fewer resources. For a 1 million operations per second workload, a 6-node Redis cluster requires 48 vCPUs and 192 GB of memory. Dragonfly performs this task with 16 vCPUs and 48 GB of memory. This configuration results in a monthly AWS cost of $600 compared to $2,500 for the Redis cluster.

Valkey 8.1 reduces memory footprints for typical workloads by 20 percent. In benchmarks testing 50 million sorted set items, Valkey 8.1 used 3.77 GB of memory, while Redis 8.2 used 4.83 GB. This represents a 28 percent memory reduction.

Performance improves with connection volume. Valkey shows 15 to 32 percent performance gains over Redis as connection concurrency increases.

Workload selection

Redis 8 remains the best option for AI and complex workloads. It integrates vector sets, JSON, and time series into the core distribution. Redis 8.4 adds hybrid search using the FT.HYBRID command to combine full-text and vector search. Valkey lacks these integrated features. You should check your codebase for RedisStack commands before you migrate.

Dragonfly is better for performance-critical scaling. It uses a multi-threaded architecture where each thread owns a subset of the keyspace. This design eliminates the need for cluster management for most workloads because a single instance can handle millions of operations per second.

For workloads where automated systems read derived data and act on it, such as fraud detection or AI agent context, Tacnode’s Context Lake provides an alternative. It replaces the cache layer by maintaining incremental materialized views inside a single transactional boundary. This ensures ACID consistency, which means reads reflect the latest committed state rather than a cached snapshot.

Valkey is safer.

Will Valkey eventually bridge the gap in vector search?