Valkey has quietly replaced Redis on Ubuntu, Debian, Homebrew and AWS ElastiCache

The fork became the default

Installing redis-server through apt on Ubuntu 24.10 or Debian 13 no longer delivers Redis. It delivers Valkey, the BSD-licensed fork born of a licensing dispute in March 2024. According to an October 2026 analysis on dev.to, Homebrew has likewise retired its redis formula in favor of valkey, and AWS ElastiCache now builds new clusters on Valkey by default. The fork has quietly become the in-memory store that ships with the platforms most developers start from.

The split was legal, not technical. Through version 7.2.4, Redis shipped under BSD 3-Clause; Redis Ltd. then moved version 7.4 onward to a dual RSALv2/SSPLv1 license, neither OSI-approved, which stopped cloud providers from selling managed Redis without a commercial deal. Days later the Linux Foundation forked the final BSD release as Valkey, backed by AWS, Google Cloud, Oracle, Ericsson and Snap. The dev.to piece also corrects a persistent myth: Redis 8.0 added AGPLv3 as a third option in May 2025, and AGPLv3 is OSI-approved, so calling Redis closed-source has been inaccurate for over a year.

The performance gap lives in pipelines

Valkey's post-fork engineering has targeted throughput. Version 8.0 moved connection handling and write-buffer flushing onto dedicated I/O threads while keeping command execution single-threaded; 8.1 redesigned the hash table and cut roughly 16 bytes of overhead per key; 9.0 added pipeline memory prefetching and automatic SIMD acceleration, AVX512 on x86 and NEON on ARM.

Citing a March 2026 migration guide on sanj.dev, the dev.to author reports that Redis 8.2 and Valkey 8.1 sit within about 8 percent on single-threaded commands, 154,000 versus 142,000 operations per second for SET. Pipelined batches of 10 tell another story: Valkey reaches roughly 812,000 ops/s on SET against 580,000 for Redis, about 40 percent higher, with a similar edge on GET and p99 latency under load falling from 4.2 ms to 2.8 ms. AWS's own figures, also cited in the article, claim up to 40 percent higher pipelined throughput on Valkey 9.0. Redis 8.8 answered with batched prefetching for MGET, MSET and HGETALL, but the throughput momentum sits with the fork. One caution from the article: Valkey's RDMA support is experimental, Linux-only, lacks TLS and is incompatible with replication links, so it is not a reason to migrate.

Snap's 60 percent saving

The headline case study in the dev.to analysis is Snap. At roughly five billion daily requests, its mix of self-managed clusters and commercial Redis licensing cost about 2.1 million dollars a year. After moving to Valkey the bill fell to 840,000 dollars, a 60 percent cut, driven by zero licensing fees, a cluster shrunk from 180 nodes to 162 thanks to the throughput gains, and fewer shard rebalances from lower per-key memory overhead. The article also reports that Valkey instances on ElastiCache cost about 20 percent less than the equivalent Redis instance classes on the same hardware.

Where Redis still wins

Module coverage is the factor many comparisons gloss over. Redis 8 absorbed the former Redis Stack into core: JSON, time series, the search query engine, vector sets and the full RedisBloom suite, including Cuckoo filters, Top-K and T-Digest. Valkey ships separate BSD-licensed equivalents, and coverage varies. valkey- is near parity with JSONPath support, but valkey-search exposes only six commands, omitting autocomplete, synonyms, spellcheck, index aliasing, query profiling, cursors and relevance ranking. Cuckoo filters, Top-K and T-Digest do not exist server-side at all, so no client library can substitute for them, a hard blocker for teams doing frequency estimation or rate limiting on those structures. Valkey holds its own cards in experimental RDMA and a roadmap for active-active CRDT replication that no Redis version offers.

The license question, stated plainly

For most teams the license question in 2026 is internal policy, not cloud-provider politics. Valkey's BSD 3-Clause carries no deployment restrictions. Redis forces a choice among three licenses: RSALv2, permissive but banning managed-service resale; SSPLv1, copyleft with broad service-side obligations; and AGPLv3, strong copyleft. Because many companies enforce blanket bans on AGPL dependencies, the dev.to author argues, Redis 8 is ruled out on paper at a large share of enterprises before any technical evaluation happens. Pinning to Redis 7.2.4, the last BSD release, is no safe harbor either, since community security backports to that line are not guaranteed.

Why it matters

Defaults shape infrastructure choices more than benchmarks do. With Ubuntu, Debian, Homebrew and ElastiCache steering new deployments toward Valkey, the fork has effectively inherited the Redis install base, while Redis keeps a defensible niche wherever probabilistic structures or full search are load-bearing. Single-command workloads see roughly 8 percent between the engines; pipelined, latency-sensitive paths see 40 percent, which decides whether you buy more nodes. The author's figures come from public sources, sanj.dev benchmarks, AWS reporting and Snap's disclosed costs, rather than a first-hand migration, so teams should validate on their own workloads. Either way, the question has moved on from whether Redis is open source to which engine fits your modules, your legal policy and your cloud bill.