Versions & roadmap
Valkey release timeline from 7.2 to 9.1, headline features, support windows, and which version to pick.
Release timeline
| Version | Released | Status | Headline features |
|---|---|---|---|
| 7.2 | 2024-04 | LTS (maintained) | First Valkey release, binary-compatible with Redis 7.2.4 |
| 8.0 | 2024-09 | maintained | Multi-threaded I/O re-architected, embedded keys, hash field TTL, around 1.2M QPS single node |
| 8.1 | 2025-03 | LTS | Experimental RDMA transport, scripting Function GA, RESP3 improvements |
| 9.0 | 2025-10 | maintained | valkey-search GA, cluster shard groups, RDB v12, IPv6 cluster bus |
| 9.1 | 2026-05 | LTS-track (current) | RDMA production-ready, per-key memory accounting, hybrid vector + BM25 search |
What each release actually means
7.2 — the safe choice
Identical feature set to Redis 7.2.4, plus security backports. If your operations team is conservative or you're running compliance-critical workloads where surprise changes are expensive, this is your line. Maintained until at least mid-2027.
8.0 — the performance jump
The big one. I/O threads were redone so multiple cores can serve client traffic concurrently. Single-node throughput roughly doubled on hardware with 8+ cores. OBJECT ENCODING changed for many small values to embed keys directly in dictEntry, saving roughly 16 bytes per key. Hash field-level TTL via HEXPIRE / HPEXPIRE / HEXPIREAT arrived.
8.1 — LTS with RDMA experiment
Long-term support line. Most important addition was experimental RDMA transport, which dropped P99 latency from around 200μs to around 50μs on RoCE / InfiniBand hardware. Function (durable Lua replacement) became GA. RESP3 push messages got cleaner.
9.0 — the AI release
valkey-search module went GA — vector indexes with HNSW, hybrid filter + KNN. Cluster shard groups let you organize a 10k-node cluster as N replication groups, reducing gossip overhead about 30 percent. RDB bumped to version 12; old Redis builds can no longer read it (one-way compatibility).
9.1 — current LTS-track
RDMA promoted to production. Per-key memory accounting (OBJECT MEMORY key) for debugging. Hybrid search query language stabilized. Default io-threads-do-reads yes. Modules in the bundle aligned on a common ABI.
Support policy
- LTS lines (7.2, 8.1, and 9.1 going forward) receive security and critical bug fixes for at least 24 months from release.
- Non-LTS lines (8.0, 9.0) receive fixes until the next LTS supersedes them.
- Cadence target is roughly one major every 6-9 months; the TSC discusses any deviations on the community call.
The full support matrix is in github.com/valkey-io/valkey/blob/unstable/RELEASE.md.
Which version to pick
| Scenario | Recommendation |
|---|---|
| New project, want long support | 9.1 (current LTS-track) |
| Conservative ops, want maximum stability | 8.1 (proven LTS) |
| Already on 7.x, no AI features needed | stay on 7.2 until 2027 |
| Migrating from Redis 7.2 | 8.1 (smooth path; RDB still compatible) |
| AI / vector workload | 9.1, modules enabled |
| RDMA-equipped hardware | 9.1, enable tls-port and rdma-port |
Upgrading across the 7.4 boundary (i.e., to 8.x or later) writes a new RDB format. Plan a downgrade path: keep the old RDB until you've verified the new version, because the new RDB cannot be read by 7.x.
Compatibility with Redis
| Boundary | Valkey says | Redis says |
|---|---|---|
| RESP wire protocol | identical | identical |
| Client libraries | unchanged | unchanged |
| Lua scripts | run unchanged | run unchanged |
| Modules ABI | mostly compatible | mostly compatible |
| RDB v11 and earlier | reads & writes | reads & writes |
| RDB v12+ | writes (8.0+) | does not read |
| Cluster shard groups | 9.0+ | not implemented |
| Hash field TTL | 8.0+ | 7.4+ added similar commands |
In practice, application code does not need to change when switching engines. Operations tooling (backups, replication, CDC) may.
Looking ahead
The 9.2 cycle is open at time of writing. Frequently-mentioned candidates: hybrid search ranking improvements, cluster bus over RDMA, in-server multimodal embedding support, and tighter Mem0 / LangGraph integration. None are committed until they merge to unstable and the TSC approves them for the release branch.
For a deeper view of where Valkey is heading, see Frontier and future (CN).