Redis 8.6 accelerates throughput fivefold

Contents

The in-memory database Redis has been released in version 8.6. The update increases throughput for caching workloads on ARM processors by more than five times compared to Redis 7.2. On a 16-core system in the AWS cloud with a Graviton4 processor, Redis 8.6 reached up to 3.5 million operations per second in benchmarks.

The developers optimized numerous components: Sorted Sets show a latency reduction of up to 35 percent, and Redis performs GET operations on short strings 15 percent faster. Regarding memory consumption, the RAM requirement for hashes decreases by 16.7 percent, and for sorted sets, it even declines by 30.5 percent. These improvements do not only affect ARM systems – similar enhancements are observed on Intel and AMD platforms.

For AI and machine learning applications, Redis 8.6 brings accelerated vector operations: the speed of inserting vectors increases by 43 percent, and queries are performed 58 percent faster. This is particularly true for binary and 8-bit quantization on x86-64 systems.

Idempotent production prevents duplicates in streams

Another key innovation concerns Redis Streams. The XADD command now supports idempotent production through the parameters IDMP and IDMPAUTO. This allows duplicates to be avoided when producers resend messages after network errors or crashes. Each producer receives a unique Producer ID, and each message an Idempotency ID – either manually assigned or automatically generated from the message content.

Administrators can control the retention period of these IDs via XCFGSET STREAM IDMP-DURATION and limit the maximum number per producer with IDMP-MAXSIZE. Idempotent production is particularly important in systems with multiple consumer groups: for example, if a producer crashes after sending an order for food, idempotence prevents the kitchen, warehouse, and delivery service from processing the order twice. However, there is a limitation: the combination of idempotent production and AOF persistence without an RDB preamble currently leads to problems – a patch is expected to follow.

New eviction strategies for write-heavy workloads

With the LRM policies (Least Recently Modified), Redis 8.6 expands the eviction strategies. Unlike the established LRU policies (Least Recently Used), LRM does not consider the last read access but exclusively write operations. This is advantageous for write-intensive applications or AI workloads where read accesses should not influence the relevance of an entry. Available are volatile-lrm for keys with TTL and allkeys-lrm for all keys.

Cluster operators benefit from the new HOTKEYS command. It identifies CPU or network-intensive keys per slot and enables targeted optimizations. In combination with the atomic slot migration introduced in Redis 8.4, hotspots in the cluster can be eliminated precisely. The analysis starts with HOTKEYS START, with metrics, sampling rate, and duration being configurable.

TLS certificates for authentication

For mTLS environments, Redis 8.6 simplifies authentication: clients can automatically log in via their TLS certificate if the common name of the certificate matches an ACL user. Configuration is done with tls-auth-clients-user CN and ACL SETUSER <CN> on nopass. This eliminates the need for separate password authentication – the certificate serves as the sole credential.

Redis 8.6 supports Time-Series databases with NaN (Not a Number) values in TS.ADD and TS.MADD. New aggregators like COUNTNAN and COUNTALL correctly handle missing measurements, while existing aggregators ignore NaN values.

All information about the new release can be found in the project's blog. Redis 8.6 is available as open-source software under the AGPLv3 license. The Unified Distribution includes all features, including RedisTimeSeries, RedisBloom, and RedisJSON, in one package. Binaries are available for Docker, Snap, Homebrew, as well as RPM and Debian packages. The developers have tested the version on Ubuntu 22.04/24.04, Rocky Linux, AlmaLinux, Debian, and macOS.

(fo)