HotKV vs Dragonfly

A Dragonfly alternative, measured head to head

Both servers are multi-threaded and speak the Redis protocol, so a Dragonfly™ alternative has to earn the switch on the details. How threads share data, how data is persisted, and what is built into the server. HotKV Ltd measured both on one 160-vCPU machine with the same load, and published every number, including the one that was a tie.

Get early access See the measurements

Dragonfly

a shard per thread

MGET a b c

HotKV

one shared store

MGET a b c
An illustration, not a measurement. Dragonfly gives each thread its own slice of the keyspace, so a command whose keys fall in several slices is coordinated across threads. HotKV runs every thread on one shared store.

This Dragonfly alternative, run as a single instance on one 160-vCPU machine, led 11 of 12 commands against Dragonfly 2.0.2, at 2.0x to 4.4x the throughput. HotKV was built from source for that CPU. Dragonfly 2.0.2 ran as its stock build. The gap is smaller on small machines.

Drop-in

What stays the same with this Dragonfly alternative

HotKV implements the Redis wire protocol. Moving from Dragonfly is usually a change of host and credentials. Check the command reference for anything outside the core set.

Same protocol

Redis-compatible. RESP2 and RESP3 with HELLO negotiation, and HotKV reports Redis version 8.4.0 to clients, so libraries pick their features by it. Connection strings, pipelines, pub/sub and transactions work as they do today.

Same clients and tools

Redis and Valkey client libraries, redis-cli, redis-cli --cluster and GUI tools connect to HotKV without code changes for the core commands.

Same core commands

300 core commands, Lua scripting and function libraries, plus HotKV's own command families when you want them.

the client you already use
$ redis-cli -h <your-hotkv-host> -p 6379 --tls --user app --askpass
> HELLO 3
1# "server" => "hotkv"
2# "version" => "8.4.0"
3# "proto" => (integer) 3
   ... (four more fields)
> SET session:4821 "same protocol, every core"
OK
> GET session:4821
"same protocol, every core"
> TAG.SET session:4821 tenant:acme   # a HotKV command, when you want it
(integer) 1

The case for a Dragonfly alternative

Why teams look for a Dragonfly alternative

Documented properties of Dragonfly that send teams looking for a Dragonfly alternative, checked on October 21, 2026.

01

Shared-nothing shards

Dragonfly splits the keyspace into a shard per thread. A command that touches keys in several shards, such as MSET or MGET, is coordinated across threads. HotKV runs every thread on one shared store. That difference is the heart of this Dragonfly alternative.

02

No append-only file

Dragonfly persists with point-in-time snapshots. Its documentation states that it does not currently support AOF. A Dragonfly alternative with an append-only file closes that gap. HotKV offers snapshots and AOF.

03

Source-available under BSL 1.1

Dragonfly is licensed under the Business Source License 1.1. Its grant allows use inside your own product or service. It does not allow use as an in-memory data store product or service, or offering Dragonfly as a hosted service that competes with DragonflyDB. The code converts to Apache 2.0 on its change date, 1 November 2030 for the current licence text. HotKV is commercial software under its own licence terms.

04

Vector search is built in, AI caches are not

Dragonfly has vector search. A prompt cache, semantic cache, agent memory and LLM cost tracking are not part of the server.

Measured, not claimed

This Dragonfly alternative, measured. HotKV vs Dragonfly 2.0.2 on one machine

HotKV Ltd ran HotKV and Dragonfly 2.0.2 one at a time, as single instances, on the same 160-vCPU machine. Each got the same load and every core. HotKV Ltd published every number with the method and raw data.

11 of 12
commands where HotKV measured the highest throughput against Dragonfly
2.0x to 4.4x
HotKV's throughput as a multiple of Dragonfly's, across the 11 commands it led (MGET to INCR)
PING
Dragonfly measured 0.6% higher, inside the load generators' limit, so it counts as a tie

HotKV Ltd built HotKV from source for this CPU. Dragonfly ran as its stock release build. Both ran as single instances. Thread counts followed each product's documented all-core setting and were not tuned per product. The gap is smaller on machines with few cores.

SET3.4x
HotKV
50.5M*
Dragonfly
15.0M
GET2.7x
HotKV
36.6M*
Dragonfly
13.4M
MGET2.0x
HotKV
4.6M*
Dragonfly
2.3M
SETNX3.8x
HotKV
55.3M*
Dragonfly
14.4M
APPEND2.8x
HotKV
25.1M
Dragonfly
8.9M
INCR4.4x
HotKV
90.8M
Dragonfly
20.6M
LPUSH3.0x
HotKV
31.6M
Dragonfly
10.6M
RPUSH4.0x
HotKV
43.8M
Dragonfly
11.0M
SADD4.2x
HotKV
50.6M*
Dragonfly
12.0M
HSET3.0x
HotKV
29.0M*
Dragonfly
9.8M
MSET3.8x
HotKV
11.3M
Dragonfly
2.9M
PING0.99x
HotKV
125.2M†
Dragonfly
126.0M

Operations per second, median of 3 runs of 10 seconds. Each pair of bars scales to the faster server on that command.

* Network-limited: the instance's network set this result, so the server's own limit is higher.

† Load-generator-limited: the load-generator machines ran above 85% CPU, so this is their ceiling, not the server's.

Azure D160alds_v7 (AMD EPYC 9V45 (Turin), 160 vCPUs), measured October 9, 2026. 256-byte values, pipeline depth 30, 5,000,000 prefilled keys read uniformly at random. Three dedicated load-generator machines. Every server ran alone as a single instance with persistence off. Each got all cores (Redis 128 I/O threads, its maximum). HotKV Ltd built HotKV from source for this CPU and ran Dragonfly as its stock release build. Full method and results · Harness and raw data

Side by side

This Dragonfly alternative and Dragonfly, feature by feature

FeatureHotKVDragonfly
Command executionMulti-threaded. Every core reads, parses and executes commands on one shared storeMulti-threaded, a keyspace shard per thread (shared-nothing)
ProtocolRedis-compatible: RESP2 and RESP3 with HELLO negotiation. Reports Redis version 8.4.0 to clientsRESP2 and RESP3
LanguageC++20, built from scratchC++
LicenceCommercial (HotKV Cloud, or an on-premise licence)BSL 1.1
PersistenceSnapshots and AOFSnapshots only, no AOF
Fork-less snapshotsYes. A background thread writes the snapshot, no forked child processYes
Vector similarity searchBuilt in (HNSW, VSIM.*), every planBuilt in
LLM prompt and semantic cacheBuilt in (PCACHE.*, SCACHE.*), Performance plan and above on HotKV CloudNo
AI agent memoryBuilt in (AMEM.*), Performance plan and above on HotKV CloudNo
LLM token and cost trackingBuilt in (LLMBUDGET.*, LLMSTATS.*), Performance plan and above on HotKV CloudNo
Prometheus endpointBuilt in. The server serves /metrics itself once metrics-port is set (on-premise)Built in
Per-command latency histogramsPer command, built in (LATENCY HISTOGRAM)Opt-in latency tracking flag. LATENCY HISTOGRAM is listed as unsupported
Key taggingYes (TAG.*)No
Nanosecond TTLYes (NEXPIRE, NTTL)No. Millisecond TTLs
Server platformsLinux (x86_64 and ARM64)Linux (x86_64 and ARM64)
DeploymentHotKV Cloud (Google Cloud, 42 regions) or on-premise under a licenceSelf-managed or Dragonfly Cloud

In fairness

Where Dragonfly is ahead of this Dragonfly alternative, or simply different

This Dragonfly alternative page is only useful when it says this part too.

  • On PING, Dragonfly measured 0.6% higher than HotKV. The load generators were above 80% CPU for both, so HotKV Ltd counts it as a tie. It is shown in the table either way.
  • Dragonfly's snapshots are fork-less, like HotKV's, and it serves Prometheus metrics itself.
  • Dragonfly is free to self-host inside your own product under the BSL 1.1, and Dragonfly Cloud is a mature managed service. This Dragonfly alternative, HotKV Cloud, is in early access.

Moving over

From Dragonfly to this Dragonfly alternative in four steps

What switching to a Dragonfly alternative involves in practice.

  1. 1

    Create an instance

    Pick a region and a plan in the HotKV Cloud console. The instance is a machine of its own, provisioned, patched and monitored for you.

  2. 2

    Point your clients at it

    Keep the client library you already use. Change the host, port and credentials from Dragonfly to the HotKV instance. TLS and password authentication are on from the start, with ACL users available. A Dragonfly alternative that speaks the same protocol needs no client changes for the core commands.

  3. 3

    Move the data

    Use the tools you already know. DUMP and RESTORE, MIGRATE, or a replay of your writes all work, because HotKV implements these commands.

  4. 4

    Turn on what was missing

    Vector search and key tagging are core commands on every plan. Prompt and semantic caches, agent memory and LLM cost tracking are commands of the same server on the Performance plan and above. No module, sidecar or second service. That is what a Dragonfly alternative should add.

Questions

Questions about this Dragonfly alternative

Both are multi-threaded. What is different about this Dragonfly alternative?

Dragonfly gives each thread its own shard of the keyspace, so a command whose keys fall in several shards is coordinated across threads. HotKV lets every thread work on one shared store, so multi-key commands and single-key commands take the same path.

How did the two compare in the benchmark?

On an Azure D160alds_v7, run as single instances, HotKV recorded the highest median throughput on 11 of 12 commands against Dragonfly 2.0.2. The multiple ran from at least 2 times on MGET to at least 4.4 times on INCR. PING was a tie within the load generators' limit. Every number, the method and the raw data are published.

Does HotKV have an AOF?

Yes. HotKV persists with snapshots and an append-only file. Dragonfly's documentation states that it does not currently support AOF.

Can I move from Dragonfly without changing my code?

Both servers speak RESP2 and RESP3, so the same client libraries connect to this Dragonfly alternative for the core commands. Move the data with DUMP and RESTORE, MIGRATE, or a replay of your writes. Check the command reference for anything outside the core set.

Put this Dragonfly alternative to the test with the clients you already have

HotKV Cloud is opening in waves. Every instance of this Dragonfly alternative is a dedicated machine with the multi-threaded engine. The AI command families come with the Performance and Scale plans.

Get early access Pricing