Benchmarks

Measured against Redis, on the same machine, from the records in the repository

Each run starts both servers on one host, drives them with the same load generator in alternating rounds, and stores the raw numbers in a JSON file next to the site. The charts below are drawn from those files when the site is built.

Apple M4 Pro1.25×geometric mean over 60 commands; 57 at 0.97× or faster; memory −26%

AMD EPYC, 2 vCPU1.09×geometric mean over 60 commands; 49 at 0.97× or faster; memory −19%

Two machines

Where each command lands on both machines

Each dot is one command: its ratio to Redis on Apple M4 Pro across, on AMD EPYC, 2 vCPU up. The shaded corner holds the commands that are faster than Redis on both. Dots near the faint diagonal behave the same on ARM and x86.

Reading it

The two machines are very different: a 12-core laptop chip with no pinning, and a 2 vCPU cloud guest with the server and the load generator pinned to separate CPUs. A command that stays top-right on both is fast because of how Ruvio runs it, not because of the hardware.

faster on bothfaster on onebehind on both

Sources: benchmark/records/mac-2026-10-14.json, benchmark/records/linux-hostinger-2026-10-14.json

Distribution

All 60 commands on one axis

The same results as a swarm: one dot per command, stacked where they crowd. Most of the mass sits right of the dashed Redis line on both machines.

Under load

Throughput and tail latency together

Each line joins Redis (hollow) to Ruvio (filled). In the first column further right is more requests per second; in the second, further left is a lower p99. Rows cover client counts, pipelining, fsync on every write, and more Ruvio shards.

RedisRuvio

ScenarioRequests per second → betterp99 latency ← better

No pipelining

SET · 50 clients, no pipeline

1.05×

0.49 ms

GET · 50 clients, no pipeline

1.00×

0.46 ms

SET · 200 clients, no pipeline

1.05×

1.8 ms

GET · 200 clients, no pipeline

1.00×

1.7 ms

Pipeline 16

SET · 200 clients, pipeline 16

1.43×

3.1 ms

HSET · 200 clients, pipeline 16

1.11×

4.1 ms

SET · 200 clients, pipeline 16 · Ruvio 2 shards

1.67×

3.1 ms

SET · 200 clients, pipeline 16 · Ruvio 4 shards

1.66×

1.3 ms

GET · 200 clients, pipeline 16 · Ruvio 2 shards

1.17×

2.7 ms

fsync on every write (always)

SET · 50 clients

2.67×

3.3 ms

SET · 200 clients

1.50×

5.4 ms

SET · 50 clients, pipeline 16

2.33×

5.4 ms

SET · 200 clients, pipeline 16 · Ruvio 2 shards

1.00×

15.2 ms

Source: benchmark/records/x86-azure-2026-10-11.json · AMD EPYC 7763, 8 vCPU (Azure VM) · 2026-10-11, commit 5520164 · server and load generator share the VM, so compare the two servers with each other, not with other hosts. Un-pipelined rows are limited by the load generator.

Whole machine

8 shards against a Redis Cluster of 8 and a single Redis

One redis-benchmark per shard, pinned to its own core, sends each command to that shard's port with the keys inside a hash tag the shard owns, so nothing crosses shards. The same load goes to each node of a Redis Cluster on the same 8 cores, and to one Redis process on one core. Rates are summed over the shards.

In-memory, pipeline 161.13×geometric mean against the cluster over 59 commands; 56 of 60 at 0.97× or faster; per core 1.09× of one Redis

In-memory, pipeline 641.23×geometric mean against the cluster over 59 commands; 56 of 60 at 0.97× or faster; per core 1.17× of one Redis

everysec against Redis AOF everysec1.14×geometric mean against the cluster over 59 commands; 55 of 60 at 0.97× or faster; per core 1.10× of one Redis

Memory0.96×per shard against one Redis over 12 data shapes (0.94× against the cluster's total)

Command

500K1.00M2.00M5.00M10.00M

vs cluster

GET

0.96×

SET

1.09×

DEL (missing key)

0.96×

SETEX

1.40×

HSET (existing field)

1.14×

HGET

1.05×

LPUSH

1.02×

ZADD (existing member, same score)

1.16×

EXPIRE (key with TTL)

1.24×

PUBLISH (no subscribers)

1.36×

Ruvio, 8 shards, one processRedis Cluster, 8 nodes, same coresone Redis process, one core

Command	Ruvio 8 shards	Redis Cluster 8	One Redis	vs cluster	per core vs one Redis
DUMP (64 B string)	6.97M	2.92M	372K	2.39×	2.34×
LINSERT (pivot missing, scans 100)	7.05M	4.67M	593K	1.51×	1.49×
HGET (10-field hash)	4.81M	3.24M	538K	1.48×	1.12×
LREM (value missing, scans 100)	7.04M	4.77M	607K	1.48×	1.45×
COPY ... REPLACE (64 B string)	9.20M	6.31M	877K	1.46×	1.31×
GEOSEARCH (radius 50 km, COUNT 10)	151K	105K	13K	1.44×	1.44×
SETEX	8.89M	6.33M	840K	1.40×	1.32×
HSET (10-field hash, existing field)	4.32M	3.14M	500K	1.37×	1.08×
GEOADD	8.14M	5.96M	772K	1.37×	1.32×
MSET (3 keys)	8.14M	5.97M	787K	1.36×	1.29×
PUBLISH (no subscribers)	2.98M	2.19M	630K	1.36×	0.59×
HDEL (10-field hash, missing field)	4.35M	3.28M	549K	1.33×	0.99×
LMOVE (rotate a 100-element list)	8.54M	6.53M	843K	1.31×	1.27×
LPOS (100 elements, hit in the middle)	6.56M	5.17M	661K	1.27×	1.24×
HGETALL (10-field hash)	3.12M	2.48M	314K	1.26×	1.24×
GETEX (set a 100 s deadline)	8.91M	7.14M	952K	1.25×	1.17×
INCRBYFLOAT	5.72M	4.62M	613K	1.24×	1.17×
EXPIRE (key with TTL)	9.13M	7.38M	977K	1.24×	1.17×
ZADD (existing member, same score)	9.07M	7.82M	1.04M	1.16×	1.09×
BITFIELD INCRBY u8 (always writes)	8.26M	7.20M	958K	1.15×	1.08×
HSET (existing field)	9.13M	8.00M	1.05M	1.14×	1.08×
LINDEX (20-item list)	8.30M	7.47M	568K	1.11×	1.83×
LSET (20-item list)	9.05M	8.16M	1.07M	1.11×	1.06×
MGET (3 keys)	7.74M	7.00M	943K	1.11×	1.03×
BITFIELD GET u8	8.76M	7.99M	1.06M	1.10×	1.03×
SET	8.78M	8.06M	1.07M	1.09×	1.02×
RENAME (same key)	9.19M	8.59M	1.17M	1.07×	0.98×
SETRANGE	9.20M	8.60M	1.14M	1.07×	1.01×
ZSCORE	9.14M	8.56M	1.12M	1.07×	1.02×
SETBIT (1 KiB bitmap, same bit)	9.29M	8.76M	1.16M	1.06×	1.00×
SADD (existing member)	9.32M	8.85M	1.17M	1.05×	0.99×
HSTRLEN	9.19M	8.74M	1.15M	1.05×	1.00×
SRANDMEMBER	9.24M	8.80M	1.17M	1.05×	0.99×
HGET	9.10M	8.68M	1.15M	1.05×	0.99×
HMGET (2 fields)	8.31M	7.95M	1.03M	1.05×	1.00×
PFADD	9.11M	8.72M	1.15M	1.04×	0.99×
GETRANGE	9.07M	8.69M	1.15M	1.04×	0.99×
LSET (100 elements)	7.84M	7.54M	968K	1.04×	1.01×
SISMEMBER	9.19M	8.90M	1.19M	1.03×	0.97×
MEMORY USAGE	9.28M	9.00M	1.19M	1.03×	0.97×
PFCOUNT	9.29M	9.02M	1.17M	1.03×	1.00×
OBJECT ENCODING	9.16M	8.94M	1.20M	1.03×	0.95×
LPUSH	8.28M	8.07M	1.06M	1.02×	0.98×
GETBIT (1 KiB bitmap)	9.29M	9.14M	1.17M	1.02×	1.00×
PERSIST (key without TTL)	9.24M	9.13M	1.18M	1.01×	0.98×
BITCOUNT (1 KiB bitmap)	8.99M	8.89M	1.18M	1.01×	0.95×
APPEND (1 byte)	8.70M	8.71M	1.15M	1.00×	0.95×
ZINCRBY	5.99M	6.00M	783K	1.00×	0.96×
TOUCH	9.34M	9.37M	1.16M	1.00×	1.00×
TTL	9.25M	9.29M	1.21M	0.99×	0.96×
LLEN	9.25M	9.32M	1.18M	0.99×	0.98×
EXISTS	9.25M	9.32M	1.17M	0.99×	0.99×
PTTL	9.20M	9.28M	1.18M	0.99×	0.97×
TYPE	9.03M	9.16M	1.16M	0.99×	0.97×
EXPIRETIME	9.23M	9.39M	1.16M	0.98×	0.99×
LRANGE 0 2	7.89M	8.02M	1.03M	0.98×	0.95×
GET	8.68M	9.02M	1.13M	0.96×	0.96×
DEL (missing key)	8.89M	9.31M	1.18M	0.96×	0.94×
LRANGE 0 9 (20-item list)	5.56M	5.82M	407K	0.96×	1.71×
FILLGET (value present)	8.49M	9.29M	1.15M	0.91×	0.92×

Data shape (8 copies on the 8-shard systems)	Ruvio 8 shards	Redis Cluster 8	One Redis, one copy	per shard vs one Redis
20,000 lists of 20 elements	46.6 MiB	42.3 MiB	5.2 MiB	1.13×
5,000 HyperLogLogs of 100 elements	15.9 MiB	14.3 MiB	2.0 MiB	1.01×
20,000 bitmaps of 128 B (SETBIT)	31.4 MiB	27.4 MiB	3.4 MiB	1.14×
100,000 BITFIELD counters (u16)	27.6 MiB	41.6 MiB	5.1 MiB	0.67×
100,000 geo points in one key	47.8 MiB	47.7 MiB	5.7 MiB	1.04×
100,000 strings of 16 B	45.0 MiB	54.0 MiB	6.7 MiB	0.84×
100,000 strings of 128 B	139.3 MiB	157.6 MiB	19.6 MiB	0.89×
100,000 strings of 16 B with TTL	79.2 MiB	77.6 MiB	9.9 MiB	1.00×
100,000 one-element lists	42.2 MiB	53.4 MiB	6.7 MiB	0.79×
50,000 small hashes (3 fields)	27.1 MiB	28.3 MiB	3.2 MiB	1.07×
50,000 small sets (4 members)	24.6 MiB	25.3 MiB	3.0 MiB	1.01×
50,000 small sorted sets (3 members)	23.7 MiB	24.6 MiB	2.8 MiB	1.07×

Sources: benchmark/records/x86-azure-2026-10-16-8-shards-final.json · benchmark/records/x86-azure-2026-10-16-8-shards-p64-after.json · benchmark/records/x86-azure-2026-10-16-8-shards-everysec-after.json · Azure F16as_v6, AMD EPYC 9V74 (virtualized), 16 cores without SMT; servers on cpus 8-15, load generators on cpus 0-7, loopback · 2026-10-16, Ruvio per-shard forwarding branch at 7376885: lean dispatch in per-shard mode, non-blocking PUBLISH, cycle-counter SLOWLOG timing, six more commands in the WAL group, no record for a no-op SADD · the record with the one-shard control run and the profiles is in the docs.

Every row

Full results

Median requests per second over the rounds, Ruvio and Redis side by side. The whisker spans the worst and best pairing of rounds, so a wide whisker means a noisy row.

Command	Ruvio ops/s	Redis ops/s	Ratio
GEOSEARCH (radius 50 km, COUNT 10)	44K	18K	2.48×
SETEX	2.50M	1.25M	2.00×
MSET (3 keys)	2.20M	1.10M	2.00×
GEOADD	1.94M	1.11M	1.74×
GETEX (set a 100 s deadline)	2.37M	1.38M	1.71×
LINSERT (pivot missing, scans 100)	1.29M	841K	1.53×
COPY ... REPLACE (64 B string)	2.09M	1.38M	1.51×
LMOVE (rotate a 100-element list)	2.05M	1.36M	1.50×
SET	2.81M	1.88M	1.50×
LREM (value missing, scans 100)	1.30M	874K	1.49×
LPOS (100 elements, hit in the middle)	1.58M	1.07M	1.47×
ZADD (existing member, same score)	2.14M	1.53M	1.40×
SETBIT (1 KiB bitmap, same bit)	2.57M	1.84M	1.40×
GETRANGE	2.57M	1.88M	1.37×
BITFIELD GET u8	2.20M	1.61M	1.37×
HSET (existing field)	2.50M	1.84M	1.36×
EXPIRE (key with TTL)	2.31M	1.70M	1.36×
BITFIELD INCRBY u8 (always writes)	1.88M	1.38M	1.35×
HGET	2.57M	1.91M	1.34×
HGET (10-field hash)	2.31M	1.73M	1.33×
LSET (20-item list)	2.20M	1.67M	1.32×
OBJECT ENCODING	2.73M	2.09M	1.30×
HSET (10-field hash, existing field)	1.91M	1.48M	1.30×
HSTRLEN	2.50M	1.96M	1.28×
RENAME (same key)	2.50M	1.96M	1.28×
SETRANGE	2.43M	1.91M	1.27×
ZSCORE	2.25M	1.80M	1.25×
DUMP (64 B string)	1.96M	1.61M	1.22×
SRANDMEMBER	2.43M	2.00M	1.22×
MGET (3 keys)	1.84M	1.53M	1.20×
HMGET (2 fields)	2.00M	1.67M	1.20×
PFADD	2.43M	2.05M	1.19×
LPUSH	2.09M	1.76M	1.19×
HDEL (10-field hash, missing field)	1.96M	1.67M	1.17×
PERSIST (key without TTL)	2.50M	2.14M	1.17×
MEMORY USAGE	2.43M	2.09M	1.16×
APPEND (1 byte)	2.43M	2.09M	1.16×
LSET (100 elements)	1.88M	1.64M	1.15×
PUBLISH (no subscribers)	2.57M	2.25M	1.14×
ZINCRBY	1.58M	1.38M	1.14×
EXISTS	2.50M	2.20M	1.14×
SADD (existing member)	2.37M	2.09M	1.13×
LRANGE 0 2	1.88M	1.67M	1.13×
INCRBYFLOAT	1.34M	1.20M	1.12×
LINDEX (20-item list)	2.14M	1.91M	1.12×
SISMEMBER	2.50M	2.25M	1.11×
TTL	2.37M	2.14M	1.11×
GETBIT (1 KiB bitmap)	2.25M	2.09M	1.08×
EXPIRETIME	2.50M	2.37M	1.06×
PFCOUNT	2.43M	2.31M	1.05×
DEL (missing key)	2.37M	2.25M	1.05×
PTTL	2.25M	2.14M	1.05×
TYPE	2.37M	2.31M	1.03×
LLEN	2.31M	2.31M	1.00×
LRANGE 0 9 (20-item list)	1.18M	1.18M	1.00×
GET	2.14M	2.20M	0.98×
BITCOUNT (1 KiB bitmap)	2.25M	2.31M	0.98×
HGETALL (10-field hash)	763K	789K	0.97×
TOUCH	2.31M	2.50M	0.92×
FILLGET (value present)	1.96M	2.31M	0.85×

Source: benchmark/records/mac-2026-10-14.json · 2026-10-14, commit eb1281c

Command	Ruvio ops/s	Redis ops/s	Ratio
DUMP (64 B string)	726K	338K	2.15×
LINSERT (pivot missing, scans 100)	789K	421K	1.88×
LREM (value missing, scans 100)	789K	474K	1.67×
LPOS (100 elements, hit in the middle)	643K	457K	1.41×
GEOSEARCH (radius 50 km, COUNT 10)	15K	11K	1.40×
GEOADD	789K	577K	1.37×
COPY ... REPLACE (64 B string)	849K	625K	1.36×
EXPIRE (key with TTL)	769K	596K	1.29×
APPEND (1 byte)	588K	471K	1.25×
SET	874K	703K	1.24×
SETEX	841K	682K	1.23×
LMOVE (rotate a 100-element list)	763K	625K	1.22×
ZINCRBY	672K	577K	1.16×
SISMEMBER	818K	703K	1.16×
MSET (3 keys)	726K	625K	1.16×
BITFIELD INCRBY u8 (always writes)	796K	709K	1.12×
HGET (10-field hash)	796K	709K	1.12×
HSET (existing field)	833K	756K	1.10×
RENAME (same key)	818K	744K	1.10×
ZADD (existing member, same score)	857K	783K	1.10×
LPUSH	657K	600K	1.09×
HSTRLEN	818K	756K	1.08×
MGET (3 keys)	692K	643K	1.08×
PERSIST (key without TTL)	738K	687K	1.07×
GETRANGE	811K	756K	1.07×
SETRANGE	865K	811K	1.07×
HGET	720K	677K	1.06×
INCRBYFLOAT	492K	464K	1.06×
HSET (10-field hash, existing field)	709K	672K	1.06×
LRANGE 0 9 (20-item list)	517K	492K	1.05×
HMGET (2 fields)	647K	616K	1.05×
LINDEX (20-item list)	865K	833K	1.04×
SADD (existing member)	714K	692K	1.03×
ZSCORE	818K	796K	1.03×
LSET (20-item list)	811K	789K	1.03×
LLEN	687K	672K	1.02×
GET	841K	826K	1.02×
EXPIRETIME	833K	818K	1.02×
HDEL (10-field hash, missing field)	662K	657K	1.01×
BITCOUNT (1 KiB bitmap)	732K	732K	1.00×
OBJECT ENCODING	738K	738K	1.00×
PTTL	826K	826K	1.00×
SRANDMEMBER	818K	818K	1.00×
LRANGE 0 2	634K	638K	0.99×
LSET (100 elements)	687K	692K	0.99×
TYPE	709K	714K	0.99×
PFCOUNT	826K	841K	0.98×
BITFIELD GET u8	726K	744K	0.98×
SETBIT (1 KiB bitmap, same bit)	826K	849K	0.97×
GETEX (set a 100 s deadline)	634K	657K	0.96×
PUBLISH (no subscribers)	738K	769K	0.96×
MEMORY USAGE	709K	744K	0.95×
PFADD	703K	744K	0.95×
TTL	750K	796K	0.94×
FILLGET (value present)	672K	720K	0.93×
EXISTS	738K	804K	0.92×
DEL (missing key)	756K	826K	0.92×
GETBIT (1 KiB bitmap)	763K	849K	0.90×
TOUCH	709K	796K	0.89×
HGETALL (10-field hash)	398K	457K	0.87×

Source: benchmark/records/linux-hostinger-2026-10-14.json · 2026-10-14, commit eb1281c

Memory

Memory added per data shape

Resident memory after loading minus resident memory at idle, for each server. Left is less, on a log axis.

RuvioRedis

Data shape

2 MiB51020

Saved

100,000 BITFIELD counters (u16)

−51%

100,000 one-element lists

−44%

100,000 strings of 16 B

−42%

50,000 small sets (4 members)

−40%

50,000 small sorted sets (3 members)

−39%

50,000 small hashes (3 fields)

−30%

100,000 geo points in one key

−21%

5,000 HyperLogLogs of 100 elements

−18%

20,000 bitmaps of 128 B (SETBIT)

−16%

100,000 strings of 128 B

−16%

100,000 strings of 16 B with TTL

−12%

20,000 lists of 20 elements

−8%

12 of 12 shapes below Redis · total 67.6 MiB against 91.6 MiB · benchmark/records/mac-2026-10-14.json

Method

How the numbers are made

Apple M4 Pro
Machine: Apple M4 Pro, 12 cores, 24 GiB, macOS 26.6.2 arm64
Redis: 8.10.2
Load: 50 clients, pipeline 16, 3 alternating rounds, median
Ruvio: memory profile, 1 shard, release build, commit eb1281c
Pinning: none

AMD EPYC, 2 vCPU
Machine: AMD EPYC 9354P 32-Core Processor (KVM guest), 2 vCPU, Linux 6.8.0-146-generic x86_64
Redis: 8.10.2
Load: 50 clients, pipeline 16, 3 alternating rounds, median
Ruvio: memory profile, 1 shard, release build, commit eb1281c
Pinning: server cpu 0, client cpu 1

The whole-machine grids come from scripts/shard_compare.py: 8 shards in one Ruvio process (port_mode = "per-shard"), a Redis Cluster of 8 masters created with redis-cli --cluster create, and one Redis process, each driven by one redis-benchmark per shard with hash-tagged keys. The 8-shard systems hold 8 copies of each data shape, so a shard holds what the single process holds.

The single-machine rows come from scripts/redis_compare_features.py, which runs redis-benchmark against each server in turn. Absolute numbers describe one host; the ratio between the two servers is the result. Shared cloud machines vary from run to run, so a row within a few percent of 1× is a tie.

Benchmarking guide All records