GitHub - unkn0wn-root/kioshun: Fast, sharded Go cache - lock-free reads, self-adaptive SieveTinyLFU, batched async writes and HTTP middleware.

Kioshun - In-Memory Cache for Go

"kee-oh-shoon" /kiːoʊʃuːn/

Fast, sharded in-memory cache for Go

Index

Installation

go get github.com/unkn0wn-root/kioshun

Quick Start

package main

import (
    "fmt"
    "time"

    "github.com/unkn0wn-root/kioshun"
)

func main() {
    // Create cache with default configuration
    c := kioshun.NewDefault[string, string]()
    defer c.Close()

    // Set with default TTL (30 min)
    c.Set("user:123", "David", kioshun.DefaultExpiration)

    // Set commits the write before returning so the key is immediately readable.
    c.Set("user:456", "John", kioshun.NoExpiration)

    // SetAsync returns early. It may commit inline when the shard is idle
    // otherwise it queues the write for that shard's worker.
    c.SetAsync("user:789", "Paul", 5*time.Minute)

    // Optional: call Sync() when committed visibility is required.
    c.Sync()

    // Get value
    if value, found := c.Get("user:123"); found {
        fmt.Printf("User: %s\n", value)
    }

}

Configuration

For most caches, start with the built-in defaults:

c := kioshun.NewDefault[string, string]()
defer c.Close()

NewDefault uses DefaultConfig(): MaxSize=10000, DefaultTTL=30m, automatic shard count, CleanupInterval=5m, SieveTinyLFU eviction, stats disabled and the default async write buffer/batch sizes.

When you only need a few changes, start from DefaultConfig() and override the specific fields:

config := kioshun.DefaultConfig()
config.MaxSize = 100000
config.DefaultTTL = time.Hour

c, err := kioshun.New[string, string](config)
if err != nil {
    // handle error
}

Every Config field, weighted capacity (MaxCost + WithWeigher) and the named cache Manager are covered in configuration.md.

API

c.Set(key, value, ttl time.Duration) error
c.SetAsync(key, value, ttl time.Duration) error
c.SetWithCallback(key, value, ttl, callback func(key, value)) error
c.Get(key) (value, found bool)
c.GetWithTTL(key) (value, ttl time.Duration, found bool)
c.Keys() []K
c.Clear()
c.Sync() error
c.Delete(key) bool
c.Exists(key) bool
c.Size() int64
c.Stats() Stats
c.PolicyStats() PolicyStats
c.Cleanup()
c.Close() error

Set is synchronous and gives immediate 'read-after-write' visibility for the key. SetAsync is optional - it may have committed inline already or it may still be queued. Use Sync when committed visibility is required.

import "github.com/unkn0wn-root/kioshun/httpcache"

config := httpcache.DefaultConfig()
config.DefaultTTL = 5 * time.Minute
config.MaxSize = 100000

middleware, err := httpcache.New(config)
if err != nil {
    // handle error
}
defer middleware.Close()

http.Handle("/api/users", middleware.Wrap(usersHandler))

See http.md for complete documentation.

Benchmarks

Hit ratio and throughput

Kioshun is compared against Ristretto, Otter and Theine by replaying ARC and LIRS request traces (P3, P8, S3, DS1, OLTP, LOOP) through every cache in the same harness at a 100,000 entry cap. Hit ratio is plotted as a percentage of the Belady optimum.

The 100k entry cap.

trace	OPT	kioshun	theine	otter	ristretto
p3	62.00%	44.94%	38.14%	36.41%	21.22%
p8	77.12%	59.85%	54.10%	56.56%	55.83%
s3	25.42%	13.46%	11.90%	9.06%	9.80%
ds1	5.16%	1.97%	1.51%	1.57%	1.36%
oltp	79.56%	79.29%	76.67%	76.41%	28.30%
loop	99.80%	99.26%	99.79%	99.60%	90.39%

Hit ratio as a percent of that ceiling:

trace	kioshun	theine	otter	ristretto
p3	71.3	61.5	58.7	34.2
p8	77.6	70.2	73.3	72.4
s3	49.3	46.8	35.6	38.6
ds1	36.8	29.3	30.4	26.4
oltp	98.8	96.4	96.0	35.6
loop	99.4	100.0	99.8	90.6
geomean	67.9	62.1	59.5	44.8

Microbenchmarks

A separate suite compares Kioshun with Ristretto, BigCache, FreeCache and go-cache using pre-generated workloads with async and strict write modes reported separately:

cd benchmarks
go test -run=TestBenchmarkComparisonGetSetup -count=1 .
go test -bench='BenchmarkCacheComparison' -benchmem -run=^$ -benchtime=1s .
go test -bench=. -benchmem -run=^$ -benchtime=1s -timeout=30m .