pure-kv

Easy-to-use embedded in-memory key-value storage with ttl and RPC interface.

Features:

doesn't depend on any third-party library;
uses concurrent hash-map inside;
time-to-live could be set to keys;
could be used both as an embedded storage and as a remote RPC server (RPC client provided);
could be dumped to disk and restored back with the full state;

Install

go get github.com/gasparian/pure-kv

Run server

make
./purekv --port 6666 \
         --shards 32 \
         --ttl_garbage_timeout 60000 \

Usage

The only thing you need to know about pure-kv - you can perform just set/get/del with time-to-live and keep byte arrays only. So you can serialize your data with gob, for example, and put the result in the store.

Embedded

You can use store directly in your go code.
Here is an example:

import (
    "log"
    "time"
    "github.com/gasparian/pure-kv/pkg/purekv"
)

// Create new store
store := purekv.NewStore(
    32,    // number of shards for concurrent map
    20000, // TTL garbage collector timeout in ms.
           // setting 0 means turn off keys removal at all
)         
defer store.Close() // Need to finalize background work
// Set key providing ttl of 1 s
err = store.Set("someKey", []byte{'a'}, 1000)
if err != nil {
    log.Fatal(err)
}
// Get current size of store
storeSize := store.Size()
log.Printf("Number of keys in store: %v", storeSize)
// Get value that has been set before
val, err := Get("someKey")
if err != nil {
    log.Fatal(err)
}
log.Printf("Value got from store: %v", val)
// Check that key became stale
time.Sleep(time.Duration(1) * time.Second)
ok = store.Has("someKey")
if ok {
    log.Fatal("Key should be already cleaned")
}
// Setting a key with 0 ttl means that it 
// will live until you delete it manually
store.Set("anotherKey", []byte{'a'}, 0)
store.Del("anotherKey")
// Check that it's not presented
ok = store.Has("anotherKey")
if ok {
    log.Fatal("Key `anotherKey` should not be presented")
}
// Dump store to disk
path := "/path/to/db/dump"
err = store.Dump(path)
if err != nil {
    log.Fatal(err)
}
// Load store back
storeLoaded, err := purekv.Load(path)
defer storeLoaded.Close()
if err != nil {
    log.Fatal(err)
}
log.Printf("Store loaded successfully, size: %v", storeLoaded.Size())
// Drop all entries in store and check it
err = store.DelAll()
if err != nil {
    log.Fatal(err)
}
storeSize = storeLoaded.Size()
if storeSize != 0 {
    log.Fatal("Store should not contain entries after deletion")
}

RPC

Run server:

package main

import (
    pkv "github.com/gasparian/pure-kv/internal/server"
)

func main() {
    flag.Parse()
    srv := pkv.InitServer(
        6666,  // port
        32,    // number of shards for concurrent map
        60000, // TTL garbage collector timeout in ms.
    )
    srv.Start() // <-- blocks
}

Client usage example:

package main

import (
    "log"
    pkv "github.com/gasparian/pure-kv/internal/client"
)

func main() {
    // creates client instance by providing server address and timetout in ms. 
    cli := pkv.New("0.0.0.0:6668", 500)
    err := cli.Open()
    defer cli.Close() 
    if err != nil {
        log.Fatal(err)
    }
    // Creates new key-value pair with ttl of 1 sec.
    err = cli.Set("someKey", []byte{'a'}, 1000) 
    if err != nil {
        log.Fatal(err)
    }
    // Returns set value
    tmpVal, ok := cli.Get("someKey") 
    if !ok {
        log.Fatal("Can't get a value")
    }    
    log.Println(val)
    // Returns number of elements in store
    size := cli.Size() 
    log.Println(size)
    // Async. delete value from the bucket
    err = cli.Del("someKey")
    if err != nil {
        log.Fatal(err)
    }
    // Recreates the entire db, deleting everything
    cli.DelAll() 
}

Tests

You can run tests by package or left argument empty to run all the tests:

make test

Run store benchmark:

make benchmark

The main idea of this benchmark is to see how concurrent access and the number of shards in store affects it's performance.

Benchmark results

I ran benchmark test on my laptop with the following configuration:

goos: darwin
goarch: amd64
cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
gomaxprocs: 12

Below you can find graphs for simulation of 100 serial/concurrent sets/gets.
First case is when ttl has been turned off:

Second case - when we use ttl:

In both cases you can see an improvement when using concurrent access, but the effect of adding more shards in our concurrent map quickly disappears, after ~4-8 shards.
As you can see, the largest improvement belongs to concurrent sets: >30% for experiment without keys ttl and >200% for the experiment with ttl (max. 32 shards).
But at the same time it looks like "gets" are more "stable", and just a single shard (which is effectively just a map with mutex lock) already does the job.
The methodology of an experiment is simple and far from ideal, of course. Possibly we could see more improvement on a larger-scale experiments.

Things to improve

Now, keys with expiry times are stored in a min heap, and this heap is being update with new entries through channels, which are unique for every shard. And on store start-up some amount of goroutines are spawned (number is equal to the number of shards), which may be not the best solution (see the code here). On the other hand - these goroutines by default just waits for the new entry in the channel, which comes from the next Set operation, in case ttl has been set, but it could a problem when there are a lot of sets and these goroutines could take more and more CPU time.