Skip to content

Latest commit

 

History

History
141 lines (105 loc) · 6.91 KB

README.md

File metadata and controls

141 lines (105 loc) · 6.91 KB

logo

barreldb

A disk based key-value store based on Bitcask.

Go Reference Go Report Card GitHub Actions


BarrelDB is a Golang implementation of Bitcask by Riak paper and aims to closely follow the spec.

Bitcask is based on a log-structured hash table to store key-value data on disk. It opens a "datafile" (term used for a Bitcask DB file) in an append-only mode and all the writes are sequentially written to this file. Additionally, it also updates an in-memory hash table which maps the key with the offset of the record in the file. This clever yet simple design decision makes it possible to retrieve records from the disk using a single disk seek.

Benefits of this approach

  • Low Latency: Write queries are handled with a single O(1) disk seek. Keys lookup happen in memory using a hash table lookup. This makes it possible to achieve low latency even with a lot of keys/values in the database. Bitcask also relies on the filesystem read-ahead cache for a faster reads.
  • High Throughput: Since the file is opened in "append only" mode, it can handle large volumes of write operations with ease.
  • Predictable performance: The DB has a consistent performance even with growing number of records. This can be seen in benchmarks as well.
  • Crash friendly: Bitcask commits each record to the disk and also generates a "hints" file which makes it easy to recover in case of a crash.
  • Elegant design: Bitcask achieves a lot just by keeping the architecture simple and relying on filesystem primitives for handling complex scenarios (for eg: backup/recovery, cache etc).
  • Ability to handle datasets larger than RAM.

Limitations

  • The main limitation is that all the keys must fit in RAM since they're held inside as an in-memory hash table. A potential workaround for this could be to shard the keys in multiple buckets. Incoming records can be hashed into different buckets based on the key. A shard based approach allows each bucket to have limited RAM usage.

Internals

You can refer to Writing a disk-based key-value store in Golang blog post to read about the internals of Bitcask which also explains how BarrelDB works.

Usage

Library

import (
	"github.com/mr-karan/barreldb"
)

barrel, _ := barrel.Init(barrel.WithDir("data/"))

// Set a key.
barrel.Put("hello", []byte("world"))

// Fetch the key.
v, _ := barrel.Get("hello")

// Delete a key.
barrel.Delete("hello")

// Set with expiry.
barrel.PutEx("hello", []byte("world"), time.Second * 3)

For a complete example, visit examples.

Redis Client

barreldb implements the API over a simple Redis-compatible server (barreldb):

127.0.0.1:6379> set hello world
OK
127.0.0.1:6379> get hello
"world"
127.0.0.1:6379> set goodbye world 10s
OK
127.0.0.1:6379> get goodbye
"world"
127.0.0.1:6379> get goodbye
ERR: invalid key: key is already expired

API

Method Description
Init(...cfg) *Barrel Returns a new instance of Barrel. Additional options can be passed like WithDir, WithReadOnly etc.
Put(string, []byte) error Store a key and value in the datastore.
PutEx(string, []byte, time.Duration) error Store a key and value with expiry in the datastore.
Get(string) []byte,error Retrieve a value by key from the datastore.
Delete(string) error Delete a key from the datastore.
Keys() []string List all keys in the datastore.
Len() int Return the total count of keys in the datastore.
Fold(func(string) error) error Fold over all K/V pairs in a Bitcask datastore. Calls a function on each key inside the datastore.
Sync() error Force any writes to sync to disk.
Shutdown() error Close a data store and flush all pending writes. Removes any lock on the data directory as well.

Benchmarks

Using make bench:

go test -bench=. -benchmem ./...
HELLO
goos: linux
goarch: amd64
pkg: github.com/mr-karan/barreldb
cpu: 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz
BenchmarkPut/DisableSync-8                385432              3712 ns/op        1103.48 MB/s          88 B/op          4 allocs/op
BenchmarkPut/AlwaysSync-8                    222           5510563 ns/op           0.74 MB/s         115 B/op          4 allocs/op
BenchmarkGet-8                            840627              1304 ns/op        3142.20 MB/s        4976 B/op          5 allocs/op
PASS
ok      github.com/mr-karan/barreldb 10.751s

Using redis-benchmark:

$ redis-benchmark -p 6379 -t set -n 10000 -r 100000000
Summary:
  throughput summary: 140845.06 requests per second
  latency summary (msec):
          avg       min       p50       p95       p99       max
        0.196     0.016     0.175     0.255     1.031     2.455

$ redis-benchmark -p 6379 -t set -n 200000 -r 100000000
Summary:
  throughput summary: 143678.17 requests per second
  latency summary (msec):
          avg       min       p50       p95       p99       max
        0.184     0.016     0.183     0.223     0.455     2.183

$ redis-benchmark -p 6379 -t get -n 100000 -r 100000000
Summary:
  throughput summary: 170068.03 requests per second
  latency summary (msec):
          avg       min       p50       p95       p99       max
        0.153     0.040     0.143     0.199     0.367     1.447

References