Home / Go Cookbook / Generating good unique ids in Go

Imagine you’re writing a note taking application.

Each note needs a unique id.

Generating unique ids is easy if you can coordinate.

The simplest way is to get database to do it: use AUTOINCREMENT column and the database will generate unique id when you insert a new note row into a table.

What if you can’t coordinate?

For example, you want the app to also generate unique note id when offline, when it cannot talk to the database.

The requirement of non-coordinated generation of unique ids comes up often in distributed systems.

A simple solution is to generate a random id. If you give it 16 bytes of randomness, the chances of generating the same random number are non-existent.

It’s such a common problem that over 30 years ago we created a standard for this called UUID/GUID.

We can do better than GUID. A good random unique id:

  • is unique; we can’t skip the basics
  • can be sorted by its string representation
  • is time-clustered i.e. ids generated at the same time are close to each other when sorted
  • string representation can be used as part of URL without escaping
  • the shorter, the better

There are few Go implementation of such id, following the same basic idea:

  • use time as part of the id to achieve time-clustering
  • fill rest of the id with random data
  • encode as a string in a way that allows lexicographic sorting and is url-safe

Here are Go packages for generating unique id and how their ids look like in string format:

package id format
github.com/segmentio/ksuid 0pPKHjWprnVxGH7dEsAoXX2YQvU 4 bytes of time (seconds) + 16 random bytes
github.com/rs/xid b50vl5e54p1000fo3gh0 4 bytes of time (seconds) + 3 byte machine id + 2 byte process id + 3 bytes random
github.com/kjk/betterguid -Kmdih_fs4ZZccpx2Hl1 8 bytes of time (milliseconds) + 9 random bytes
github.com/sony/sonyflake 20f8707d6000108 ~6 bytes of time (10 ms) + 1 byte sequence + 2 bytes machine id
github.com/oklog/ulid 01BJMVNPBBZC3E36FJTGVF0C4S 6 bytes of time (milliseconds) + 8 bytes random
github.com/chilts/sid 1JADkqpWxPx-4qaWY47~FqI 8 bytes of time (ns) + 8 random bytes
github.com/satori/go.uuid 5b52d72c-82b3-4f8e-beb5-437a974842c UUIDv4 from RFC 4112 for comparison

You can see how the values change over time by refreshing this test page a couple of times.

How to generate unique ids using different libraries:

import (
	"github.com/chilts/sid"
	"github.com/kjk/betterguid"
	"github.com/oklog/ulid"
	"github.com/rs/xid"
	"github.com/satori/go.uuid"
	"github.com/segmentio/ksuid"
	"github.com/sony/sonyflake"
)

// To run:
// go run main.go

func genXid() {
	id := xid.New()
	fmt.Printf("github.com/rs/xid:           %s\n", id.String())
}

func genKsuid() {
	id := ksuid.New()
	fmt.Printf("github.com/segmentio/ksuid:  %s\n", id.String())
}

func genBetterGUID() {
	id := betterguid.New()
	fmt.Printf("github.com/kjk/betterguid:   %s\n", id)
}

func genUlid() {
	t := time.Now().UTC()
	entropy := rand.New(rand.NewSource(t.UnixNano()))
	id := ulid.MustNew(ulid.Timestamp(t), entropy)
	fmt.Printf("github.com/oklog/ulid:       %s\n", id.String())
}

func genSonyflake() {
	flake := sonyflake.NewSonyflake(sonyflake.Settings{})
	id, err := flake.NextID()
	if err != nil {
		log.Fatalf("flake.NextID() failed with %s\n", err)
	}
	// Note: this is base16, could shorten by encoding as base62 string
	fmt.Printf("github.com/sony/sonyflake:   %x\n", id)
}

func genSid() {
	id := sid.Id()
	fmt.Printf("github.com/chilts/sid:       %s\n", id)
}

func genUUIDv4() {
	id := uuid.NewV4()
	fmt.Printf("github.com/satori/go.uuid:   %s\n", id)
}

func main() {
	genXid()
	genKsuid()
	genBetterGUID()
	genUlid()
	genSonyflake()
	genSid()
	genUUIDv4()
}

Full example: generate-unique-id/main.go

Which one to use?

All of them are good.

I would pick either rs/xid or segmentio/ksuid.

oklog/ulid allows custom entropy (randomness) source but pays for that with complex API.

sony/sonyflake is the smallest but also the least random. It’s based on Twitter’s design for generating IDs for tweets.

For simplicity the example code serializes sony/snoflake in base16. It would be even shorter in base62 encoding used by other libraries, but other libraries provide that out-of-the-box and for sony/snoflake I would have to implement it myself.

The last one is UUID v4 from RFC 4112, for comparison.

To learn more:

Code for this chapter: https://github.com/kjk/go-cookbook/tree/master/generate-unique-id

Part of Go Cookbook