Matthew Brown's Blog

The factory method pattern in Go

I have been writing web services in Go for approaching 2 years now. The factory pattern is something I have found particularly useful, especially for writing clean, concise, maintainable and testable code. Originally I used it just for the data access layer so I could swap in and out upon MySQL, PostgreSQL, etc without changing the application layer code but it has proved useful in many other places.

Starting at the very beginning we define our package and import the required libraries.

// Package definition and import the required stdlib packages.
package main

import (
	"fmt"
	"log"
	"errors"
	"strings"
	"database/sql"
	"sync"
)

For the purposes of illustration we will create a DataStore interface that is implemented by PostgreSQLDataStore and MemoryDataStore. The DataStore interface will specify 2 methods: Name and FindUserNameById.

var UserNotFoundError = errors.New("User not found")

type DataStore interface {
	Name() string
	FindUserNameById(id int64) (string, error)
}

Then create 2 structs that implement the the interface.

//The first implementation.
type PostgreSQLDataStore struct {
	DSN string
	DB sql.DB
}

func (pds *PostgreSQLDataStore) Name() string {
	return "PostgreSQLDataStore"
}

func (pds *PostgreSQLDataStore) FindUserNameById(id int64) (string, error) {
	var username string
	err := pds.DB.Query("SELECT username FROM users WHERE id=$1", id).Scan(&username)
	if err != nil {
		if err == sql.ErrNoRows {
			return "", UserNotFoundError
		}
		return "", err
	}
	return username, nil
}

//The second implementation.
type MemoryDataStore struct {
	sync.RWMutex
	Users map[int64]string
}

func (mds *MemoryDataStore) Name() string {
	return "MemoryDataStore"
}

func (mds *MemoryDataStore) FindUserNameById(id int64) (string, error) {
	mds.RWMutex.RLock()
	defer mds.RWMutex.RUnlock()
	username, ok := mds.Users[id];
	if !ok {
		return "", UserNotFoundError
	}
	return username, nil
}

Now we can start to explore the power of the Factory method pattern.

…create objects without having to specify the exact ‘type’ of the object that will be created…

First we must create factory methods for all our implementations that return the common interface. These are essentially constructors that accept a common argument. In our case this common argument is a map[string]string.

type DataStoreFactory func(conf map[string]string) (DataStore, error)

func NewPostgreSQLDataStore(conf map[string]string) (DataStore, error) {
	dsn, ok := conf.Get("DATASTORE_POSTGRES_DSN", "")
	if !ok {
		return nil, errors.New(fmt.Sprintf("%s is required for the postgres datastore", "DATASTORE_POSTGRES_DSN"))
	}

	db, err := sqlx.Connect("postgres", dsn)
	if err != nil {
		log.Panicf("Failed to connect to datastore: %s", err.Error())
		return nil, datastore.FailedToConnect
	}

	return &PostgresDataStore{
		DSN: dsn,
		DB:  db,
	}, nil
}

func NewMemoryDataStore(conf map[string]string) (DataStore, error) {
	 return &MemoryDataStore{
	 	Users: &map[int64]string{
	 		1: "mnbbrown",
	 		0: "root",
	 	},
	 	RWMutex: &sync.RWMutex{},
	 }, nil
}

Now we must store these factory methods somewhere to be called upon as needed. I’ll create a Register helper method to add factories to datastoreFactories. The init function registers both the factories we created above using easy to remember names.

var datastoreFactories = make(map[string]DataStoreFactory)

func Register(name string, factory DataStoreFactory) {
	if factory == nil {
		log.Panicf("Datastore factory %s does not exist.", name)
	}
	_, registered := datastoreFactories[name]
	if registered {
		log.Errorf("Datastore factory %s already registered. Ignoring.", name)
	}
	datastoreFactories[name] = factory
}

func init() {
	Register("postgres", NewPostgreSQLDataStore)
	Register("memory", NewMemoryDataStore)
}

Now the magic happens. Using the Create function below the appropriate factory method will be called using the the conf argument to create an instance of the DataStore interface.

func CreateDatastore(conf map[string]string) (DataStore, error) {

	// Query configuration for datastore defaulting to "memory".
	engineName := conf.Get("DATASTORE", "memory")

	engineFactory, ok := datastoreFactories[engineName]
	if !ok {
		// Factory has not been registered.
		// Make a list of all available datastore factories for logging.
		availableDatastores := make([]string, len(datastoreFactories))
		for k, _ := range datastoreFactories {
			availableDatastores = append(availableDatastores, k)
		}
		return nil, errors.New(fmt.Sprintf("Invalid Datastore name. Must be one of: %s", strings.Join(availableDatastores, ", ")))
	}

	// Run the factory with the configuration.
	return engineFactory(conf)
}

You can use the above CreateDatastore method in your application code as follows:

datastore, err := CreateDataStore(&map[string]string{
	"DATASTORE": "postgres",
	"DATASTORE_POSTGRES_DSN": "dbname=factoriesareamazing",
})

of use another datastore:

datastore, err := CreateDataStore(&map[string]string{
	"DATASTORE": "memory",
})

This interface driven approach allows you to do injection and makes mocking simple for unit tests. i.e you can test against MockDataStore which implements the DataStore interface rather than having to spin up an instances of PostgreSQL for your unit tests. That said, you should still be writing integration tests against a real version of PostgreSQL to account for SQL errors and other quirks you could have missed.

Like what you read? Or think I’m wrong - let me know via @mnbbrown.

Thanks to @smitec and @davidghooper for reading my draft.