pboyd.io

An experiment on code structure - Part 2

This is a follow-up to an experiment on code structure. To recap, I built two versions of a back-end for a simple web app with statistics about commercial airline flights. backendA had no real design behind it, it was just whatever fell out of keyboard. backendB was a structure that I’ve been using for the past couple years, it isolates dependencies, and uses interfaces with dependency injection. The two versions produce identical output, only the structure of the code is different.

I goofed up with my experiment and never chose a criteria to judge it by. Needless to say, it ended inconclusively and all I could do was stick with my status quo.

Fortunately, a reader sent me a recommendation to read “A philosophy of software design” by John Ousterhout. The book argues that good software design produces so called “deep modules”, that is powerful implementations hidden behind simple interfaces. Complexity is therefore the enemy, but complexity gets a special definition: “anything related to the structure that makes the code hard to understand or modify”. So not reducing complexity in general, just the complexity created by the structure of the code.1

It may not sound like it at first, but “modules should be deep” contradicts a lot of common advice about software design (a module, by the way, is any division of code–function, class, package, etc.). Such as, making functions small. Ousterhout actually encourages large functions. Short functions may be simple by themselves, but lots of them together create a complicated system.

I’m not sure that that simplicity is better than every other software design virtue. But I certainly agree that simplicity is better than complexity, and it at least gives me a reasonable way to judge my experiment. The winning design is the simpler one.

It isn’t even a close contest by this standard. Take the resolver function for the airport GraphQL query. It handles queries like this: {airport(code:"JFK"){name,city,state}} which looks up an airport by code.

Here’s the version from backendA:

func resolveAirportQuery(db *sql.DB) graphql.FieldResolveFn {
	return graphQLMetrics("airport",
		func(p graphql.ResolveParams) (interface{}, error) {
			code := getAirportCodeParam(p, "code")
			if code == "" {
				return nil, nil
			}

			row := db.QueryRow(`
				SELECT
					code, name, city, state, lat, lng
				FROM
					airports
				WHERE
					is_active=1 AND
					code=?
			`, code)

			var a airport
			err := row.Scan(&a.Code, &a.Name, &a.City, &a.State, &a.Latitude, &a.Longitude)
			if err != nil {
				if err == sql.ErrNoRows {
					return nil, nil
				}
				return nil, err
			}

			return &a, nil
		},
	)
}

It’s easy to see how this works. It isn’t extremely short, it operates on more than one level of abstraction, and it is concerned with things unrelated to it’s objective (e.g. the graphQLMetrics wrapper). But there are no mysteries here. If you need to add another field or make some other small change it won’t get in your way.

Now compare that with the equivalent in backendB:

func (p *Processor) resolveAirportQuery(params graphql.ResolveParams) (interface{}, error) {
	code := p.getAirportCodeParam(params, "code")
	if code == "" {
		return nil, nil
	}

	return p.config.AirportStore.Airport(params.Context, code)
}

It’s much shorter. If you’re content with the abstraction provided by AirportStore.Airport then this is everything you need. If not, you’ll need to dig deeper. If you know this codebase pretty well you can probably jump right to the relevant code. Otherwise, you’ll have to figure out what p.config is:

type Processor struct {
	config ProcessorConfig
	schema graphql.Schema
}

It’s a ProcessorConfig:

type ProcessorConfig struct {
	AirportStore     app.AirportStore
	FlightStatsStore app.FlightStatsStore
}

I’ll spare you the details and just give a summary:

  • p.config.AirportStore is an app.AirportStore, so go look in the app package
  • app.AirportStore is an interface with an Airport method
  • AirportStore isn’t implemented in app, the programmer has to go back and see how the Processor gets a ProcessorConfig
  • The ProcessorConfig is passed in when the Processor is created. It’s up to the caller to supply a valid ProcessorConfig with a valid AirportStore instance.
  • Hopefully the programmer thinks to check main, where they’ll see that AirportStore is a *mysql.Store

Now, finally, in the mysql package we’ll find:

func (s *Store) Airport(ctx context.Context, code string) (*app.Airport, error) {
	code = strings.ToUpper(code)

	row := s.db.QueryRow(`
		SELECT
			code, name, city, state, lat, lng
		FROM
			airports
		WHERE
			is_active=1 AND
			code=?
	`, code)

	var a app.Airport
	err := row.Scan(&a.Code, &a.Name, &a.City, &a.State, &a.Latitude, &a.Longitude)
	if err != nil {
		if err == sql.ErrNoRows {
			return nil, nil
		}
		return nil, err
	}

	return &a, nil
}

That doesn’t seem simple. backendA is the simpler design.

backendB is more flexible though. Everything is isolated and used through an interface, so I can just replace those interface implementations. I could support an additional database MySQL, or add a caching layer just by writing a new Store implementation. If I actually needed to do any of that, backendB might have had a good design, but I didn’t. And that flexibility caused complexity and that’s why backendA is simpler.

I made another version of the flightranker backend, backendC, that followed Ousterhout’s book. I settled on three packages:

  • store - high-level application functions that pull data from MySQL
  • server - HTTP server handling GraphQL queries
  • main - just an entry point

server is responsible for everything related to HTTP. This is a break from backendB where main started the server, http made a handler and graphql had most of the functionality.

All the code in backendB that reads environment variables is in the main package. That made main aware of things it didn’t really need (e.g. MySQL connection information). With backendC I decided to read the environment variables from deeper down, so a module is responsible for its’ whole problem.

backendB’s main package also used dependency injection to put the right implementation of an interface in the right place. But, again, this was needless complexity since there there were single implementations. For backendC, there’s only one store implementation, and since it can read it’s own config, it’s simple enough for server to create an instance directly.

Between pushing the configuration down and removing dependency injection, backendC’s main function is really short:

func main() {
	server.Run()
}

That’s the equivalent of ~30 lines of code in backendB.

Earlier we looked at the airport resolver in backendA and backendB. Here’s the equivalent for backendC:

// airportQuery defines a GraphQL query that accepts an airport code and
// responds with information about the airport.
func airportQuery(st *store.Store) *graphql.Field {
	return &graphql.Field{
		Type:        airportType,
		Description: "get airport by code",
		Args: graphql.FieldConfigArgument{
			"code": airportCodeArgument,
		},
		Resolve: func(params graphql.ResolveParams) (interface{}, error) {
			code, _ := params.Args["code"].(string)
			airport, err := st.Airport(params.Context, code)

			if err == store.ErrInvalidAirportCode {
				return nil, nil
			}

			return airport, err
		},
	}
}

I decided to in-line the Resolve functions in the graphql.Field definitions. I think they’re easier to understand together.

This isn’t as complete as backendA, but the store.Airport abstraction is good enough for some purposes. And if it’s not good enough, you can at least see right away that store.Airport comes from store.Store.

backendC is simpler than backendB, that’s pretty obvious. It’s not as clear of a comparison between backendA and backendC, but I think backendC is simpler than backendA too. There’s a limit to what a person can keep in their head at one time. backendA doesn’t have many abstractions and it can be hard to follow the logic at times.

For instance, the airport resolver in backendA was simple enough to not be a problem, but the more complicated resolvers are a challenge to keep straight. Like the all-time airline stats resolver in backendA. It has very few abstractions, and those that are there (e.g. isAirportCode) are “shallow”, and don’t remove many details from the programmer.

Now look at the equivalent in backendC. It’s not perfect, but the store.FlightStats call does a big portion of the job, and let’s the programmer get that much of the problem out of their head.

The main thing I learned through all this is that I actually have to design software. I was relying too much on a “standard” design which usually didn’t fit. A standard structure has to be flexible, but every bit of flexibility costs a bit of simplicity.


  1. If the special definitions bug you, just remember the title of the book contains the word “philosophy”. [return]