Permalink
Branch: master
Find file Copy path
788 lines (582 sloc) 23.7 KB

2. Language Rules

2.1. MUST NOT use "return"

The return statement from Java signals a side-effect - unwind the stack and give this value to the caller. In a language in which the emphasis is on side-effect-full programming, this makes sense. However Scala is an expression oriented language in which the emphasis is on controlling/limiting side-effects and return is not idiomatic.

To make matters worse, return probably doesn't behave as you think it does. For example in a Play controller, try doing this:

def action = Action { request =>
  if (someInvalidationOf(request))
    return BadRequest("bad")

 Ok("all ok")
}

In Scala, a return statement inside a nested anonymous function is implemented by throwing and catching a NonLocalReturnException. It says so in the Scala Language Specification, section 6.20.

Besides, return is anti structural programming, as functions can be described with multiple exit points and if you need return, like in those gigantic methods with lots of if/else branches, the presence of a return is a clear signal that the code stinks, a magnet for future bugs and is thus in urgent need of refactoring.

2.2. SHOULD use immutable data-structures

Immutable data-structures are facts that can be compared and reasoned about. Mutable things are error-prone buckets. You should never use a mutable data-structure unless you're able to defend it and there are really, really few places in which a mutable data-structure is defensible.

Lets exemplify:

trait Producer[T] {
 def fetchList: List[T]
}

// consumer side
someProducer.fetchList

Question - if the List returned above is mutable, what does that say about the Producer interface?

Here are some problems:

  1. if this List is produced on another thread than the consumer, one can have both visibility and atomicity problems - you can't know whether that will happen, unless you take a look at the Producer's implementation.

  2. even if this List is effectively immutable (i.e. still mutable, but no longer modified after being signaled to the Consumer), you don't know if it will be signaled to other Consumers that may modify it by themselves, so you can't reason about what you can do with it.

  3. even if it is described that access to this List must be synchronized, problem is - on which lock are you going to synchronize? Are you sure you'll get the locking order right? Locks are not composable.

So there you have it - a public API exposing a mutable data-structure is an abomination of nature, leading to problems that can be worse than what happens when doing manual memory management.

2.3. SHOULD NOT update a var using loops or conditions

It's a mistake that most Java developers do when they come to Scala. Example:

var sum = 0
for (elem <- elements) {
  sum += elem.value
}

Avoid doing this, prefer the available operators instead, like foldLeft:

val sum = elements.foldLeft(0)((acc, e) => acc + e.value)

Or even better, know the standard library and always prefer to use the built-in functions - the more expressive you go, the less bugs you'll have:

val sum = elements.map(_.value).sum

In the same spirit, you shouldn't update a partial result with a condition. Example:

def compute(x) = {
  var result = resultFrom(x)

  if(needToAddTwo) {
    result += 2
  }
  else {
  	result += 1
  }

  result
}

Prefer expressions and immutability. The code will be more readable and less error-prone, since it makes the branches more explicit and that's a good thing:

def computeResult(x) = {
  val r = resultFrom(x)
  if (needToAddTwo)
    r + 2
  else
  	r + 1
}

And you know, as soon as the branches get too complex, just as was said in the discussion on return, that's a sign that the code smells and is in need of refactoring, which is a good thing.

2.4. SHOULD NOT define useless traits

There was this Java Best Practice that said "program to an interface, not to an implementation", a best practice that has been cargo-cult-ed to the point that people started defining completely useless interfaces in their code. Generally speaking, that rule is healthy, but it refers to the general engineering need of hiding implementation details especially details of modifying state (encapsulation) and not to slap interface declarations that often leak implementation details anyway.

Defining traits is also a burden for readers of that code, because it signals a need for polymorphism. Example:

trait PersonLike {
  def name: String
  def age: Int
}

case class Person(name: String, age: Int)
  extends PersonLike

Readers of this code might come to the conclusion that there are instances in which overriding PersonLike is desirable. That couldn't be further from the truth - Person is perfectly described by its case class as a data-structure without behavior. In other words it describes the shape of your data and if you need to override this shape for some unknown reason, then this trait is badly defined because it imposes the shape of your data and that's about the only thing you can override. You can always come up with traits later, if you're in need of polymorphism, after your needs evolve.

And if you're thinking that you may need to override the source of this (as in to fetch the person's name from the DB on first access), OMG don't do that!

Note that I'm not talking about algebraic data structures (i.e. sealed traits that are signaling a closed set of choices - like Option).

Even in those cases in which you think the issue is clear-cut, it may not be. Lets take this example:

trait DBService {
  def getAssets: Future[Seq[(AssetConfig, AssetPersistedState)]]

  def persistFlexValue(flex: FlexValue): Future[Unit]
}

This snippet is taken from real-world code - we've got a DBService that handles either queries or persistence in a database. Those two methods are actually unrelated, so if you only need to fetch the assets, why depend on things you don't need in components that require DB interaction?

Lately my code looks a lot like this:

final class AssetsObservable
    (f: => Future[Seq[(AssetConfig, AssetPersistedState)]])
  extends Observable[AssetConfigEvent] {

  // ...
}

object AssetsObservable {
  // constructor
  def apply(db: DBService) = new AssetsObservable(db.getAssets)
}

See, I do not need to mock an entire DBService in order to test the above.

2.5. MUST NOT use "var" inside a case class

Case classes are syntactic sugar for defining classes in which - all constructor arguments are public and immutable and thus part of the value's identity, have structural equality, a corresponding hashCode implementation and apply/unapply auto-generated functions provided by the compiler.

By doing this:

case class Sample(str: String, var number: Int)

You just broke its equality and hashCode operation. Now try using it as a key in a map.

As a general rule of thumb, structural equality only works for immutable things, because the equality operation must be stable (and not change according to the object's history). Case classes are for strictly immutable things. If you need to mutate stuff, don't use case classes.

In the approximate words of Fogus in "The Joy of Clojure" or Baker in his paper from 1993: if any two mutable objects resolve as being equal now, then there’s no guarantee that they will a moment from now. And if two objects aren’t equal forever, then they’re technically never equal ;-)

2.6. SHOULD NOT declare abstract "var" members

It's a bad practice to declare abstract vars in abstract classes or traits. Do not do this:

trait Foo {
  var value: String
}

Instead, prefer to declare abstract things as def:

trait Foo {
  def value: String
}

// can then be overridden as anything
class Bar(val value: String) extends Foo

The reason has to do with the imposed restriction - a var can only be overridden with a var. The way to allow freedom to choose on inheritance is to use def for abstract members. And why would you impose the restriction to use a var on those that inherit from your interface. def is generic so use it instead.

2.7. MUST NOT throw exceptions for validations of user input or flow control

Two reasons:

  1. it goes against the principles of structured programming as a routine ends up having multiple exit points and are thus harder to reason about - with the stack unwinding happening being an awful and often unpredictable side-effect
  2. exceptions aren't documented in the function's signature - Java tried fixing this with the checked exceptions concept, which in practice was awful as people simply ignored them

Exceptions are useful for only one thing - signaling unexpected errors (bugs) up the stack, such that a supervisor can catch those errors and decide to do things, like log the errors, send notifications, restarting the guilty component, etc...

As an appeal to authority, it's reasonable to reference Functional Programming with Scala, chapter 4.

2.8. MUST NOT catch Throwable when catching Exceptions

Never, never, never do this:

try {
 something()
} catch {
 case ex: Throwable =>
   blaBla()
}

Never catch Throwable because we could be talking about extremely fatal exceptions that should never be caught and that should crash the process. For example if the JVM throws an out of memory error, even if you re-throw that exception in that catch clause, it may be too late - given that the process is out of memory, the garbage collector probably took over and froze everything, with the process ending in a zombie unrecoverable state. Which means that an external supervisor (like Upstart) will not get an opportunity to restart it.

Instead do this:

import scala.util.control.NonFatal

try {
 something()
} catch {
 case NonFatal(ex) =>
   blaBla()
}

2.9. MUST NOT use "null"

You must avoid using null. Prefer Scala's Option[T] instead. Null values are error prone, because the compiler cannot protect you. Nullable values that happen in function definitions are not documented in those definitions. So avoid doing this:

def hello(name: String) =
  if (name != null)
    println(s"Hello, $name")
  else
    println("Hello, anonymous")

As a first step, you could be doing this:

def hello(name: Option[String]) = {
  val n = name.getOrElse("anonymous")
  println(s"Hello, $n")
}

The point of using Option[T] is that the compiler forces you to deal with it, one way or another:

  1. you either have to deal with it right away (e.g. by providing a default, throwing an exception, etc..)
  2. or you can propagate the resulting Option up the call stack

Also remember that Option is just like a collection of 0 or 1 elements, so you can use foreach, which is totally idiomatic:

val name: Option[String] = ???

for (n <- name) {
  // executes only when the name is defined
  println(n)
}

Combining multiple options is also easy:

val name: Option[String] = ???
val age: Option[Int] = ???

for (n <- name; a <- age)
  println(s"Name: $n, age: $a")

And since Option is seen as an Iterable too, you can use flatMap on collections to get rid of None values:

val list = Seq(1,2,3,4,5,6)

list.flatMap(x => Some(x).filter(_ % 2 == 0))
// => 2,4,6

2.10. MUST NOT use Option.get

You might be tempted to do this:

val someValue: Option[Double] = ???

// ....
val result = someValue.get + 1

Don't ever do that, since you're trading a NullPointerException for a NoSuchElementException and that defeats the purpose of using Option in the first place.

Alternatives:

  1. using Option.getOrElse
  2. using Option.fold
  3. using pattern matching and dealing with the None branch explicitly
  4. not taking the value out of its optional context

As an example for (4), not taking the value out of its context means this:

val result = someValue.map(_ + 1)

2.11. MUST NOT use Java's Date or Calendar, instead use java.time (JSR-310)

Java's Date and Calendar classes from the standard library are awful because:

  1. resulting objects are mutable, which doesn't make sense for expressing a date, which should be a value (how would you feel if you had to work with StringBuffer everywhere you have Strings?)
  2. months numbering is zero based
  3. Date in particular does not keep timezone info, so Date values are completely useless
  4. it doesn't make a difference between GMT and UTC
  5. years are expressed as 2 digits instead of 4

Instead, always use the java.time API introduced in Java 8 - or if you're stuck in pre-Java 8 land, use Joda-Time, which is its spiritual ancestor.

2.12. SHOULD NOT use Any or AnyRef or isInstanceOf / asInstanceOf

Avoid using Any or AnyRef or explicit casting, unless you've got a really good reason for it. Scala is a language that derives value from its expressive type system, usage of Any or of typecasting represents a hole in this expressive type system and the compiler doesn't know how to help you there. In general, something like this is bad:

val json: Any = ???

if (json.isInstanceOf[String])
  doSomethingWithString(json.asInstanceOf[String])
else if (json.isInstanceOf[Map])
  doSomethingWithMap(json.asInstanceOf[Map])
else
  ???

Often we are using Any when doing deserialization. Instead of working with Any, think about the generic type you want and the set of sub-types you need, and come up with an Algebraic Data-Type:

sealed trait JsValue

case class JsNumber(v: Double) extends JsValue
case class JsBool(v: Boolean) extends JsValue
case class JsString(v: String) extends JsValue
case class JsObject(map: Map[String, JsValue]) extends JsValue
case class JsArray(list: Seq[JsValue]) extends JsValue
case object JsNull extends JsValue

Now, instead of operating on Any, we can do pattern matching on JsValue and the compiler can help us here on missing branches, since the choice is finite. This will trigger a warning on missing branches:

val json: JsValue = ???
json match {
  case JsString(v) => doSomethingWithString(v)
  case JsNumber(v) => doSomethingWithNumber(v)
  // ...
}

2.13. MUST serialize dates as either Unix timestamp, or as ISO 8601

Unix timestamps, provided that we are talking about the number of seconds or milliseconds since 1970-01-01 00:00:00 UTC (with emphasis on UTC) are a decent cross-platform serialization format. It does have the disadvantage that it has limits in what it can express. ISO-8601 is a decent serialization format supported by most libraries.

Avoid anything else and also when storing dates without a timezone attached (like in MySQL), always express that info in UTC.

2.14. MUST NOT use magic values

Although not uncommon in other languages to use "magic" (special) values like -1 to signal particular outcomes, in Scala there are a range of types to make intent clear. Option, Either, Try are examples. Also, in case you want to express more than a boolean success or failure, you can always come up with an algebraic data type.

Don't do this:

val index = list.find(someTest).getOrElse(-1)

2.15. SHOULD NOT use "var" as shared state

Avoid using "var" at least when speaking about shared mutable state. Because if you do have shared state expressed as vars, you'd better synchronize it and it gets ugly fast. Much better is to avoid it. In case you really need mutable shared state, use an atomic reference and store immutable things in it. Also checkout Scala-STM.

So instead of something like this:

class Something {
  private var cache = Map.empty[String, String]
}

If you can't really avoid that variable, prefer doing this:

import java.util.concurrent.atomic._

class Something {
  private val cache =
    new AtomicReference(Map.empty[String, String])
}

Yes, it introduces overhead due to the synchronization required, which in the case of an atomic reference means spin loops. But it will save you from lots and lots of headaches later. And it's best to avoid mutation entirely.

2.16. Public functions SHOULD have an explicit return type

Prefer this:

def someFunction(param1: T1, param2: T2): Result = {
  ???
}

To this:

def someFunction(param1: T1, param2: T2) = {
  ???
}

Yeah, type inference on the result of a function is great and all, but for public methods:

  1. it's not OK to rely on an IDE or to inspect the implementation in order to see the returned type
  2. Scala currently infers the most specialized type possible, because in Scala the return type on functions is covariant, so you might actually get a really ugly type back

For example, what is the returned type of this function:

def sayHelloRunnable(name: String) = new Runnable {
  def sayIt() = println(s"Hello, $name")
  def run() = sayIt()
}

Do you think it's Runnable? Wrong, it's Runnable{def sayIt(): Unit}.

As a side-effect, this also increases compilation times, as whenever sayHelloRunnable changes implementation, it also changes the signature so everything that depends on it must be recompiled.

2.17. SHOULD NOT define case classes nested in other classes

It is tempting, but you should almost never define nested case classes inside another object/class because it messes with Java's serialization. The reason is that when you serialize a case class it closes over the "this" pointer and serializes the whole object, which if you are putting in your App object means for every instance of a case class you serialize the whole world.

And the thing with case classes specifically is that:

  1. one expects a case class to be immutable (a value, a fact) and hence
  2. one expects a case class to be easily serializable

Prefer flat hierachies.

2.18 MUST NOT include classes, traits and objects inside package objects

Classes, including case classes, traits and objects do not belong inside package objects. It is unnecessary, confuses the compiler and is therefore discouraged. For example, refrain from doing the following:

package foo

package object bar {
  case object FooBar
}

The same effect is achieved if all artifacts are inside a plain package:

package foo.bar

case object FooBar

Package objects should only contain value, method and type alias definitions, etc. Scala allows multiple public classes in a single file, and the convention is to have the first letter of the filename be lowercase in such cases.

Implicit value classes can be defined in a package object

In one rare circumstance, it makes sense to include classes defined directly in a package object. The reason for that is that implicit classes need to be nested inside another object/class, you cannot define a top level implicit in Scala. Nesting implicit value classes inside a package object also allows us to create a great importing experience for a library, as the single import of the package object will bring in all the necessary implicits.

There is also no way around this, as defining the implicit value class means we cannot effectively define the implicit and the class separately, the whole point is to let the compiler avoid the runtime boxing by generating all the "right code" at compile time, by predicting the runtime boxing. This is the optimal way to achieve a pimp my library pattern performance wise, and we need the entire syntax "in one place".

package object dsl {
  implicit class DateTimeAugmenter(val date: Datetime) extends AnyVal {
    def yesterday: DateTime = date.plusDays(-1)
  }
}

2.19 SHOULD use head/tail and init/last decomposition only if they can be done in constant time and memory

Example of head/tail decomposition:

def recursiveSumList(numbers: List[Int], accumulator: Int): Int =
  numbers match {
    case Nil =>
      accumulator

    case head :: tail =>
      recursiveSumList(tail, accumulator + head)
  }

List has a special head/tail extractor :: because Lists are made by always appending an element to the front of the list:

val numbers = 1 :: 2 :: 3 :: Nil

This is the same as:

val numbers = Nil.::(3).::(2).::(1)

For this reason, both head and tail on a list need only constant time and memory! These operations are O(1).

There is another head/tail extractor called +: that works on any Seq:

def recursiveSumSeq(numbers: Seq[Int], accumulator: Int): Int =
  numbers match {
    case Nil =>
      accumulator

    case head +: tail =>
      recursiveSumSeq(tail, accumulator + head)
  }

You can find the implementation of +: here. The problem is that other collections than List do not necessarily head/tail-decompose in constant time and memory, e.g. an Array:

val numbers = Array.range(0, 10000000)

recursiveSumSeq(numbers, 0)

This is highly inefficient: each tail on an Array takes O(n) time and memory, because every time a new array needs to be created!

Unfortunately, the Scala collections library permits these kinds of inefficient operations. We have to keep an eye out for them.


An example for an efficient init/last decomposition is scala.collection.immutable.Queue. It is backed by two Lists and the efficiency of head, tail, init and last is amortized constant time and memory, as explained in the Scala collection performance characteristics.

I don't think that init/last decomposition is all that common. In general, it is analogue to head/tail decomposition. The init/last deconstructor for any Seq is :+.

2.20 MUST NOT use Seq.head

You might be tempted to this:

val userList: List[User] = ???

// ....
val firstName = userList.head.firstName

Don't ever do this, as this will throw NoSuchElementException if the sequence is empty.

Alternatives:

  1. using Seq.headOption possibly combined with getOrElse or pattern matching

    Example:

    val firstName = userList.headOption match {
        case Some(user) => user.firstName
        case _ => "Unknown"
      }
  2. using pattern matching with the cons operator :: if you're dealing with a List

    Example:

    val firstName = userList match {
     case head :: _ => head.firstName
     case _ => "Unknown"
    }
  3. using NonEmptyList if it is required that the list should never be empty. (See cats, scalaz, ...)

2.21 Case classes SHOULD be final

Extending a case class will lead to unexpected behaviour. Observe the following:

scala> case class Foo(v:Int)
defined class Foo

scala> class Bar(v: Int, val x: Int) extends Foo(v)
defined class Bar

scala> new Bar(1, 1) == new Bar(1, 1)
res25: Boolean = true

scala> new Bar(1, 1) == new Bar(1, 2)
res26: Boolean = true
// ????
scala> new Bar(1,1) == Foo(1)
res27: Boolean = true

scala> class Baz(v: Int) extends Foo(v)
defined class Baz

scala> new Baz(1) == new Bar(1,1)
res29: Boolean = true //???

scala> println (new Bar(1,1))
Foo(1) // ???

scala> new Bar(1,2).copy()
res49: Foo = Foo(1) // ???

Credits:Why case classes should be final

So by default case classes should always be defined as final.

Example:

final case class User(name: String, id: Long)