These are the slides of the talk I gave on May 22, 2014 to the San Francisco Scala user group. Similar talks were given before at GOTO Chicago, keynote, at Gilt Groupe and Hunter College in New York, at JAX Mainz and at FlatMap Oslo.
1 of 54
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
More Related Content
Scala - The Simple Parts, SFScala presentation
1. Scala - The Simple Parts
Martin Odersky
Typesafe and EPFL
3. Grown Up?
Scala’s user community is pretty large for its age group.
~ 100’000 developers
~ 200’000 subscribers to Coursera online courses.
#13 in RedMonk Language Rankings
Many successful rollouts and happy users.
But Scala is also discussed more controversially than usual
for a language at its stage of adoption.
Why?
3
4. Controversies
Internal controversies: Different communities don’t agree
what programming in Scala should be.
External complaints:
“Scala is too academic”
“Scala has sold out to industry”
“Scala’s types are too hard”
“Scala’s types are not strict enough”
“Scala is everything and the kitchen sink”
Signs that we have not made clear enough what the
essence of programming in Scala is.
4
5. 1-5
The Picture So Far
Agile, with lightweight syntax
Object-Oriented Functional
Safe and performant, with strong static tpying
= scalable
6. What is “Scalable”?
1st meaning: “Growable”
– can be molded into new languages
by adding libraries (domain
specific or general)
See: “Growing a language”
(Guy Steele, 1998)
2nd meaning: “Enabling Growth”
– can be used for small as well as
large systems
– allows for smooth growth from
small to large.
6
7. A Growable Language
• Flexible Syntax
• Flexible Types
• User-definable operators
• Higher-order functions
• Implicits
...
Make it relatively easy to build new DSLs on top of Scala
And where this fails, we can always use the macro system
(even though so far it’s labeled experimental)
7
9. Growable = Good?
In fact, it’s a double edged sword.
– DSLs can fracture the user
community (“The Lisp curse”)
– Besides, no language is liked
by everyone, no matter whether
its a DSL or general purpose.
– Host languages get the blame
for the DSLs they embed.
Growable is great for experimentation.
But it demands conformity and discipline for large scale
production use.
9
10. A Language For Growth
• Can start with a one-liner.
• Can experiment quickly.
• Can grow without fearing to fall off the cliff.
• Scala deployments now go into the millions of lines of
code.
– Language works for very large programs
– Tools are challenged (build times!) but are catching up.
“A large system is one where you do not know that some of
its components even exist”
10
11. What Enables Growth?
Unique combination of Object/Oriented and Functional
Large systems rely on both.
Object-Oriented Functional
Unfortunately, there’s no established term for this
object/functional?
11
13. But unfortunately it’s often more like this
OOP
FP
And that’s where we are
How many OOP
people see FP
How many FP
people see OOP
14. Scala’s Role in History
14
(from:James Iry: A Brief, Incomplete, and
Mostly Wrong History of Programming
Languages)
15. 15
Another View: A Modular Language
Large Systems
Object-Oriented Functional
Small Scripts
= modular
16. Modular Programming
Systems should be composed from modules.
Modules should be simple parts that can be combined in
many ways to give interesting results.
– (Simple: encapsulates one functionality)
But that’s old hat!
– Should we go back to Modula-2?
– Modula-2 was limited by the Von-Neumann Bottleneck (see John
Backus’ Turing award lecture).
– Today’s systems need richer models and implementations.
16
17. FP is Essential for Modular Programming
Read:
“Why Functional Programming Matters”
(John Hughes, 1985).
Paraphrasing:
“Functional Programming is good because it leads to
modules that can be combined freely.”
17
18. Functions and Modules
Functional does not always imply Modular.
Some concepts in functional languages are at odds with
modularity:
– Haskell’s type classes
– OCaml’s records and variants
– Dynamic typing? (can argue about this one)
18
19. Objects and Modules
Object-oriented languages are in a sense the successors of
classical modular languages.
But Object-Oriented does not always imply Modular either.
Non-modular concepts in OOP languages:
– Smalltalk’s virtual image.
– Monkey-patching
– Mutable state makes transformations hard.
– Weak composition facilities require external DI frameworks.
– Weak decomposition facilities encourage mixing domain models
with their applications.
19
20. Scala – The Simple Parts
Before discussing library modules, let’s start with the
simple parts in the language itself.
They correspond directly to the fundamental actions that
together cover the essence of programming in Scala
compose – match – group – recurse – abstract – aggregate – mutate
As always: Simple ≠ Easy !
20
21. 1. Compose
Everything is an expression
everything can be composed with everything else.
if (age >= 18) "grownup" else "minor"
val result = tag match {
case “email” =>
try getEmail()
catch handleIOException
case “postal” =>
scanLetter()
}
21
22. 2. Match
Pattern matching decomposes data.
It’s the dual of composition.
trait Expr
case class Number(n: Int) extends Expr
case class Plus(l: Expr, r: Expr) extends Expr
def eval(e: Expr): Int = e match {
case Number(n) => n
case Plus(l, r) => eval(l) + eval(r)
}
Simple & flexible, even if a bit verbose.
22
23. The traditional OO alternative
trait Expr {
def eval: Int
}
case class Number(n: Int) extends Expr {
def eval = n
}
case class Plus(l: Expr, r: Expr) extends Expr {
def eval = l.eval + r.eval
}
OK if operations are fixed and few.
But mixes data model with “business” logic.
23
24. 3. Group
Everything can be grouped and nested.
Static scoping discipline.
Two name spaces: Terms and Types.
Same rules for each.
def solutions(target: Int): Stream[Path] = {
def isSolution(path: Path) =
path.endState.contains(target)
allPaths.filter(isSolution)
}
24
25. Tip: Don’t pack too much in one expression
I sometimes see stuff like this:
jp.getRawClasspath.filter(
_.getEntryKind == IClasspathEntry.CPE_SOURCE).
iterator.flatMap(entry =>
flatten(ResourcesPlugin.getWorkspace.
getRoot.findMember(entry.getPath)))
It’s amazing what you can get done in a single statement.
But that does not mean you have to do it.
25
26. Tip: Find meaningful names!
There’s a lot of value in meaningful names.
Easy to add them using inline vals and defs.
val sources = jp.getRawClasspath.filter(
_.getEntryKind == IClasspathEntry.CPE_SOURCE)
def workspaceRoot =
ResourcesPlugin.getWorkspace.getRoot
def filesOfEntry(entry: Set[File]) =
flatten(workspaceRoot.findMember(entry.getPath)
sources.iterator flatMap filesOfEntry
26
27. 4. Recurse
Recursion let’s us compose to arbitrary depths.
This is almost always better than a loop.
Tail-recursive functions are guaranteed to be efficient.
@tailrec
def sameLength(xs: List[T], ys: List[U]): Boolean =
if (xs.isEmpty) ys.isEmpty
else ys.nonEmpty && sameLength(xs.tail, ys.tail)
27
28. 5. Abstract
Functions are abstracted expressions.
Functions are themselves values.
Can be named or anonymous.
def isMinor(p: Person) = p.age < 18
val (minors, adults) = people.partition(isMinor)
val infants = minors.filter(_.age <= 3)
(this one is pretty standard by now)
(even though scope rules keep getting messed up
sometimes)
28
29. 6. Aggregate
Collection aggregate data.
Transform instead of CRUD.
Uniform set of operations
Very simple to use
Learn one – apply everywhere
29
30. “The type of map is ugly / a lie”
30
Collection Objection
31. “The type of map is ugly / a lie”
31
Collection Objection
32. Why CanBuildFrom?
• Why not define it like this?
class Functor[F[_], T] {
def map[U](f: T => U): F[U]
}
• Does not work for arrays, since we need a class-tag to
build a new array.
• More generally, does not work in any case where we
need some additional information to build a new
collection.
• This is precisely what’s achieved by CanBuildFrom.
32
33. Collection Objection
“Set is not a Functor”
WAT? Let’s check Wikipedia:
(in fact, Set is in some sense the canonical functor!)
33
34. Who’s right, Mathematics or Haskell?
Crucial difference:
– In mathematics, a type comes with an equality
– In programming, a single conceptual value might have more than
one representation, so equality has to be given explicitly.
– If we construct a set, we need an equality operation to avoid
duplicates.
– In Haskell this is given by a typeclass Eq.
– But the naive functor type
def map[U](f: T => U): F[U]
does not accommodate this type class.
So the “obvious” type signature for `map` is inadequate.
Again, CanBuildFrom fixes this.
34
35. 7. Mutate
It’s the last point on the list and should be used with
restraint.
But are vars and mutation not anti-modular?
– Indeed, global state often leads to hidden dependencies.
– But used-wisely, mutable state can cut down on boilerplate and
increase efficiency and clarity.
35
36. Where I use Mutable State
In dotc, a newly developed compiler for Scala:
– caching lazy vals, memoized functions,
interned names, LRU caches.
– persisting once a value is stable, store it in an object.
– copy on write avoid copying untpd.Tree to tpd.Tree.
– fresh values fresh names, unique ids
– typer state 2 vars: current constraint & current diagnostics
(versioned, explorable).
36
37. Why Not Use a Monad?
The fundamentalist functional approach would mandate that
typer state is represented as a monad.
Instead of now:
def typed(tree: untpd.Tree, expected: Type): tpd.Tree
def isSubType(tp1: Type, tp2: Type): Boolean
we’d write:
def typed(tree: untpd.Tree, expected: Type):
TyperState[tpd.Tree]
def isSubType(tp1: Type, tp2: Type):
TyperState[Boolean]
37
38. Why Not Use a Monad?
Instead of now:
if (isSubType(t1, t2) && isSubType(t2, t3)) result
we’d write:
for {
c1 <- isSubType(t1, t2)
c2 <- isSubType(t2, t3)
if c1 && c2
} yield result
Why would this be better?
38
39. A Question of Typing
Clojure Scala Haskell Idris Coq
syntax arguments effects values correctness
Statically checked properties
None of the 5 languages above is “right”.
It’s all a question of tradeoffs.
39
40. From Fundamental Actions to Simple Parts
Compose
Nest
Match
RecurseAbstract
Aggregate
Mutate
40
Expressions
Scopes
Function Values
Collections
Vars
Patterns
Functions
41. Modules
Modules can take a large number of forms
– A function
– An object
– A class
– An actor
– A stream transform
– A microservice
Modular programming is putting the focus on how modules
can be combined, not so much what they do.
In Scala, modules talk about values as well as types.
41
42. Scala’s Modular Roots
Modula-2 First language I programmed intensively
First language for which I wrote a compiler.
Modula-3 Introduced universal subtyping
Haskell Type classes Implicit parameters
SML modules
Object ≅ Structure
Class ≅ Functor
Trait ≅ Signature
Abstract Type ≅ Abstract Type
Refinement ≅ Sharing Constraint
42
43. Features For Modular Programming
1. Our Vocabulary: Rich types with static checking and
functional semantics
– gives us the domains of discourse,
– gives us the means to guarantee encapsulation,
– see: “On the Criteria for Decomposing Systems into Modules”
(David Parnas, 1972).
2. Start with Objects − atomic modules
3. Parameterize with Classes − templates to create modules
dynamically
4. Mix it up with Traits − mixable slices of behavior
43
44. 5. Abstract By Name
Members of a class or trait can be concrete or abstract.
Example: A Graph Library
44
45. Where To Use Abstraction?
Simple rule:
– Define what you know, leave abstract what you don’t.
– Works universally for values, methods, and types.
45
47. 6. Abstract By Position
Parameterize classes and traits.
class List[+T]
class Set[T]
class Function1[-T, +R]
List[Number]
Set[String]
Function1[String, Int]
Variance expressed by +/- annotations
A good way to explain variance is by mapping to abstract
types.
47
48. Modelling Parameterized Types
class Set[T] { ... } class Set { type $T }
Set[String] Set { type $T = String }
class List[+T] { ... } class List { type $T
}List[Number] List { type $T <: Number }
Parameters Abstract members
Arguments Refinements
49. 7. Keep Boilerplate Implicit
Implicit parameters are a rather simple concept
But they are surprisingly versatile!
Can represent a typeclass:
def min(x: A, b: A)(implicit cmp: Ordering[A]): A
49
50. Implicit Parameters
Can represent a context:
def typed(tree: untpd.Tree, expected:
Type)(implicit ctx: Context): Type
def compile(cmdLine: String)
(implicit defaultOptions: List[String]): Unit
Can represent a capability:
def accessProfile(id: CustomerId)
(implicit admin: AdminRights): Info
50
52. Summary
A fairly modest (some might say: boring) set of parts that
can be combined in flexible ways.
Caveat: This is my selection, not everyone needs to
agree.
52
53. Other Parts
Here are parts that are either not as simple or do not work as
seamlessly with the core:
– Implicit conversions
– Existential types
– Structural types
– Higher-kinded types
– Macros
All of them are under language feature flags or experimental
flags.
– This makes it clear they are outside the core.
– My advice: Avoid, unless you have a clear use case.
– e.g, you use Scala as a host for a DSL or other language.
53