Tuesday 19 February 2019

Functions, Methods, and Interfaces in Go - Notes on Coursera course


Why Use Functions?

  • function - set of instructions grouped together, usually with a name
  • all programs in Go have to have function main() - where program's execution begins
func main() {
   fmt.Printf("Hello, world!")
  • main() function is a special in a sense we don't call it explicitly which is not the case for other functions
func PrintHello() {
   fmt.Printf("Hello, world!")

func main() {
  • function declaration starts with keyword func, function name, arguments, return type and body
  • Why using functions?
    • reusability (within the same project or across multiple projects via libraries)
    • abstraction
      • hiding details of implementation; only input-output behaviour is what we need to know (we look at function as a black box)
      • improves understandability
        • naming 
        • grouping of function calls

Function Parameters and Return Values

  • functions need some data to work on - they can be passed via function parameters
  • parameters are listed in parentheses after function name
  • arguments are supplied in the call
func foo(x int, y int) {
   return x * y

func main() {
   foo(2, 3)
  • Parameter Options
    • if no parameters are needed, put nothing in parentheses; you still need parentheses
    • list arguments of same type
func foo(x, y int) {
  • Return Values
    •  Functions can return value as result
    • type of return value after parameters in declaration
    • function call is on the right side of an assignment
func foo(x int) int {
   return x + 1

y := foo(1) // y gets assigned value 2
  • Functions can have multiple return values
    • their types must be listed in the declaration
func foo2(x int) (int, int) {
   return x, x + 1

a, b := foo2(3) // a is assigned 3, b 4

Call by Value, Reference

  • how are arguments passed to parameters during the function call
  • Call by Value
    • arguments are copied to parameters
    • data used is the copy of the original;
    • called function can't interfere with the original variables in the calling function
    • modifying parameters has no effect outside the function
func foo(y int) {
   return y + 1

func main() {
   x := 2
   fmt.Print(x) // still 2
  • tradeoffs of call by value
    • advantage: 
      • data encapsulation
      • function variables are changed only inside the function
    • disadvantage:
      • copying time
      • large object may take a long time to copy
  • Call by Reference
    •  programmer can pass a pointer as an argument
    • called function has direct access to caller variable in memory
    • we don't pass data but a pointer to data (the address of data)
func foo(y *int) {
   *y = *y + 1

func main() {
   x := 2
   foo(&x) // foo gets a copy of the address where x is in memory; foo is modifying x
   fmt.Print(x) // 3
  • tradeoffs of Call by Reference
    • advantages:
      • copying times; only the address is copied, not data
    • disadvantage:
      • (no) data encapsulation
      • called function can change data/variables in the calling function

Passing Arrays and Slices

  • how to pass an array to a function?
  • array arguments are copied
    • array can be big so this can be a problem
func foo(x [3]int) int { // NOTE size of array => argument is of type array (not slice)
   return x[0]

func main() {
   a := [3]int{1, 2, 3}
   fmt.Print(foo(a)) // entire array is copied!
  • instead, we can pass a reference - a pointer to the array (array pointer)
    • this can be messy for having to reference and dereference
func foo(x *[3]int) {
   (*x)[0] = (*x)[0] + 1

func main() {
   a := [3]int{1, 2, 3}
   fmt.Print(a) // [2 2 3]
  • proper Go approach is to pass slices instead of arrays
  • in Go: get used to use slices instead of arrays! you can almost always use slices instead of arrays
  • slice is a structure that contains 3 things:
    • pointer to an underlying array
    • length
    • capacity
  • if we pass a slice, we actually pass a copy of the pointer to an array
    • function can use that pointer directly, without the need to dereference/reference explicitly
func foo(sli int) int {
   sli[0] = sli[0] + 1

func main() {
   a := []int {1, 2, 3} // NOTE no size is specified between [] => this is a SLICE declaration!!!
   fmt.Print(a) [2 2 3]
  • instead of arrays or pointers to arrays pass slices to functions!!!

Well-Written Functions

  • how you should write functions so your code is well-organized and understandable
  • Understandability
    • code = functions + data
    • if you are asked to find a feature, you should be able find it quickly; also, your peer reviewers should also be able to find it quickly
    • if you are asked where data is used (defined, accessed, modified), you should be able to find it quickly
  • Debugging Principles
    • e.g. code crashes inside a function
    • two options for a cause:
      • function is written incorrectly
        • e.g. sorts a slice in a wrong order
      • data that function uses is incorrect
        • e.g. sorts slice correctly but slice has wrong elements in it
  • Supporting Debugging
    • functions need to be understandable
      • determine if actual bahaviour matches desired behaviour - this shall be easy to do
    • data needs to be traceable
      • we should be able to trace where did data come from
      • global variables complicate this
        • anybody can write into them
        • otherwise we know that calling function passed data...

Guidelines for Functions

  • Function Naming
    • behaviour should be understandable at a glance
    • parameter naming counts too
func ProcessArray(a []int) float {}  // BAD! What kind or processing? What is the meaning of a?

func ComputeRMS(samples []float) float {} // GOOD. We know what f-n does, and what is the argument

  • Functional Cohesion
    • function should perform only one "operation"
    • an "operation" depends on context
    • even from a function name we can see that it does only one thing
      • if you need to put two or more actions in a function name, that should raise an alert
    • merging behaviours makes code complicated
  • Reduce number of parameters
    • use few parameters
    • debugging requires tracing function input data
      • it's more difficult with large number of parameters
    • if you function has many parameters it may have bad functional cohesion - it does too many things; each functionality ("operation") needs its own set of inputs; 
    • group related arguments into a structure
      • e.g. if we want to pass 3 points in space to a function we might need to pass 2 x 3 = 9 arguments
      • improvement: define a struct point and pass 3 points: 
type point struct {x, y, z float}
  • best: define a struct triangle and pass a triangle as a single argument:
type triangle struct {p1, p2, p3 point}

Function Guidelines

  • Function Complexity
    • function length is the most obvious measure
    • functions should not be complex; easier to debug
    • short functions can be complex too
  • Function length
    • write complicated code with simple functions
    • function call hierarchy
  • Control-flow Complexity
    • how many control-flow paths are there in a function?
    • paths from the start to the end of function
    • if there are no if statements: one control-flow path
    • if there is if statement: two control-flow paths
    • control-flow describes conditional paths
    • functional hierarchy can reduce control-flow complexity: separate conditional code into separate functions


First-Class Values

  • Go treats functions as a first-class values
  • functions are First-class
  • Go implements some features of functional programming
  • Go treats functions as any other type like int, float...
    • variables can be declared to be a function type and then assigned a function
    • functions can be created dynamically, on the fly
      • so far we've been creating them dynamically, in the global space we'd use func
      • but they can be created dynamically, inside other functions
    • functions can be passed as arguments to functions
    • functions can be returned from functions
    • functions can be stored in structs

Variables as Functions 

  • declare variable as function
    • variable becomes an alias (another name) for that function
var funcVar func(int) int // "func(int) int" is function signature

func incFn(x int) int {
   return x + 1

func main() {
   funcVar = incFn // in assignment just use function name, without () as we're not calling function here

Functions as Arguments

  • functions can be passed to other functions as arguments
    • we have to use keyword func
func applyIt(afunc func(int) int, val int) int {
   return afunc(val)

func incFn(x int) int { return x + 1 }
func decFn(x int) int { return x - 1 }

func main() {
   fmt.Println(applyIt(incFn, 2)) // 3
   fmt.Println(applyIt(decFn, 2)) // 1

Anonymous Functions

  • functions don't need to have names
  • functions with no name are called anonymous 
  • when passing function to another function you usually don't need to name passed function
    • function is created right there at the call 
    • this comes from lambda calculus
func main() {
   v := applyIt(func (x int) int { return x + 1}, 2);
   fmt.Println(v) // 3

Returning Functions

Functions as Return Values

  • functions can create functions and return them
    • new function can have a different set of parameters; controllable parameters
  •  example: Distance to Origin function
    • takes a point (x, y coordinates)
    • returns distance to origin
    • what if I want to change the origin?
      • option1: origin becomes a parameter
      • option 2: we create a function for each origin (o_x, o_y)
        • origin is build in the returned function
        • func (float64, float64) float64 is the type of the function returned
func makeDistOrigin(o_x, o_y float64) func (float64, float64) float64 {
   fn := func (x, y float64) float64 {
      return math.Sqrt(math.Pow(o_x, 2) + math.Pow(o_y, 2))
  return fn

Special-Purpose Functions

  • we make special-purpose functions by giving them parameters (e.g. Dist1 and Dist2 have different origins)

func main() {
   Dist1 := MakeDistOrigin(0, 0)
   Dist2 := MakeDistOrigin(2, 2)
   fmt.Println(Dist1(2, 2))
   fmt.Println(Dist2(2, 2))

Environment (Scope) of a Function

  • every function has an environment ("scope")
  • set of all names that are valid inside a function; that you can refer inside the function
  • environment includes names defined locally, in the function
  • Lexical Scoping: Go is lexically scoped 
    • environment includes names defined in block where the function is defined
    • BK: this is called variable capturing 
    • when you start passing around functions as arguments, the environment goes along with functions 
var x int
func foo(y int) {
   z := 1


  • function + its environment, together
  • in Go, it is implemented as a structure which contains pointer to function and pointer to environment
  • when you pass function as an argument to another function, you pass its environment with it
  • at the place where this function is executed, it still has an access to variables from the place where it was defined
  • e.g. o_x and o_y are carried with returned function, and are accessible when its called later, wherever and whenever is called
  • variables are coming from the closure, from the environment where function was defined
func makeDistOrigin(o_x, o_y float64) func (float64, float64) float64 {
   fn := func (x, y float64) float64 {
      return math.Sqrt(math.Pow(o_x, 2) + math.Pow(o_y, 2))
   return fn

Variadic and Deferred

Variable Argument Number

  • it is possible to pass variable number or arguments to function; such function is called variadic
  • to specify this use ellipsis character: ...
  • such argument is treated as a slice inside the function
func getMax(vals ...int) int {
   maxV := -1
   for _, v := range vals {
      if v > maxV {
         maxV = v 
   return maxV

  • How to pass list of arguments to variadic function?
    • you can pass a comma-separated list of arguments
    • you can pass a slice
      • need a ... suffix
func main() {
   fmt.Println(getMax(1, 2, 6, 4))
   vslice := []int {1, 3, 6, 4}

Deferred Function Calls

  • call can be deferred until surrounding function completes
  • they don't get executed where they are explicitly called but after the surrounding function is done
  • typically used for cleanup activities
  • use keyword defer
func main() {
   defer fmt.Println("Bye!") // "Bye!" printed after "Hello!"
  • the arguments are NOT evaluated in a deferred way, they are evaluated immediately but the call is deferred
  • if you pass an argument, it is evaluated right there where defer statement is 
func main() {
   i := 1
   defer fmt.Println(i + 1) // 2 is printed second time
   i++ // 2
   fmt.Println(i) // 2 is printed first time


Classes and Encapsulation


  • What is OOP?
  • Go supports OOP
  • It does not have classes but something equivalent (structs)
  • What is class? Collection of data fields and functions that share a well-defined responsibility (they are all related to the same concept)
    • function in a class is called a method
  • Example: Point class
    • used in geometry program
    • data: x and y coordinate
    • functions:
      • DistToOrigin(), Quadrant()
      • AddXOffset(), AddYOffset()
      • SetX(), SetY()
  • class is a template; contains fields, not data


  • instance of the class
  • contains data
  • Example: instances of Point class


  • associated with OOP (and generally, with abstraction)
  • if there is a program using your class, you want to hide details
  • you want to prevent someone changing internal data; therefore we provide public methods that shall be used to modify the state of the object from the outside
  • Example: double distance to origin (double x and y)
    • option 1 (safe): expose method DoubleDist() which doubles x and y internally
    • option 2 (not safe): allow programmer to access x and y directly; but programmer can make mistake if for example they double x but forget to double y
    • by exposing methods we prevent such mistakes and object will always be in good state

Support for Classes (1)

No "class" keyword

  • there is no "class" keyword in Go
  • most OO languages have class keyword
  • Data fields and methods are defined inside a class block
  • example in Python:

class Point:
   def _init_(self, xval, yval):
      self.x = xval
      self.y = yval

Associating Methods with Data

  • Go has different way of associating methods with data
  • Go is using "receiver types"
  • data is some type 
  • method has a receiver type that it is associated with
  • BK: the approach is the same as in C: we have some struct type and if some function has to deal with it, we just pass to it a pointer to that struct. The terminology reminds me of Objective-C
  • type and function have to be defined in the same package
  • when we call a method we use dot notation
  • example: we want to associate function Double with our custom type MyInt
    • MyInt is the receiver type - it is specified before the name of the function
    • mi is the receiver object (instance of the receiver type) that double would be called on
type MyInt int

func (mi MyInt) Double () int { 
   return int(mi *2) 

func main() {
   v := MyInt(3)

  • Double() could be defined for multiple types (to have multiple receiver types); Go looks what's the type left of the dot (.) operator to find out what is the receiver type (and so which Double() function to call)
  • in the example above mi becomes an implicit argument of function Double() (just like this - pointer to the current instance of the class - is an implicit argument in C++ classes' methods)

Implicit Method Arguments

  • although it seems that Double() takes no arguments, there is one implicit (hidden) argument: and instance (object) of the receiver type
  • object v is an implicit argument to the method
    • call by value (that's how argument passing is done in Go)
    • a copy of v is made and passed to the function

Support for Classes (2)

  • in a normal OOP language lots of different data (fields in class) is associated with any number of methods
  • the same can be done in Go: we'll just make a struct with lots of various data and make it to be a receiver type
  • in struct you can group together arbitrary number of arbitrary data 

Structs, again

  • struct types compose data fields 
    • this is traditional feature of classes
type Point struct {
   x float64
   y float64

Structs with Methods

  • structs and methods together allow arbitrary data and functions to be composed
func (p Point) DistToOrig() {
   t := math.Pow(p.x, 2) + math.Pow(p.y, 2)
   return math.Sqrt(t)

func main() {
   p1 := Point(3, 4)


Controlling Access

  • Go provides lots of different support of encapsulation and how to keep data private
  • we want to be able to control data access
    • we want people to use data in a way we define - via functions/methods
    • we can define a set of public functions that allow another/external package to access the data
package data

var x int = 1

func PrintX() {
package main

import "data"

func main() {
  • PrintX function starts with capital leter => it gets exported
  • package main can access (see) x only through that exported function 
  • x can't be modified externally in the example above but we can allow that if we export another method which allows that

Controlling Access to Structs

  • we can do something similar to struct members
  • hide fields of structs by starting field name with lower-case letter
  • define public methods which access hidden data
  • example: need InitMe() to assign hidden data fields
package data

type Point struct {
   x float64
   y float64

func (p *Point) InitMe(xn, yn float64) {
   p.x = xn
   p.y = yn

func (p *Point) Scale (v float64) {
   p.x = p.x * v
   p.y = p.y * v

func (p *Point) PrintMe() {
   fmt.Println(p.x, p.y)

package main
import "data"

func main() {
   var p data.Point
   p.InitMe(3, 4)

  • access to hidden fields only through public methods

Point Receivers

Limitations of Methods

  • there are some limitations to the process of associating methods with receiver types we described above
(1)  Method cannot change the state of receiver object as it's passed by value 
  • receiver type/object is implicitly passed to the method
  • it is passed by value (like any function argument in Go) => method receives only its copy => method can't change receiver object (method can't modify  the data inside the receiver)
  • example: OffsetX() should increase coordinate x in object p1
func main() {
   p1 := Point(3, 4)
   p1.OffsetX(5)  // only temp copy of p1 inside OffsetX() is changed
(2) Large Receivers

  • if receiver object is large, lots of copying is required when you make a call
  • all object is copied onto the stack
type image [100][100]int
func main(){
   i1 := GrabImage()
   i1.BlurImage() // 10000 bytes gets copied on stack - this can be slow


Pointer Receivers

  • instead of passing objects by value we can pass by reference (pointer)
  • instead of using regular types for receiver types, we can use pointer to those types as receiver type
func (p *Point) OffsetX(v float64) {
   p.x = p.x + v

Point Receivers, Referencing, Dereferencing

No Need to Dereference

  • when using a pointer receiver there is no need to perform explicit dereferencing (as in the previous example where Point is referenced as p, not *p)
  • dereferencing is automatic with . operator

No Need to Reference

func main() {
   p := Point(3, 4)
   p.OffsetX(5)  // no need to do something like (&p).OffsetX()

Using Pointer Receivers

  • Good programming practice
    • all methods for a type have pointer receivers, or 
    • all methods for a type have non-pointer receivers
  • This is for mixing pointer/non-pointer receivers for a type will get confusing
  • pointer receiver allows modification



  • one of OOP properties
  • ability of an object to have different "forms" depending on the context
  • example: Area() function - the function with the same name can do the same thing but in a different way, depending on the context
    • rectangle: area = base * height
    • triangle: area = 0.5 * base * height
  • these two functions
    • at high level of abstraction, they are identical in a way what they do; they do the same thing - compute the area
    • at low level of abstraction, they are different, in a way how do they compute the area
  • We need Go's support for polymorphism to achieve this
  • How is polymorphism implemented in traditional OOP languages?


  • Go does NOT have inheritance
  • there are parent-child (base-derived; superclass-subsclass) relations between classes
    • superclass is a top level class
    • subclass extends from superclass, subclass inherits data and methods from a superclass
    • Example: 
      • Speaker superclass - represents anything that can make noise/speak
        • Speak() method prints "<noise>"
      • Subclasses Cat and Dog
        • Also have the Speak() method, inherited from the Speaker superclass
        • Cat and Dog are different forms of Speaker


  • subclass redefines a method inherited from the superclass 
  • example: Speaker, Cat, Dog
    • Speaker Speak() prints "<noise>"
    • Cat Speak() prints "meow"
    • Dog Speak() prints "woof"
  • without overriding Cat and Dog Speak() methods would print "<noise>" but overriding allows them to redefine Speak() methods to print what they want
  • Speak() is polymorphic: in context of Cat it prints "meow" and in context of Dog, it prints "woof"
  • although overriden in subclasses, the method keeps the same signature


  • interface is a concept used in Go to help us get polymorphism
  • we don't need inheritance or overriding to get polymorphism, we can get it with interfaces
  • interface is a set of method signatures (name, parameters, return values)
    • implementation is not defined
  • it is used to express conceptual similarity between types
  • example: Shape2D interface has two methods: Area() and Perimeter()
    • all 2D shapes must have Area() and Perimeter()
    • any type that has these two methods can be considered to be 2D shape

Satisfying the Interface

  • types satisfies the interface if type defines all methods specified in the interface
    • same method signatures (names, args, return values)
    • example: Rectangle and Triangle types satisfy the Shape2D interface if:
      • have Area() and Perimeter() methods
      • Additional methods are OK
  • similar to inheritance with overriding

Defining an Interface Type

  • use keyword interface

type Shape2D interface {
   Area() float64
   Perimeter() float64

type Triangle {

func (t Triangle) Area() float 64 {

func (t Triangle) Perimeter() float 64 {

  • in the example above, we can say that Triangle implements (satisfies) interface Shape2D
  • we don't state explicitly that Triangle implements interface; we just present to compiler the interface and methods and it infers which types can are satisfying which interfaces
  • we don't care what data is in Triangle, which fields/properties it has; it only matters that that type is set as receiver type for functions with the same signature as defined in the interface

Interface vs. Concrete Types

Concrete vs Interface Types

  • concrete and interface types are fundamentally different
  • Concrete Type
    • a regular type
    • Specifies the exact representation of the data and methods 
    • fully specified
    • complete method implementation included
    • it has data which is associated with it
  • Interface type:
    • just specifies some method signatures
    • not data
    • implementations are abstracted
    • interface type eventually gets mapped to a concrete type

Interface Values

  • can be treated like other values
    • assigned to variables
    • passed, returned
  • interface values have two components
    • Dynamic Type: concrete type which it is assigned to (like a class which implements an interface in classic OOP languages)
    • Dynamic Value: value of the dynamic type (like an instance of the class which implements an interface in classic OOP languages)
  • interface value is actually a pair (dynamic type, dynamic value)

Defining an Interface Type

type Speaker interface {

type Dog struct {
   name string

func (d Dog) Speak() {

func main() {
   var s1 Speaker // interface value
   var d1 Dog{"Brian"} // 
   s1 = d1 // legal as Dog satisfies interface Speaker; 
  • interface type is Speaker
  • interface value is s1
  • dynamic type is Dog
  • dynamic value is d1
  • pair is (Dog, d1)

Interface with Nil Dynamic Value

  • an interface can have a nil dynamic value (no dynamic value)
var s1 Speaker
var d1 *Dog // pointer
s1 = d1 // legal
  • d1 is pointer to Dog, it is not concrete object, has no data in it; d1 has no concrete value yet
  • s1 has dynamic type - Dog, but has NO dynamic value

Nil Dynamic Value

  • interface with dynamic type but not dynamic value; it is legal to call interface methods on nil dynamic value
  • can still call the Speak() method of s1
  • doesn't need a dynamic value to call interface methods
  • need to check inside the method
func(d *Dog) Speak() {
   if d == nil { // does it have dynamic value or not?
   } else {

var s1 Speaker
var d1 *Dog
s1 = d1
s1.Speak() // it is legal to call function on a non-assigned pointer!

Nil Interface Value

  • interface with nil dynamic type
  • very different from an interface with a nil dynamic value
  • we can't call interface methods as there is no underlying concrete type with methods to call, there is no method implementations
  • Nil dynamic value and valid dynamic type
    • can call a method since type is known
    • example:
var s1 Speaker
var d1 *Dog
s1 = d1
  • Nil dynamic type
    • cannot call a method, runtime error
var s1 Speaker
s1.Speak() // error!

Using Interfaces

  • why would we need interfaces? why are they used?

Ways to Use an Interface

  • need a function which takes multiple types of parameter
    • e.g. we want function to take either type int or type float
  • we want function foo() to accept parameter of type X or type Y
  • we can:
    • define interface Z
    • make types X and Y satisfy Z (BK: this is something like class extensions...as any type can at any time be extended to satisfy any interface; this can happen after some concrete type is defined)
    • make Z to be the type of the foo() argument
    • interface methods must be those needed (called) by foo()

 Pool in a Yard

  • I need to put a pool in my yard
  • Pool needs to fit in my yard
    • total area must be limited
  • Pool needs to be fenced
    • total perimeters must be limited
  • Need to determine if a pool shape satisfies criteria
  • FitInYard() bool
    • takes a shape as argument
    • returns true if the shape satisfies criteria


  • many possible shape types
    • rectangle, triangle, circle, etc...
  • FitInYard() should take many shape types
  • Valid shape types must have:
    • Area()
    • Perimeter()
  • Any shape with these methods is OK

Interface for Shapes

type Shape2D interface {
   Area() float64
   Perimeter float64

type Triangle {...}
func (t Triangle) Area() float64 {...}
func (t Triangle) Perimeter() float64 {...}

type Rectangle {...}
func (r Rectangle) Area() float64 {...}
func (r Rectangle) Perimeter() float64 {...}

  • Rectangle and Triangle satisfy interface Shape2D
func FitInYard(s Shape2D) bool {
   if (s.Area() < 100 && s.Perimeter() < 100) {
      return true}
   else {
      return false
  • parameter is any type that satisfies Shape2D interface

Empty Interface

  • specifies no methods
  • all types satisfy the empty interface
  • use it to have a function accept any type as a parameter
  • use interface{} to specify it
  • val can be any type
func PrintMe(val interface{}) {

Type Assertions

Concealing Type Differences

  • a lot of point of interfaces is to hide differences (or highlight similarities) between types
    • example: Triangles and Rectangles are treated in the same way in FitInYard() - as long as they satisfy Shape2D interface
    • different types which has some similarities are treated in the same way
  • sometimes you need to treat different types in different ways
  • sometimes we need to differentiate based on the type, to figure out what is the concrete type
  • in FitInYard() it does not matter what is the concrete type; it can be Rectangle or Triangle

Exposing Type Differences

  • example: Graphics program
  • DrawShape() will draw any shape, it can take any Shape2D as an argument
    • func DrawShape(s Shape2D) { ...
  • Underlying API has different drawing functions for each shape so they have to take particular, specific, concrete types as arguments:
    • func DrawRect (r Rectangle) { ...
    • func DrawTriangle (t Triangle) { ...
  • Inside DrawShape() we need to find out what is the concrete type of s so we know which underlying function to call; concrete type must be determined
  • type assertions are used for that

Type Assertions for Disambiguation

  • type assertions can be used to determine and extract the underlying concrete type 

func DrawShape(s Shape2D) bool {
   rect, ok := s.(Rectangle)
   if ok {
   tri, ok := s.(Triangle)
   if ok {

  • type assertions extract Rectangle from Shape2D
    • concrete type has to be specified in parentheses
  • If interface contains specified concrete type 
    • rect == specified concrete type
    • ok == true
  • If interface does not contain concrete type 
    • rect == zero value for that type
    • ok == false

Type Switch 

  • Interface can be satisfied by many concrete types; we might be interested only in some of them
  • Another way to do this disambiguation is to use switch
  • switch statement used with a type assertion
    • use keyword type in parentheses: .(type)
    • if s is Triangle then sh will be Triangle
func DrawShape(s Shape2D) bool {
   switch := sh := s.(type) {
   case Rectangle:
   case Triangle:

 Error Handling

  • common use of Error interface in Go

Error Interface

  • there are lot of Go functions built in packages where return two values: 
    • value
    • error (error interface objects which indicates an error)

type error interface {
   Error() string // prints the error message

  • correct / successful operation: error == nil
  • incorrect / failed operation: error != nil
  • if Go function returns an error (usually as a second value) you should check that error and handle it! (BK: compiler will complain if returned error is not checked explicitly in the code)
f, err := os.Open("harris/test.txt")
if err != nil {

  • check whether error is nil
  • if it's not nil, handle it

No comments: