Good Value Object Conventions for Ruby

In computer science, a value object is a small object that represents a simple entity whose equality is not based on identity: i.e. two value objects are equal when they have the same value, not necessarily being the same object. →

Creating good, reusable and idiomatic value objects for Ruby is not that simple.

This repository provides a checklist for a good value object design. Currently it is in "RFC" (request for comments) state, gathering experience, agreements and convention. In future, it will also have automated tests, so you can just

# in RSpec
it_behaves_like "good value object",
  arithmetic: false,
  ordered: false,
  sample_values: [
    {lat: 1, lng: 2},
    {lat: 50, lng: 40}
    ...
  ]

Examples

We are using imaginary, yet real-life-alike Quantity { amount: Numeric, unit: String } type for most of the examples. And, eventually, other types that demonstrate some points better.

Definitions

We can think about most value objects as a Struct (not Ruby's particular implementation, but generic programming concept: group of named fields). The fields of this structure we further will call structural elements. It is logical concept rather than implementational.

Example: for Date value type, "structural" values are probably (year, month, day of month) (maybe calendar too, depending of fanciness of your date). That does not imply that Date instance stores them in instance variables, neither the fact that it is the only instance variables:

• Date may be internally represented by one integer value, and calculate components back and force on construction and parts accessors;
• Date may have weekday as an accessor and instance variable. But it is probably derived value, because it indeed can be derived from year, month and day, and there are almost no situations where it can be used to specify the date (e.g. 2018, March, Monday is ambiguous, and 2018, March, 5th doesn't need weekday to be specific).

Note: 2018, 10th week, Monday is a thing in some business contexts, but probably it is better to have specialized constructor or even type for it.

Construction

• #initialize should have type's structural parts as an arguments, the TypeName.new(...) should be the most straightforward ("just validate and store in instance variables") way to construct value; all other ways to construct should go to specialized class methods

  # Bad
  Quantity.new('10 m')
  # Good
  Quantity.new(10, 'm')
  Quantity.parse('10 m')
  

  • See also "Conversions"" section
  • Note: there are still cases where string representation is the most natural for "default" constructor:

  # Probably OK
  IPAddress.new('192.168.0.1')
  # Does it bring additional clarity? YMMV
  IPAddress.new(192, 168, 0, 1)
  # Please don't!
  IPAddress.new(byte1: 192, byte2: 168, byte3: 0, byte4: 1)
  

• It is acceptable to have structural elements converted or wrapped on construction

  q = Quantity.new(10, 'm')
  q.amount # => #<BigDecimal 10>
  q.unit # => #<Quantity::Unit m>
  

• Prefer keyword arguments over positional ones in most cases, especially if there are more than 2 arguments for constructor, or order is not obvious (is it GeoPoint.new(lat, lng) or GeoPoint.new(lng, lat)?)

  • (Obvious yet mandatory: please, use real keyword arguments, not pre-Ruby 2.1 params = {} hack)

• Value construction options could be provided by keyword arguments, but it is undesirable to have both main argument and options as keyword arguments, or having both as positional arguments

  # OK
  Quantity.new(10, 'm')
  Quantity.new(amount: 10, unit: 'm')
  Quantity.new(10, 'm', system: Quantity::SI)
  # Questionable
  Quantity.new(amount: 10, unit: 'm', system: Quantity::SI)
  Quantity.new(10, 'm', Quantity::SI)
  

  • Note: Probably, specialized constructors are better than options in the generic constructor;

• Sometimes it is useful (but not required) to provide construction method synonymous with the type name, e.g. Quantity(amount, unit); it brings no additional functionality yet emphasizes the fact that value "just exists", and we are referencing to existing concept of "10 meters", not constructing it (which "new" implies);

  • As an alternative, consider providing {Type}.call class method: it allows to have almost the same look-and-feel, yet semantically belongs to the same module instead of being a global method:

  # Shortcut for .call, available since Ruby 1.9
  Quantity.(10, 'm')
  # Some prefer this alternative:
  Quantity[10, 'm']
  

• If there are expected to be a lot of similar objects created during the lifecycle of the application, consider caching objects (having exactly one object for one value). Type.new can be redefined for this purpose:

  10.times.map { Quantity.new(10, 'm') }.map(&:object_id).uniq.count # => 1
  

  Another approach seen in use is making Type.new private, and making Type(...) or Type.[] (with caching inside) the primary construction method.

  • Consider global caching with a great thoroughness: it may be not threadsafe, and may lead to a lot of memory eaten if never gets flushed.

• Avoid redefining .new for other purposes, especially to return value of type different from requested:

  # Really bad
  Quantity.new(10, 'm') # => #<Quantity::Physics::Length 10 m>
  # Something like this would be better
  Quantity.coerce(10, 'm') # => #<Quantity::Physics::Length 10 m>
  # or even
  Quantity['m'].new(10)
  

Basic properties

• All structural elements of the value should be exposed as attr_readers (or methods with the same behavior)

• Value object should be absolutely immutable, no attr_writers and no other way to change value of the object

  • It is wise to freeze all structural elements that belong to mutable Ruby types, to prevent code like this:

    q = Quantity.new(10, 'm')
    q.unit.upcase!
    # Or, more believable:
    q = Quantity.new(10, 'm')
    u = q.unit
    # ...later...
    u.upcase! # => Unexpectedly makes q to have unit == 'M'
    

• As immutability makes this code impossible:

  new_value = value.dup
  new_value.property = x
  

  consider providing some reasonable methods to "produce a value like this, with some parts changed"

  • Consider (but mindfully) merge(property: value, property: value) interface for it

    # Good
    FancyDate.now.merge(month: 12) # produces new FancyDate: "same day, but in December"
    # Not really useful
    Quantity.new(10, 'm').merge(unit: 's') # what's the semantics of "same value but in seconds"?..
    # Probably better
    Quantity.new(10, 'm').unit.create(20) # => Quantity(20, 'm')
    

    • #with is another frequently used option instead of #merge.

• No global option should change behavior of value objects. Consider providing "context" or "environment" to constructor or instance method:

  # Unforgivable bad
  Quantity.new(10, 'm').normalize # => Quantity.new(32.8, 'feet')
  Quantity.system = Quantity::SI
  Quantity.new(10, 'm').normalize # => Quantity.new(10, 'm')
  
  # Still pretty questionable
  Quantity.new(10, 'm').normalize # => Quantity.new(32.8, 'feet')
  Quantity.new(10, 'm', system: Quantity::SI).normalize # => Quantity.new(10, 'm')
  
  # Good
  Quantity.new(10, 'm').normalize # => Quantity.new(32.8, 'feet')
  Quantity.new(10, 'm').normalize(system: Quantity::SI) # => Quantity.new(10, 'm')
  
  # Best ;)
  Quantity.new(10, 'm').normalize # => Quantity.new(10, 'm')
  Quantity.new(10, 'm').normalize(system: Quantity::IMPERIAL) # => Quantity.new(32.8, 'feet')
  

#inspect and #pp

• You should implement #inspect for your types, it is really helpful for debugging
• By convention, #inspect for value types should look like #<TypeName value representation>
• Value representation should be full (without losing important details) yet concise (without variable names and unimportant clarifications)

  # Good
  Quantity.new(10, 'm').inspect # => "#<Quantity 10 m>" or #<Quantity(10 m)>`
  # Bad
  Quantity.new(10, 'm').inspect # => "#<Quantity(m)>"
  Quantity.new(10, 'm').inspect # => "10 m" - it is unhelpful to not be able to distinguish from string while debugging
  Quantity.new(10, 'm').inspect # => "#<Quantity amount=10 unit=\"m\">" - unnecessary verbosity
  # Also bad: Ruby's stdlib Date
  Date.today.inspect # => "#<Date: 2018-03-04 ((2458182j,0s,0n),+0s,2299161j)>" -- ((2458182j,0s,0n),+0s,2299161j) anybody?
  

• If it can be created, it should be possible to inspect; #inspect should try hard to never raise and never return anything except string

  # Good
  Quantity.new(INFINITY, 'm') # => ArgumentError on attempt to create, no problems with inspect
  # Acceptable
  Quantity.new(INFINITY, 'm').inspect # => "#<Quantity [UNREPRESENTABLE]>"
  # Bad
  Quantity.new(INFINITY, 'm').inspect # => ArgumentError or nil
  

• If it is known beforehand about some possible basic values the value object will try to represent, it is advisable to try providing nicer inspects, immediately readable

  # Not really helpful
  Quantity.new(10_000_000, 'm').inspect # => #<Quantity 10000000 m>
  # Good
  Quantity.new(10_000_000, 'm').inspect # => #<Quantity 10,000,000 m>
  # Could be acceptable in some contexts
  Quantity.new(10_000_000, 'm').inspect # => #<Quantity 1e7 m>
  

• As since Ruby 2.5 pp is required by default, consider implementing multiline #pretty_print for the value, especially if it contains lots of data that is reasonable to print in multiple lines
  • Documentation on implementing #pretty_print (pretty terse, yet enough to start) could be found here

Comparison

• Provide == method for values
  • Values should be equal if, and only if, all of their structural elements are equal or could be converted in a tuple of equal structured elements

  # example of the latter:
  Quantity.new(1000, 'm') == Quantity.new(1, 'km') # => probably true, unless the domain is some formal reporting system
  

  • == should NOT raise on attempt to compare with incompatible type: in Ruby, 1 == "1" is just false, not a deadly sin punished by exception
  • Value of other type could be considered equal, if it could be converted into value of current type without loosing context:

  # Good
  Quantity.new(1, 'm') == Unitwise(1, 'm')
  Date.parse('2017-05-01') == Time.parse('2017-05-01') # Doesn't work in Ruby though ;)
  # Bad
  Quantity.new(10, 'm') == 10 # could be helpful in some particular case yet source of hidden bugs
  

• See "Behavior in hashes" about overriding #eql?
• Never override #equal?
• Provide order comparison for values (<, > and so on) if, and only if, order on all acceptable values is defined and unambiguous
  • It is strongly advised to provide those methods by implementing <=> and including Comparable (and it will give you == for free)
  • <=> should NOT raise on attempt to compare with incompatible type, just return nil, Comparables implementation of other method will behave the most reasonable way: == will return false and < and other similar methods would raise ArgumentError
  • if implementing < and > by yourself, don't forget about <= and >=; and make them raise ArgumentError on incompatible types
• Consider providing positive?, negative? and zero? for the value if, and only if, their meaning is clear and semantically unambiguous
• If the order on values is strictly defined, consider providing Type::INFINITY constant or class method, for using in expressions like:

  ranges = {
    Quantity.new(1, 'm')...Quantity.new(10, 'm') => 'near',
    Quantity.new(10, 'm')...Quantity.new(100, 'm') => 'far',
    Quantity.new(100, 'm')...Quantity::INFINITY => 'nowhere'
  }
  ranges.select { |r, _| r.cover?(value) }....
  # and this
  value.clamp(Quantity.new(100, 'm'), Quantity::INFINITY) # "not lower the 100" one-side clamp
  

  Possible infinity concept interfaces:

  # Probably OK if used rarely, and constructor should not fail on this
  Quantity.new(Float::INFINITY, 'm')
  # Pretty clear yet no explicit type, can be hard to implement <=>
  Quantity::INFINITY
  # Also clear and typed, needs mindful implementation
  Quantity.infinity('m')
  

  • See also "Behavior in ranges" for notes about Range implementation quirks

Other operators

• Consider providing a subset of math operators (+, -, *, / and so on) if their meaning is obvious and unambiguous
• Try to follow "natural" intuition of mathematical operators (a + b == b + a, a - b = a + (-b) and so on)
  • Note that Ruby's intuition also redefines some of operators base qualities, when acceptable, for example, using + for concatenation (of strings and arrays), which is not commutative
• Don't override operators just because it is cool: using, say ~Quantity.new(10, 'm') to say "something about this quantity" (for example, producing range Quantity.new(9.5, 'm')..Quantity.new(10.5, 'm')) is witty yet leads to unguessable code
• Consider implementing | and & if:
  • value object is some kind of pattern, for this operators to mean "or" and "and"
  • value object represents some kind of range(s), for this operators to mean "union" and "intersection"

  Dates::Period.parse('2017-02') | Dates::Period.parse('2016-12')
  # => #<Dates::Period Dec 1-31 2016, Feb 1-28 2017>
  Dates::Period.parse_range('2017-01-30'..'2017-02-12') & Dates::Period.parse('2017-01')
  # => #<Dates::Period Jan 30-31 2017>
  

• Consider implementing === if value can be used as some kind of pattern

  # Messy
  if quantity.unit == 'm'
  elsif quantity.unit == 's'
  else ...
  
  quantities.select { |q| q.unit == 'm' }
  
  # Nice
  case quantity
  when Quantity::Unit('m')
  when Quantity::Unit('s')
  ...
  
  quantities.grep(Quantity::Unit('m'))
  

Conversions

To other types

• Consider providing #to_<type> to convert value object to other types
• #to_<type> protocol should be used only when format or precision of value is changed, but not when context is lost

  # Good
  BigDecimal('100').to_i # => it is the same number, just loses precision
  # Bad
  Quantity.new(10, 'm').to_i # => context is lost, Quantity#amount is much better convention
  
  # Acceptable
  Dates::Period.to_activercord # => may have sense in some context
  # Questionable
  Dates::Period.to_regexp # => probably, just #regexp would be better
  

To Ruby's core types

• Never provide "implicit conversion" methods (#to_str, #to_ary, #to_hash, #to_int) unless you really know what you do (= type is really kind of string/array/hash/integer); they'll convert values violently and unexpectedly;
• Never provide to_a either (unless it is kind of collection), as it will unexpectedly deconstruct the value on Array(value) call
  • This means that if the type is descendant of Struct, you should explicitly undef :to_a
  • Even for objects "somewhat resembling collection", it is better to provide one or more #each_<something> methods, returning Enumerator
• Always try to provide #to_h, it is really good for serialization:
  • #to_h should probably return hash with symbolic keys, containing exactly all the structural elements of value object and nothing more;
  • If value object's constructor uses keyword arguments, ValueType.new(**value.to_h) == value should be always true
• Always provide #to_s, as Ruby's default #to_s will expose object_id and look really unhelpful on string interpolations
  • for value objects that represent typed values (time, geometry, quantities) consider providing as "human-readable" #to_s as possible, without any quoting and type names;
  • for value objects that represent complicated domain structures, consider making #to_s just an alias to #inspect (see "#inspect and #pp" section above)

  # Good
  puts Quantity.new(10 'm') # "10 m"
  # Also good
  puts StoreId.fetch('xyz') # => "#<StoreId xyz>"
  # Questionable
  puts StoreId.fetch('xyz') # => "xyz" -- Loses too much of domain context
  # if you needed this to interpolate sql, probably #to_sql method would be better
  

  • If there are a lot of way to represent value as a string, consider providing #format(lot: of, **options) or strf<typename>

From other types

• Consider providing Type.from_<othertype>() methods for as much of basic Ruby types, and domain types, as possible;
• As with to_<othertype>, the from_ naming convention can ONLY be used if format or precision of data is changed, but not when context is lost or attached:

# Good
Quantity.from_a([10, 'm']) # => #<Quantity 10 m>
# Bad
Quantity.from_f(10, unit: 'm') # It is constructor (maybe specialized one), not "converter from Float"!

• Most of the time, Type.from_othertype(value.to_othertype) == value should be true;
• Sometimes, it is useful to provide two methods for conversion: one raising on incorrect input, and other just returning nil:

Quantity.from_a([10, 'm']) # => #<Quantity 10 m>
Quantity.from_a([10]) # => ArgumentError: expected 2-element array
Quantity.try_from_a([10]) # => nil

• If the domain data can have very variable string representation, consider providing two ways to parse:
  • Typename.parse(string) that accepts any input, tries to guess how to parse it, and returns nil if it absolutely can not;
  • Set of methods, or set of options, or pattern DSL allowing user to specify how data should be parsed:

  # set of methods:
  Quantity.amount_unit('10m') # => #<Quantity 10 m>
  Quantity.unit_amount('$10') # => #<Quantity 10 $>
  # set of options
  Quantity.from_s('$&nbsp;10', order: :unit_amount, separator: '&nbsp;')
  # pattern DSL
  Quantity.strpquantity('%amount (%unit)', '20 (m)')
  

  Note: strp<typename> is probably not the best convention, but it is like Ruby's Date.strptime

Behavior in hashes

• If it is a slightest possibility the value type could be used as a key in hashes, implement #hash, returning different number for different combinations of structural elements, and same number for same combination. The easiest implementation is probably

  def hash
    [each, of, structural, elements, self.class].hash
  end
  

• In this case #eql? method also should be implemented, as Hash uses it to decide on key's equality on #hash values collision (as number of possible integer values could be lower than number of possible value object values). Typically, it can be just an alias to #==, but if #== is forgiving, #eql? should be strict.

  # Imagine Paragraph class, which is just a wrapper around String, but with some fancy interface
  # It can have...
  def ==(other)
    @string == other.to_s
  end
  
  # In this case...
  h = {'test' => 1, Paragraph.new('test') => 2}
  # ...may lead to only ONE key being stored
  

• If #eql? implementation is different from #=='s, never implement it as a #hash comparison

  # Really bad: on hash collision Hash will have no means of telling two values one from other
  def eql?(other)
    hash == other.hash
  end
  

Behavior in ranges

For most of its functionality, Ruby's Range currently relies on value providing #succ (next value in ordered values space). Unfortunately, this includes case equality === too.

UPD: Since Ruby 2.6, #=== uses cover? underneath, so "not working" or "too slow" examples of case below are working correctly. So, rule about implementing #succ becomes simpler: "Implement it, if it makes unambigous sense for your type".

For code expecting to work under Ruby < 2.6, there stay two opposite rules:

• Consider providing #succ method if value space is small and has unambiguous granularity, to allow code like this:

  case DayOfWeek.current
  when DayOfWeek('Mon')..DayOfWeek('Thu')
  

• Consider consciously NOT providing #succ to explicitly disallow code like this:

  # Idiomatic, yet slow: calculates thousands of IPs inside range
  case ip
  when IP("172.16.10.1")..IP("172.16.11.255")
  ...
  
  # Can't be used in `case`, yet fast:
  if (IP("172.16.10.1")..IP("172.16.11.255")).cover?(ip) ...
  
  # Another case
  #
  # Ruby will try to do #succ on start value, but what it should be?
  # "Obvious" from the first sight Quantity.new(2, 'm') will leave Quantity.new(1.5, 'm') outside the comparison
  case quantity
  when Quantity.new(1, 'm')..Quantity.new(10, 'm')
  ...
  
  # The only solution, again:
  if (Quantity.new(1, 'm')..Quantity.new(10, 'm')).cover?(quantity)
  

Serialization/deserialization

• Consider providing reasonable #to_json implementation. For lot of cases, this should be enough (if you have provided #to_h which is strongly advised above):

  def to_json(*opts)
    to_h.to_json(*opts)
  end
  

• Consider your value object's YAML-friendliness
  • Default YAML implementation will dump all object's instance variables on YAML.dump, and just set them all to an uninitialized allocated object on YAML.load. You can alter this behavior by redefining methods with inventive and memoizable names encode_with(coder) and init_with(coder)

  # good
  - !ruby/object:Quantity
    amount: 1
    unit: m
  
  # not so good
  - !ruby/object:Quantity
    amount: 1
    unit: !ruby/object:Quantity::Unit
      name: meter
      synonym: metre
      plural: metres
      short: m
      domain: distance
      base: true
      system: !ruby/object:Quantity::System
        ...
    _memoized_method_cache_: # memoist was here....
    ...
  

Inheritance friendliness

For small value objects it is always a temptation to inherit from, to add several more methods, change constructor or formatting, required by current domain. Your types should be ready to be inherited, which most of the time, means not hardcoding class (by name or by value) in methods (or, sometimes, vice versa, hardcoding it, look at examples)

class FancyQuantity < Quantity
end

# Bad
FancyQuantity.new(10, 'm').inspect # => #<Quantity 10 m>, because #inspect hardcodes "#<Quantity" part
# solution is
def inspect
  "#<#{self.class} .... >"
end

# Probably bad
FancyQuantity.new(10, 'm') == Quantity.new(10, 'm') # => false, because #== has self.class == other.class

# solution?
def ==(other)
  # Bad: only values of exactly same type are compatible
  self.class == other.class && ...
  # Bad: Quantity#==(FancyQuantity) would work, but not vice versa
  other.kind_of?(self.class) && ...

  # Good: just hardcode the base
  other.is_a?(Quantity) && ...
  # ...or, sometimes, duck type
  other.respond_to?(:amount) && other.respond_to?(:unit) && ...
end

Pattern Matching

Value objects are a prime candidate for use with the pattern matching syntax introduced in Ruby 2.7.

The easiest way to integrate with pattern matching is to return a hash of attributes from #decontruct_keys. The simplest implementation is to just call #to_h, if that method is already available. This allows the object to be pattern matched using the hash syntax:

class Quantity
  ...

  def decontruct_keys(_keys)
    to_h
  end
end

case Quantity.new(10, 'm')
in unit: 'm', amount:
  puts "Metric: #{amount} meters"
in unit: 'ft', amount:
  puts "Imperial: #{amount} feet"
end

The _keys argument provided to #decontruct_keys can be ignored, but if generating the returned hash is an expensive operation, this argument can be used to optimize performance.

Consider also supporting the array syntax if the attributes have an obvious or intuitive ordering (See also: keyword arguments vs positional arguments in Construction). This is done by returning an array of attributes from the #deconstruct method:

class Quantity
  ...

  def deconstruct
    [@amount, @unit]
  end
end

case Quantity.new(10, 'm')
in amount, 'm'
  puts "Metric: #{amount} meters"
in amount, 'ft'
  puts "Imperial: #{amount} feet"
end