(This article was first posted on my personal blog on 2018-08-04 at https://bbs-software.com/blog/2018/08/04/using-lambdas-to-simplify-varying-behaviors-in-your-code.html.)
Lambdas are anonymous (unnamed) functions. Unlike methods in object oriented programming languages, lambdas are not bound to any object. Although they are best known in the context of functional programming languages, object oriented programming languages that support lambdas, such as Ruby, can greatly benefit from their use. This is especially true when it comes to implementing varying behaviors(1).
By varying behaviors, I mean situations where the code that needs to run is not fixed, but rather specified, using parameters, subclassing, or configuration. Examples would include cases for which you would use the Command and Strategy design patterns.
Many developers are already familiar with using lambdas for event handling (implemented with the function keyword in JavaScript and also as arrow functions in ES6) , but there are many other cases in which they are helpful, a couple of which I will describe below.
The Problem with Non-Lambda Approaches
The procedural if-elsif-end or case clauses work when you have a small number of conditions and actions that are known in advance, but if you don't, they're pretty useless.
And although the object oriented approach of polymorphism by inheritance (2) can produce a correct result, in many cases it is unnecessarily verbose, ceremonial, and awkward.
To further embellish on this point, @kayess pointed me to a discussion on a Scheme forum where Anton van Straaten said about closures (lambdas are closures): "A closure's simplicity can be an asset: classes and interfaces can get in the way of simple parameterization of behavior. Anyone who's tried functional programming in Java or C++ has encountered this - it can be done, but it's more tedious." Anton goes on to provide an entertaining koan about the subject.
Furthermore, though we're accustomed to thinking about this problem in the context of a single customizable behavior, what if there are several?
Let's say we have a class that contains 3 varying behaviors. As an admittedly contrived example, let's say we have classes for each of hundreds of different species of animals, and they each have a move, sleep, and vocalize behavior. As a simplifying assumption, let's say that each of these behaviors has 7 possible variations, each of which is shared by many species. If we were to write a class to implement each possible set of the 3 behaviors, we would need the Cartesian product of classes, (7 * 7 * 7), or 343 classes! That would be a silly monstrosity for several reasons of course, one of which being we could simplify the design by providing a class hierarchy for each of the three kinds of behavior, and plug those into the larger class -- but then we would still need (7 + 7 + 7), or 21 classes! (Probably 24 really, as pure design would dictate an additional class as an abstract superclass for each set of 7 implementations).
If these behaviors are truly complex enough to justify a class of their own, this is not a problem. However, often they are not, and the solution is many times as verbose and complex as it needs to be.
A better solution is using callables such as lambdas.
Callables as a Superset of Lambdas
In traditional object oriented languages such as Java and C++, polymorphism is (in general) implemented by inheritance. Ruby does this also (3), but in addition, Ruby uses duck typing, meaning that any object that responds to the method name can be used, regardless of its position in the class hierarchy.
This means that in Ruby, since the method used to call a lambda is call, any object that responds to call can be used in place of a lambda.
It could be a lambda, an instance of a class, or even a class or module. This provides great flexibility in implementing varying behavior. You can choose what kind of object to use based on your situation. For complex behaviors you may want modules or classes, and for simpler behaviors a lambda will work just fine.
Since any object responding to call can be used in place of a lambda, I will use the term callable instead of lambda where applicable.
The Buffered Enumerable
I once worked on a project where I needed to implement buffering of multiple kinds of things received over network connections. I started writing the first one, and noticed how the code could be cleanly divided into two kinds of tasks: 1) knowing when to fetch objects into the buffer and other buffer management tasks, and 2) knowing how to fetch each block of objects and what else to do each time that fetch is performed (e.g. logging, displaying a message to the user, updating some external state).
Realizing that #1 would be common and identical to all cases, and only #2 would vary, I thought about how wasteful it would be to implement #1 separately in all cases. I thought about the admonition about high cohesion / low coupling, and the Unix axiom "do one thing well", and decided to separate the two.
The most natural way to design this functionality in Ruby is with an Enumerable, which will have access to all kinds of functional wizardry thanks to the methods it gets for free by including the Enumerable module. In addition, it can easily be used to generate an array by calling its to_a method.
This is the origin of the BufferedEnumerable class in my trick_bag gem. This class manages buffering but has no idea how to fetch chunks of data, nor what else to do at each such fetch; for that, the caller provides callables such as lambdas. (Upon user request, the ability to subclass it and override its methods was also added.) The result is a dramatic simplification, where the logic of buffering is defined in only one place, and the places it is used need not be concerned with its implementation (or its testing!).
To create an instance of this with callables, we call the BufferedEnumerable class method create_with_callables, which is defined as follows(4):
def self.create_with_callables(chunk_size, fetcher, fetch_notifier = nil)
instance = self.new(chunk_size)
instance.fetcher = fetcher
instance.fetch_notifier = fetch_notifier
instance
end
When a fetcher callable has been defined (as opposed to the use of the subclassing approach, where an overriding method is used), it is called with the empty data buffer and the number of objects requested as shown below whenever the buffer needs to be filled, as follows:
fetcher.(data, chunk_size)
(In Ruby, .( is an abbreviation for .call(.)
A trivial fetcher that merely fills the array of the requested chunk size with random numbers could look like this:
fetcher = ->(data, chunk_size) do
chunk_size.times { |index| data[index] = Random.rand }
end
Below is a pry example that illustrates the call to that fetcher, and its effect on the passed array:
[7] pry("")> a = []; fetcher.(a, 2)
2
[8] pry("")> a
[
[0] 0.4885287976297428,
[1] 0.5493143769524284
]
After the buffer is filled, if a fetch notifier callable has been defined (unlike the fetcher, this is optional), it too is called, with the data buffer:
fetch_notifier.(data)
A trivial fetch notifier might look like this:
->(data) { puts "#{Time.now} Fetched #{data.size} objects" }
(data.size will not necessarily be equal to chunk_size, especially on the last fetch.)
This notifier might produce something looking like this:
2018-07-26 17:19:47 +0700 Fetched 1000 objects
After defining the fetcher and fetch_notifier lambdas, we could call the class method for creating a BufferedEnumerable shown above as follows:
buffered_enumerable = BufferedEnumerable.create_with_callables( \
1000, fetcher, fetch_notifier)
This enumerable can be used to call each or any other of the rich set of methods available on the Enumerable module.
By parameterizing the behaviors with callables, we have increased the simplicity of the implementation by separating the two orthogonal tasks into separate code areas, and avoided the unnecessary overhead of the inheritance approach, which would have packaged these functions in classes.
Using Predicate Callables to Implement Filters
Predicates are functions that return a Boolean value, that is, either true or false. There are many uses for predicates in software: filters, boundaries, triggers, authentication results...again, anything that produces a true or false value.
Configurable predicates are another natural fit for using callables.
I once had to write a DNS mock server for network testing that could be configured to respond with specific behaviors based on the characteristics of the incoming request. In another situation more recently, I was writing some accounting software and wanted to be able to filter the working set of transactions based on date, category, etc.
Both cases were an excellent fit for using callables as filters.
In the case of the mock DNS server, there were multiple criteria for the DNS request filters, such as protocol (TCP vs. UDP), qtype (question type), qclass (question class), and qname (question name). I provided methods that return lambdas that filter for specific values for those attributes; for example, to create a filter that will return true only for the qname example.com, you would do the following:
MockDnsServer::PredicateFactory.new.qname('example.com')
The qname method (i.e. the method that returns a filter for exactly one qname value) is defined as (roughly):
def qname(qname)
->(message, _protocol = nil) do
eq_case_insensitive(message.qname, qname)
end
end
If you view the method body from left to right, you will notice the prominent -> and its corresponding end two lines beneath it, which tell you that the value returned by this method is a lambda. The leading underscore of the _protocol parameter is a convention that indicates that that parameter is unused by the lambda.
Notice that the qname parameter's value (e.g. "example.com") is effectively stored in the lambda that the method returns? This technique is called partial application, and is extremely useful when working with lambdas.
Does storing state in the lambda make it any less functional? Not really; the state is immutable and used only for comparison.
If we were to refine the filter by adding the requirement that the qtype be 'A' and the protocol be 'TCP', then we could call methods to return those filters as well, and combine all three using the all compound filter. (Other compound filters are any and none.) Here is what that would look like:
pf = MockDnsServer::PredicateFactory.new
filter = pf.all(
pf.qtype('A'),
pf.qname('example.com'),
pf.from_tcp
)
The all compound filter is nothing more than a simple wrapper around Ruby's Enumerable's all? method:
def all(*predicates)
->(message, protocol = nil) do
predicates.all? { |p| p.call(message, protocol) }
end
end
Why does all this work? The filters are interchangeable because they all have uniform inputs and outputs. That is, they take the same parameter list ([message, protocol] in this example), and they all return a value usable by the caller (true or false in this example). We've already seen one implementation, the lambda returned by the qname method shown above. Here is another one; this one returns true if and only if the message was sent over TCP:
def from_tcp
->(_message, protocol) { protocol == :tcp }
end
In this case, we only care about the protocol so we can ignore the first (message) parameter.
Conclusion
I hope I have been successful in persuading you to consider using callables for implementing variable predicates and actions.
These are just two examples. I will stop here for the sake of brevity, but if you have any questions or suggestions for elaboration, please contact me.
You can find me on:
Also, for more information about lambdas, feel free to check out my Ruby lambdas presentation (Speakerdeck slideshow and YouTube video).
Footnotes
(1) One might ask what else is there about lambdas other than varying behavior, but there are other reasons to use them in Ruby. For example, they can be used:
- as nested functions
- as truly private functions (which, unlike private methods, cannot be accessed with
send - for self invoking anonymous functions, to hide variables
- for defining functions where methods cannot be defined (e.g. in some RSpec code)
- for defining classes dynamically with the
class...endnotation which may be easier to understand than theClass.newapproach. - for chaining operations, as in an array of lambdas
(2) Polymorphism by inheritance is a key characteristic of object oriented design whereby, by virtue of having a common ancestor in the class hierarchy that contains the method in question, objects of different classes can respond to the same message (typically identified by a method or function name) differently. This can be a nice design in some cases, but in others it is an overly heavy handed solution to a simple problem, as it forces the developer to create multiple classes in a class hierarchy.
(2 (cont'd)) @kayess points out that the more precise terms in software engineering are nominal typing for polymorphism by inheritance, and structural typing for duck typing.
(3) To be more precise, Ruby supports polymorphism by inheritance, but not by checking the class hierarchy like most OO languages do. Instead, it is using duck typing, and merely calls the method by name; since a subclass will be able to call its superclass' method by default, it works.
(4) The omission of the fetcher and fetch_notifier parameters from the constructor is intentional. The constructor is intended to be used directly by the caller only when the subclassing approach is used; for lambdas the static method create_with_callables should be used.
This article may be improved over time. To see its revisions you can go to its Github commit history.
Top comments (0)