James Dabbs

Responsible Metaprogramming

21 Jul 2020

When done well, metaprogramming enables a very high-level, expressive style of coding. Some of the real gems of the Ruby ecosystem - like activerecord and rspec - lean heavily (and mostly successfully) on metaprogramming. It’s a powerful technique, but easily misused. Here’s what I’ve learned about metaprogramming so far - both the how-tos and and when-not-tos.

Caveat: this post is informed by my role at Procore writing Boring Business-class Software on a large team. If you’re hacking on a personal project, feel free to selectively disregard the Principles, skip on down to the Tools, and metaprogram your heart out. There are some absolutely perfect and delightful DSLs like Sonic Pi out there that break most of these “rules”.


Declarative Shell, Object-Oriented Core

On balance, Metaprogramming tends to make it easier for developers to express their intent, but harder to follow what’s really happening at runtime. As such, metaprogramming is well suited for declarations about object configuration, that happen at “compile” time (or whatever your application’s “startup” phase is), and not “runtime”. ActiveRecord::Base.has_many is a prime example here: it’s incredibly complex to trace what that actually does at runtime, but that rarely matters; we just declare our data model, and trust that the DSL will work it out.

Most of the complexity in day-to-day development in a large, existing system is about coordinating interactions of objects at runtime (and especially recombining them in new ways that they weren’t necessarily designed for). Metaprogramming aside, this is why it’s often helpful to keep your objects simple and lift the complexity into the factories that assemble them.

Altogether, I generally find that DSLs work very well in factory methods - where a high-level, grokkable declaration of the configuration is valuable - and tend not to work as well for describing lower-level, runtime object interactions. Most commonly, my DSLs are very thin factory layers that are only responsible for newing up an object and it’s collaborators. You could always just use the objects directly, but the DSL gives developers an easy-to-write and easy-to-read way to declare that object configuration (and gives me a place to validate that declaration).

Metaprogramming is like an onion - if find yourself spending hours on end hacking away at it, you and everyone around you are going to cry. Also, there are layers.

Keep It Composed

Most of the time, when metaprogramming goes bad, it’s because someone built a domain-specific language without really reflecting on the fact that they were writing a (little) language. Most object-oriented programmers have spent a considerable amount of time mastering the tools of their trade - e.g. inheritance and composition - and depriving them of those tools is a high cost that has to be minimized or justified.

Good OO design is successful precisely because it allows you to easily re-use your objects in new, novel contexts. Spend some time considering: what happens when a developer wants to extend your DSL? Re-use parts of it in new ways? Are there escape hatches to get back down to Plain Ol’ Ruby? Or is your language executed in a separate environment, with narrowly defined interface points to the host language? Either is fine, but it should be clear to your users what the boundaries between your DSL and the rest of the Ruby system are.

Setting Expectations

Many of the guidelines above are shorthand for encouraging the principle of least astonishment. If you’re building a language for other developers, you want it to be intuitive and ergonomic for them. rspec is a rather astonishing library, but it’s “unsurprising” in the sense that you express yourself in code more-or-less as you would express yourself to a co-worker.

If you can find natural reference points to help make your DSL “unsurprising”, great. But don’t overlook the other ways that you can set clear expectations with your users - like good documentation, examples, and consistent interfaces.


Ruby has a dizzying array of tools available for metaprogramming. What follows is an overview of the ones that I tend to find most helpful - when I reach for them, and what the sharp edges are.

Programming generally involves inspecting, creating and calling classes and methods. Unsurprisingly, there are metaprogramming tools to do all the same things.

Defining Classes

Ruby is strongly object-oriented. That means that everything in Ruby is an Object, including Classes and Modules. You can define new Classes on the fly, return them from methods, and don’t have to assign them to named constants.

class Base
  # ...

def base_with_field(name)
  # define a specialized subclass of `Base` with a named field
  Class.new(Base) do
    attr_accessor name

Similarly, many things that you might not think of as methods on classes, are still. For instance

class Collection
  include Enumerable
  # ...

is equivalent to

class Collection
  # ...



Defining Methods

You’re probably already familiar with a few methods-that-define-methods, like attr_reader or alias_method. There are a few other really handy options for defining methods on the fly.

define_method &c

The most straightforward way to dynamically define a method is the aptly-named define_method, which defines methods on a class, just like a normal def.

class Base
  def self.logging_attr_accessor(logger, *names)
    names.each do |name|
      define_method(name) do
        value = instance_variable_get(:"@#{name}")
        logger.info("Read #{name}=#{value}")

      # Defined methods can take arguments - even *splats,
      # keywords:, or &blocks - just like normal methods.
      define_method(:"#{name}=") do |value|
        instance_variable_set(:"@#{name}", value)
        logger.info("Set #{name}=#{value}")

  logging_attr_accessor Logger.new(STDOUT), :foo, :bar

There’s also the less commonly used define_singleton_method which can be used to define class methods (or methods on any single object).


method_missing allows you to hook into Ruby’s default NoMethodError handling, and supply your own fallback logic.

class Object
  def method_missing(name, *args)
    match = name.to_s.match(/^try_to_(.*)/)
    if match
        send(match[1], *args)

"test".try_to_upcase # => "TEST"
"test".try_to_floop  # => nil

“Methods” defined with method_missing will always have some limitations - e.g. they don’t show up in pry, and they may conflict with other named methods. A few rules here:

As an example, rspec uses method_missing to make these two equivalent

expect(result).to be_correct
expect(result.correct?).to eq true

which is definitely cool, but not really necessary. I tend to find it clearer and easier to maintain to have an single explicit method dedicated to handling those dynamic message

expect(result).to be(:correct?)

I’m serious about “sparingly”. I have exactly one method_missing implementation in production right now. It’s similar to the MagicMap example below, and I still periodically wonder if it was a good idea.

Calling Methods

send & public_send

If you don’t know what method you’re calling until runtime, it’s common to use send

def format(user, order=[:first_name, :last_name])
  order.map { |field| user.send(field) }.join(', ')

The terminology “send” comes from thinking of calling a method M on an object O (and it returning a value) as equivalent to sending O the message M (and checking its response).

send will allow you to call methods that should be private - that’s great if you’re in a REPL, but almost always a Bad Idea otherwise. If you’re sending a message that the receiver expects, you should be using public_send instead.

instance_variable_get & instance_variable_set

I almost always prefer to access the state of an object through its public interface, but there are occasions where you may need to get or set the internal state of an object directly. Ruby’s got you covered there.

object.instance_variable_set(:@variable_name, value)

This can be especially handy in constructors

class ValueObject

  def initialize(json_hash)
    json_hash.each do |key, value|
      instance_variable_set(key.to_sym, value)

instance_eval & instance_exec

Less common, but you may want to dynamically change the receiver of a block (the implicit self and the @ivar context). You can do that using instance_exec or [instance_eval]((https://apidock.com/ruby/BasicObject/instance_eval)

class MagicMap
  attr_reader :declarations

  def initialize
    @declarations = {}

  def method_missing(name, *args)
    @declarations[name] = args

  def self.declare(&dsl)
    new.tap do |m|

MagicMap.declare do
  foo :bar
  baz :quux
# => { foo: :bar, baz: :quux }

Do note that this gives you another options for accessing instance variables, if you know the name of them

object.instance_exec do
  @foo = :bar

This can similarly can be used to define methods, but I tend to prefer using the more purpose-tailored and intention-revealing methods.

The only differences between instance_exec and instance_eval are that instance_exec allows you to pass arguments to your block (which is good) and instance_eval allows you to pass in a string to be evaluated as Ruby code (which is bad). As such, I generally use instance_exec alone, and avoid instance_eval.

instance_exec is powerful, but should be used sparingly. It can be very surprising when self isn’t what a user expects it to be.


Don’t. Just don’t. Seriously - with a modern Ruby, I’ve never seen a use case where a literal eval was necessary. It just opens up a host of security concerns.

Inspecting Methods

Object#respond_to? and Module#method_defined? let you tell if a method already exists on an object or module, respectively. Object#methods lets you get the full list of all messages that an object responds to.

If you want more information about an existing method, you can use tools like Module#instance_method or Object#method to get an object representing the method and ask it questions about things like its source, source_location, parameters or arity.

Inspecting Classes

Object.is_a? and Class.ancestors let you check where in the inheritance hierarchy any given object sits.

ObjectSpace is an interesting sledgehammer for inspecting all of the classes in your system. This is heavy enough that I tend not to use it in production, but I wrote an earlier post with some examples of what’s possible here.

Hooks and Extension Points

One useful programmatic option that’s not available when programming by hand is extending some of Ruby’s built in inheritance mechanisms, e.g. using Class.inherited or Module.included to run extra code when these events happen.


Taking extension points a step further, Ruby allows you to hook into the VM layer and react to events like methods being defined or called. This is another big, heavy tool that I’ve written about separately (and tend to avoid in production code).


Declarative Factories

I have a fairly standard snippet that I use for factory methods

module Definable
  def self.included(other) # 1

  module ClassMethods
    def define(*args, &setup) # 2
      dsl = self::DSL.new(*args) # 3
      setup.arity.zero? ? dsl.instance_exec(&setup) : setup.call(dsl) # 4
      dsl.build # 5

Unpacking the parts here:

  1. We use the included hook to allow this module to define class methods when included. (If you have access to ActiveSupport::Concern, you should probably just use that instead.)
  2. Definable classes have a define method that takes a block (the DSL to run), and possibly some other arguments.
  3. Definable classes must define an inner DSL class that provides the execution context for that DSL block.
  4. Usually it’s easiest if the DSL block is executed with its implicit self being the DSL instance. This line does that by default, but provides users with an escape hatch:

     Thing.define do
       # `self` is a `Thing::DSL` instance
     Thing.define do |dsl|
       # `dsl` is the `Thing::DSL` instance
       # `self` is still `Thing`
  5. The inner DSL class must define a build method, which is responsible for calling new on the defined class.

I find that this setup encourages several nice effects:

Wrapping Module

A pattern I’ve seen pop up a few times now involves needing to dynamically wrap some-or-all methods on an existing object with a block - e.g. to add timing, science experiments, or error handling. By putting together a few of the techniques above, we can package up this pattern:

class MethodWrapper
  def initialize(&handler)
    @handler = handler # 1

  def wrap(target_class, methods=target_class.instance_methods(false)) # 2
    handler = @handler # 3

    wrapper = Module.new do # 4
      methods.each do |name|
        define_method(name) do |*args, &block|
          handler.call(name) do # 6
            super(*args, &block) # 5

    target_class.prepend(wrapper) # 5
  1. We instantiate the wrapper with a block that we’ll use to decorate methods.
  2. We supply a target_class that we want to decorate and a set of method names (by default, all methods defined on the class).
  3. What follows will execute in a different object context, so we won’t have a reference to our @handler instance variable. We assign it to a local variable, so that subsequent blocks close over and include it.
  4. Modules are objects too, so we can instantiate a new one which dynamically defines all the given methods.
  5. We’re going to prepend the module, so calls to super will then call the existing implementation in target_class.
  6. Altogether, the module method implementations wrap each named method in the provided handler, which receives the wrapped method name as a parameter.

With that in place, we can add decorated behavior like so

wrapper = MethodWrapper.new do |name, &method|
  start  = Time.now
  result = method.call
  puts "#{name} took #{Time.now - start}s"

# upcase took 2.0e-06s
# => "HELLO"

send Off

Ruby’s superpower is that is lets developers express themselves in any way they can dream up. When working with other developers, remember that you’re writing for your teammates as much as for yourself and the interpreter. Used well, metaprogramming allows you to develop a shared language that’s well-tailored to the problem at hand, and a joy to work with. Take your cues from Ruby, and try to build little languages that make people happy.