Robert Heaton

Software Engineer /
One-track lover / Down a two-way lane

PFAB #10: First-class functions and dependency injection

23 Feb 2020

This post is part of my “Programming Feedback for Advanced Beginners” series, which helps you make the leap from knowing how to cobble messy programs toether to writing clean, elegant code. Subscribe now to receive PFAB in your inbox, every fortnight, entirely free.

In the previous edition of PFAB we began looking at a program that analyzes WhatsApp message logs, written by Adarsh Rao. We discussed some of the tradeoffs between taking a batch approach - where we load and process many records at once - and a streaming approach - where we load one record at a time, fully process it, and only then load and process the next record. We focussed in particular on how to write a clean batch pipeline. This week we’re going to look at how to make a modular streaming pipeline by using first-class functions and dependency injection.

This edition contains advanced, challenging material - this series isn’t called “Programming Feedback for Advanced Beginners” for nothing. If after reading this post you think “OK, that’s nice and all but I have no idea how to use any of it in my own code” then that’s fine and normal. My hope is that the next time you come across some code written in this style then you’ll think “oh I see what they’re doing here.” Then the time after you’ll start to really see what’s going on, and eventually you’ll be writing this type of code yourself in your sleep.

If you haven’t read the previous PFAB then make sure to read it before going any further. This project is written in Ruby, but should be broadly understandable even if you haven’t come across Ruby before.

Moaning about our un-modular streaming code

At the end of last week’s PFAB, we were attempting to update our streaming code to deal with the fact that WhatsApp messages can sometimes spread over multiple lines in a log file, like so:

7/28/19, 11:07 PM - Kimi: Here's an idea for the next time you return. 
Sample Message that is split over two-lines. 
7/30/19, 11:03 PM - Gaurav: More messages

We were lamenting the fact that our updated streaming code was starting to look quite gnarly and un-modular. Our code that parses log lines into messages was becoming deeply entwined with our code that processes messages once they had been parsed. Look at the call to update_stats buried deep inside our log-parsing code:

stats = {}
current_message = nil

# File.foreach reads the given file and runs
# the given "block" of code on each line in turn.
File.foreach("samplelog.txt") do |l|
  # Assume that we've written a function called
  # parse_raw_log_line that parses out the date, time,
  # message, and whether this line is the start of a new
  # message.
  parsed_line = parse_raw_log_line(l)

  # If this line is a new message, add the previous
  # message to our stats and reset current_message
  # to be this line.
  if parsed_line[:start_of_new_message]
    if !current_message.nil?
      stats = update_stats(current_message, stats)
    current_message = parsed_line

  # If this line is not a new message, it must be a
  # continuation of the previous line's message. We
  # therefore add its contents to the current message.
    current_message[:contents] += parsed_line[:contents]

# We need to manually add the final message to our stats,
# since it won't be added by the above loop (urgh)
stats = update_stats(current_message, stats)

By contrast, the code for our batch version of the program remained nicely separated, with all of its logic wrapped up inside well-separated functions:

raw_log ="samplelog.txt")
parsed_messages = parse_raw_logs(raw_log)
message_stats = calculate_stats(parsed_messages)

I promised that we would refactor our streaming code so that it looked as aesthetically pleasing as our batch code. Let’s see how.

Start from the end

When refactoring code to make it cleaner and tidier, I like to start from the end. How do I want my code to look to another programmer who is using it?

For the batch case this is easy. Our code is a pipeline with clearly separated steps - load data, parse data, analyze data. We want the results of one step to be passed directly into the next, and we want the code for each step to be entirely separate and isolated in different functions. When I was sketching out my version of the batch program, the first thing that I wrote was the high-level structure of the program:

raw_log ="samplelog.txt")
parsed_messages = parse_raw_log(raw_log)
message_stats = calculate_stats(parsed_messages)

This helped me see what functions I would need to write, the arguments that each function would need to accept, and the values that each function would need to return. With this chunked-up specification in hand, the rest of the program was just a matter of filling in the pre-designed blanks.

Designing our streaming code

However, I found that coming up with this kind of high-level layout for streaming code was rather harder. Have a quick think about how you might do it. Don’t worry if you can’t come up with anything.

The reason that streaming is harder is that we don’t have the simple, sequential execution that we had in the batch case. We have to read a line, parse it, check if we have finished composing a new message, add it to our stats if so, then go back, read another line, add it to our stats, and so on. We can’t simply assemble a line of functions that feed data into each other.

The root cause of my discontent with our current streaming code is that our update_stats function is buried deep inside our code that parses log lines. This troubles me because if we wanted to reuse our log-parsing code to parse a different log file and perform a different action on each message (for example, write it to a message database), we would have to go inside our log-parsing code and conduct deep invasive surgery to change the call to update_stats. By contrast, to make a similar update to our batch code we would just have to write a new write_to_db function and swap it in for our calculate_stats function, leaving the code that parses logs from files untouched.

Some plausible but unsatisfactory solutions to this problem include:

  • Having two almost-duplicate copies of the same function, called (for example) parse_log_file_and_calculate_stats and parse_log_file_and_copy_to_database. However, copy-pasting code is always a shame, and it’s particularly unfortunate that we’ll have to make a new function for every single new processing action that we may want to perform in the future too.
  • A slightly better alternative would be to pull our log-parsing code into a function called (say) parse_log_file, and to pass a string into parse_log_file that describes the processing action that we want to perform. Inside parse_log_file we would use a long if-statement to match the string to a function. For example:
def parse_log_file(filename, processing_action)
  # <lots of parsing code goes here>

  case processing_action
  when 'update_stats'
    stats = update_stats(message, stats)
  when 'write_to_db'
  # ...etc...

  # <lots more parsing code goes here>

However, this approach still has the disadvantage of requiring us to update parse_log_file every time we want to add a new processing action.

The solution to this conundrum is to use first-class functions and dependency injection.

First-class functions

First-class functions are a feature of many (but not all) programming languages. If a language has first-class functions, this means that you can pass its functions around as variables. To demonstrate what this means and why it matters, let’s switch for a second from Ruby to Python, where the syntax for first-class functions is simpler:

def say_hello():

x = say_hello
# => prints "HELLO!"

In Python, the line x = say_hello means “assign the function say_hello to the variable x”. It is very different to the more familiar x = say_hello() (note the () after say_hello), which means “evaluate the function say_hello, then assign the result to the variable x.”

After executing x = say_hello, x points to the say_hello function. This means that on the next line, x can be called like any other function by writing x(). This executes the code inside say_hello, printing “HELLO!”.

Dependency injection

Variables that point to functions can be passed around like any other type of variable. This allows you to pass around instructions and logic without having to immediately execute it. Passing around logic in this way is known as dependency injection, because you’re injecting logic that the receiving function depends on. Take a look at the following snippet, which demonstrates a function that takes another function as an argument, and then runs that function twice:

def say_hello():

def run_function_twice(my_function):
  """run_function_twice takes another function
  as an argument and runs that function twice"""

x = say_hello
# => prints "HELLO!" "HELLO!"

# Or simply:
# => prints "HELLO!" "HELLO!"

Translating first-class functions into Ruby

(If you don’t care about the specifics of Ruby and only care about the general principles of first-class functions, you can safely skip this section)

Strictly speaking, Ruby doesn’t have first-class functions. You can’t assign a method to a variable in the same way that you can in Python:

def say_hello()

# In Python this would point `x` at the
# `say_hello` function:
x = say_hello
# But in Ruby it executes `say_hello`
# immediately and prints "HELLO!"

However, even though Ruby doesn’t have true first-class functions, it does have a concept called a Proc, which is a different, slightly more verbose way of achieving the same thing (you don’t need to worry too much about the details of Procs unless you’re interested and want to look them up). In order to perform dependency injection and produce the same behavior as our Python example above, we need to:

  • Wrap the function we want to pass around inside
  • Invoke the Proc using .call(), instead of simply ()

For example:

def say_hello()

def run_proc_twice(my_proc):

x = { say_hello() }
# => prints "HELLO!"

# => prints "HELLO!" "HELLO!"

Don’t worry too much about the syntactic specifics for now - if the general idea of dependency injection and passing around logic as variables makes sense then you’re doing great. If it doesn’t then please do let me know.

How does any of this help us?

Recall the second band-aid solution that we proposed to our problem: pass parse_log_file a string describing the processing action that we want to perform, and then use a big if-statement inside parse_log_file to pick out and execute the logic that corresponds to this string.

This solution was OK, but how about we skip out the middleman? Instead of passing in a string that tells our function which logic to run, how about we pass in the logic itself in the form of a Proc?

At the top-level, this might look like this:

update_stats_proc = { |m| update_stats(m) }
parse_log_file(fname, update_stats_proc)

The parse_log_file function now accepts 2 arguments - a filename, and a Proc containing the logic that parse_log_file should perform whenever it finishes parsing a message. This means that if we want to add new functionality to our system, for example to parse messages and add them to a database, we can reuse the existing parse_log_file method and pass in a different Proc. Excitingly, we won’t have to update a single line inside parse_log_file.

write_to_db_proc = { |m| write_to_db(m) }
parse_log_file(fname, write_to_db_proc)

Side-note: Ruby has a more elegant “block” syntax that we can use to write the same logic in a more stylish way:

parse_log_file(fname) do |m|

(Google “ruby block” and “ruby yield” for more details)

Now that we’ve fleshed out how we want our program to look, all that remains is to implement the individual pieces that we’ve designed. We want parse_log_file to be a function that takes 2 arguments: a filename and a message-processing Proc. It should read the given file, parse WhatsApp messages from it, and invoke the message-processing Proc on every message as it is parsed.

Something like this ought to do the trick:

def parse_log_file(filename, message_processing_proc)
  current_message = nil
  File.foreach(filename) do |l|
    parsed_line = parse_log_line(l)

    if parsed_line[:new_message]
      if !current_message.nil?
        # If the line starts a new message, invoke
        # message_processing_proc on current_message
        # to process it.
      current_message = parsed_line
      current_message[:body] += parsed_line[:body]

  # Manually invoke message_processing_proc on the
  # final message.

message_count = 0
def update_stats(message)
  message_count += 1

update_stats_proc = { |m| update_stats(m) }
parse_log_file(fname, update_stats_proc)

And there we have it - a streaming pipeline with the different steps of the pipeline as cleanly separated as they were in the batch pipeline.

Future work

We can use first-class functions and dependency injection to make our code even more powerful and more modular.

Configurable batch sizes

In the previous edition of PFAB, we mentioned that it is possible to combine streaming and batch approaches. We’ve so far thought of “streaming” as loading and processing one record at a time, and “batch” as loading and processing every record at once. However, we can also load and process “some” records at a time, in batches of a configurable size.

There may be interesting tensions at play in our choice of batch size. We might want to keep our batches small to reduce the amount of memory that we use (see last week’s PFAB). But we might also want to keep our batches large to reduce the number of separate times we have to load data from our data source. In between these competing priorities there is likely a happy middle-ground. Where? That all depends.

Let’s once again take our top-down approach, and consider how we would like our batch-size code to look to other programmers who are using it. I would like it if our users simply passed in a batch_size parameter to our existing functions, and then we take care of the rest:

parse_log_file(fname, batch_size) do |m|

parse_log_file’s contract then becomes “you give me a filename, a message-processing function, and a batch size. I’ll load messages from your file in batches of batch_size, and run each record through the message-processing function.”

Go back to the previous definition of parse_log_file and think about how you might update it to work with a batch-size.

More modularity

The parse_log_file function is currently responsible for both:

  • Loading log data from a file
  • Parsing log data into messages

For our current purposes, it’s not a problem that parse_log_file has two jobs. However, suppose that we wanted to extend our system to work with additional data sources. In this expanded system, users can load their log data from not just a file, but also from a database or a website. The data in these new sources is formatted in the same way in each new source type as it is in the log file. We therefore want to be able to reuse our log-parsing code, and just want to be able to configure the source from which the logs are originally loaded.

We could write a band-aid solution to this problem, similar to those proposed for our earlier problems. We could make multiple functions called parse_log_database, parse_log_website, parse_log_file, although this would require us to duplicate some code. Or we could pass a sourcetype argument into parse_log_file that indicated whether the function should load its data from a database, website, or file.

However, I would prefer to create additional modular building blocks that users of our code can string together themselves, in the same way that users can already string together parsing and processing. I’d like our streaming code to look like this:

load_data_from_db(dbname) do |line|
  parse_and_transform_line(line) do |message|

The new component in the mix is load_data_from_db. Recall how our line-parsing function (now called parse_and_transform_line) works - it parses lines one-by-one, and passes each line into a user-provided Proc for further processing. In the same way, load_data_from_db loads lines from a database, and passes each line into a user-provided Proc for further processing.

This allows us to use the same line-parsing function, whether we’re reading the original from databases or websites. At the same time, we can seamessly swap out the piece of the pipeline that deals with loading data. In fact, now every step of our pipeline can be replaced and customized, independent of all the other steps. For example, if we wanted to load our data from a website instead of a database, and to translate the messages into French instead of calculating statistics, we could swap out our first and last components of our pipeline and write:

load_data_from_website(url) do |line|
  parse_and_transform_line(line) do |message|
    translate_message(message, 'french')

Once again, implementing these postulated functionse is left as a difficult exercise for another day or an especially motivated reader. This process is nonetheless a good example of the power of top-down design.

Should you actually write code like this?

Probably not. At least, not yet.

I’ve been somewhat less than honest in my presentation. The top-level code that I’ve sketched out is all very clean and delightful, but the lunch that it gives you is not free. Despite my hand-waving, actually implementing all of those methods will get complex quickly.

If you’re writing a library that will be used by many other programmers then this is just part of your job. It’s worth stuffing as much complexity as you can inside your library in order to make it as easy as possible for people on the outside to use. But if you’re writing a small project that is currently only worked on by you and maybe one or two other people, embellishments like this might be more trouble than they’re worth.

In particular, one of the biggest risks in erecting a framework inside a project is that it might turn out to be the wrong one. Adding new functionality to a framework that isn’t designed to handle it is much harder than adding functionality to a project with no framework at all.

For example, all of the above code has assumed that message records are processed one at a time, in a complete vacuum. This sounds like a reasonable assumption. But suppose that Adarsh (the original author of this program) wanted to be able to measure the time difference in between messages. This will require us to be able to peek back at past messages in order to read their timestamps, which is not a feature that our current framework has contemplated. Adding it to our now-opinionated framework will be substantially harder than adding it to a more freeform, unstructured program.

Don’t undervalue flexibility. Every clever flourish that you add to your code makes some use-cases easier to handle, but probably also makes others much harder. When you know for sure(ish) the types of tasks that your system will and won’t need to handle, then you can start to make assumptions and specialize. Until then, keep your options open, even if it makes your code a little more ragged around the edges for the time being.

Keep reading:

Subscribe to my new work on programming, security, and a few other topics. Published a few times a month.
Follow me on Twitter ➜ RSS ➜

More on Programming Projects for Advanced Beginners