A tech stack is more often inherited than it is chosen. On the rare occasion someone gets the chance to decide which tech the company gets to use for the next decade. I've seen this go well and I've seen this go belly up.

If you're in the right club, you know the answer; stick with boring.

The Postgres Paradox

When thinking about how long a technology will last I often use the current age of the software to determine how much longer it will be around. So if it's 2 years old, it might be around 2 more years. Postgres dates back 35 years, so it will probably be around for 35 more years.¹

We can view this in terms of experienced employees; we are more likely to find someone with 20 years of Postgres experience than we are to find someone with 20 years in a tech that's only existed for 2 years.

Taking the age of the tech into consideration allows us to earmark the hiring pool and the likelihood that fixes or workarounds exist for problems we might run into. Diving into newer tech increases the chance we are the first to come across certain problems.

The bleeding edge might appear shiny and new but we aren't here to ship blog posts we are here to ship features and delight customers.

These avoidable issues are hard to solve because of two reasons:

1. They're novel, we're the first to run into them.
2. The people that have faced this issue are expensive, because not many of them exist.

Make a Difference

Starting a business is a risk, picking fancy new tech is a risk. These risks add up. Where you decide to take the risk depends on where you are trying to drive innovation.

In 2015, London, two neobanks were founded; this is the tech they are built on and the tech's age at the time. Revolut is built on Postgres (25yo) and Java (19yo), the other is built on Cassandra (7yo) and Golang (6yo). The former has 50million customers, the later has 10million.

Risks are allowed to be taken, but we should limit the number of unknowns and not make things harder than they need to be. If we pick our entire stack on the principle of shiny and new we are going to be fighting our tech more often that we are solving customer use-cases.

Get excited about delighting users, not the fancy tech.

Don't be First, Be Fast

Early market leaders have much greater long-term success and enter an average of 13 years after pioneers.

Golder and Tellis (1993)

People make the assumption that if they're not using the latest technology they're not staying competitive, when they are instead just becoming a case-study for the industry.²

The study by Golder and Tellis tells us that first movers have a failure rate that is 6 times higher than fast followers.

If the tech stack mattered, you'd see it in advertising, but you don't, because the customer doesn't care. Your tech stack doesn't determine your product market fit³

Your investors might care (🤷).

If I had to choose between customers and investors I'd pick customers.⁴

The Right Tool for The Job

Ask a graduate what tool or language they would use for their next project and they'll give you a political answer: "I would pick the best tool for the job". Here's a list of programming languages. There are 50 that start with the letter A. Let me know when you've found the one that's best suited for the job.

Perhaps we should define what "best for the job", means. If we are in a team of 12 engineers that are maintaining 2 Postgres databases; introducing a MySQL database on the basis that it addresses a specific use-case better than postgres is not a good idea.

If we hire a DB expert in the future are we going to ensure they're versed in both Postgres and MySQL? It's niche but we could find them at a price.

We don't have to hire this DB expert, our team has built up experience with Postgres, now they'll need to familiarise themselves with the maintenance process of a new db. Just bumping your database version is now slightly tricker and you're monitoring news and bug fixes for both releases.

We could define "best for the job" as perhaps the one we are most familiar with, and this is often the best choice. If you can skip learning a new language or learning the framework step, you're going to get to market faster than others. Don't let learning the tech be the procrastination for learning the market.

In a startup we are testing our market assumptions, this is the largest risk to the company, we shouldn't slow down the speed we get to market. It's better to know we are wrong sooner rather than later.

Getting Experience

Engineers at startups might use the job as an opportunity to focus on CV driven develop instead of focusing on the needs of the business.

I've seen a graduate write every new project in a new programming language.

Anyone know Elixir? We're trying to fix a bug in production and the developer that wrote this is no longer at the company.

From a slack thread.

Starting every new project in a new language isn't a fast way to learn how to deliver. It's a fast-way to get 3 months of experience 5 times. Learn how to deal with a language's pain points and how to solve issues with a specific framework is what makes you valuable. If you're looking to stretch yourself, then use the same language and framework for larger projects.

If you've only got 3 months of experience in a language and framework you're probably not that hard to replace.

The new thing won’t be better, you just aren’t aware of all of the ways it will be terrible yet

Slide 84

Propaganda

Propaganda doesn't just occur in war or between competing states, corporate propaganda is a well established strategy available to large technical institutions.

Let's say you are working on an internal framework that holds up your billion dollar social media platform, you wish to hire more engineers to maintain and build the project. Training new hires on this framework is time consuming, wouldn't it be great if the engineers we hired were already familiar with the framework?

They can also hire developer advocates, these are the influencers of the software engineering world, corporate mercenaric missonaries that are very good at making the work they're doing exciting.

Ignore the hype, your job is to keep things simple.

Bleeding Edge

One reason you might wish to be in the bleeding edge of tech is to gain the expertise and then build around some of the fundamentals you've learnt. Perhaps your customers are others trying to enter the market, this is perhaps a safety-net for the risk you take, but the market can also flatten out and you're left holding a bag of air.

There are a few parts that make up the blog, the static site generator, the API and the repo that hosts all the HTML. This article provides an overview of these pieces and how they're all brought together to create the website.

Framework

The framework is called Pelican a python based static site generator. Each post is written in markdown which starts with meta tags such as Title and Category. Pelican then uses this and a custom theme template to create HTML. The theme uses jinja templating to construct template HTML pages.

Pelican also provides a config file which I have used for feature flags and to enable extensions. The extensions I've enabled include footnotes and code-highlighting. Additionally Pelican hosts a local version of my blog which it recompiles whenever it detects a change to any of the files.

In the config there's a list which contains all the links rendered in the sidebar. Extending or updating these links is only a matter of making changes to this list.

There are two directories inprogress/ and content/. inprogress/ stores drafts that are half complete, completed or minor thoughts for potential blog posts. When these are ready to be published they are moved over to content/ and when these are pushed to the repo, using a github action, pelican automatically renders the entire blog into HTML. The action then creates a commit for any changes and pushes the changes to a public repo.

The public repo contains the entire rendering of the blog and theme. Relying on Github pages, this is then hosted.

The CV

Within the same repo as the drafts and published content there's a CV written in Latex. This latex file can be rendered to a PDF. I have a bash script which allows me to recompile this PDF whenever changes are made to the Latex script.

The API

There's an API that contains edge functions written in Go. These serve requests for the comment sections that sit below each blog post. There are three endpoints, one fetches the count of comments given a list of article URLs. This is used on index pages.

There's an endpoint to fetch comments given an article URL and an endpoint to post a comment to an article URL. The repo containing these endpoints has an action which deploys updates on each new commit.

Finally the API handles the mailing list. It handles registering new emails, verifying the email is valid and lastly triggers an edge function to look for new posts using the blog's RSS feed in order to send updates to recipients in the mailing list.

Software is planning, and as the adage goes, failing to plan is planning to fail. This article goes into the development of Git and UNIX with the goal of dispelling the myth that the genius software engineer exists and when we hear something was written entirely over night there's key information being omitted.

Everyone likes the story of the programmer that disappears and resurfaces after a week with some revolutionising software. We all enjoy a compelling narrative but in reality great software takes time, additionally for most great projects the code is actually the smallest part of the project.

Planning

Projects don't go wrong they start wrong

How Big Things Get Done (2023)

We shouldn't view software creation from the point when coding starts. The process should include the upfront planning. In 1975 Brooke advocated that the coding portion of a software project should amount to 17% of the overall time and it should be noted that in the 1970s the most dominant programming languages were FORTRAN and COBOL.

If a software requirements error is detected and corrected during the plans and requirements phase, its correction is a relatively simple matter of updating the requirements specification.

Software Engineering Economics (1981)

When we think of planning we often think of planning permissions, regulation and endless bureaucracy. We don't have time for that, we are trying to move fast and break things. However, in software, the planning stage is usually when most of the thinking happens and the hard problems get ironed out. Thinking is hard and most people avoid it, make thinking easier and you can stand out from other developers.

The planning phase of a project is also the point in time when large changes are the cheapest.¹

UNIX

One fable within software is that UNIX was developed over the weekend by Ken Thompson. Ken is quite the mythical figure in the software world, Brian Kernighan attests to this in multiple interviews.

We must keep in mind though that Ken worked for three years on Multics which he carried many features forward to UNIX. Here's not saying Ken isn't a great programmer, but we shouldn't discount that his mind was focused on an operating system for quite an amount of time and probably made some contribution towards formulating a better system.

Imagine your best programmer said to you, "sure let me work on the project for two to three years then we will scrap it; and start again. You are bound to have the best in class software." Instead businesses want every project to have a fabled programmer and to continue the trope.

Ken Thompson developed UNIX over 3 weeks, there's a video where Brian Kernighan remarks that modern software engineers aren't as productive².

Git

At this point Git is the most popular form of version control for software, it's also famously known to have been written in a fairly short amount of time. Linus states that he wrote Git in 10 days and we mostly assume the timeline for the project was - I have an idea and ten days later we have Git.

However this is not the case. Linus wrote Git in 10 days but it took him 4 month of thinking about the problem until he had a solution that he was satisfied with. 10 days after that we had Git.

Being a Creator

When it comes to creation, sometimes we are too excited by the answer that we fail to think about the question. A lot of software gets written without getting to the root of the problem and thus fail to materialise any value.

Large projects fail when creation comes first and planning comes never. Even after you have complete knowledge of the problem, solutions may have challenges which require addressing upfront, addressing them when you are midway through a project is going to cost both time and money.

For some problems you can have an attitude of "we will work it out when we get there", but you don't want to have this attitude to all problems as when you get there you might be 95% of the way through your time and budget and you have crossed the point of no return and getting over this hurdle requires going over budget.

Use your planning to categorise problems into "it won't work unless we know this" and "we can figure this out when we get there".

Board Games

"Is it fun" is the most important question in board game design. Ideally you should find this out before you hire an artist, write storyboards or even print and design cards. In board game creation circles they advocate tearing up pieces of paper and scribbling out a prototype to answer this question. The same thing applies to software.

Determine if the question you are answering is the correct one and determine if your software will actual solve the problem, do this; Wizard of Oz style, before putting a shovel into the ground.

Assumptions

All projects start with assumptions. These are typically the first things that need to be addressed before you start coding. Sometimes asking someone if a solution exists or how something works can knock off some unknowns and you'll be in a better position to start than had you remained silent.

Design it Twice

A Philosophy of Software Design (2018)

John Ousterhout advocates that designing software once is not enough and when you are planning something out you should make multiple attempts at it before you move ahead. He's found this leads to better design.

Every project comes with its unknowns and challenges, we are probably never going to rebuild the exact same solution twice. So packing in as much learning as you can upfront about the challenges ahead will leave you at least the most prepared you could have been.

The beginning of project is the cheapest time to learn so we should maximise and front load learning. Learning midway through a project is an expensive way of learning. It's fine if we can afford these learnings but on big and critical projects you don't want to bring up the point that the work we've been doing for the last 6 months was all for nothing especially if we could have learnt this earlier on.

Break things down, explore and figure out how each piece of the project will work and that we are actually providing the right answer to the correct question, this is the only way to deliver on time and on budget.

If we jump straight into code our design strategy is hope.

Are our tests good? Are they bad? Do we have enough tests?

Beyond writing isolated and targeted unit tests, there are methods that ensure our tests are appropriate. This article covers some strategies that answer these questions. Namely; using test coverage metrics, include mutation testing, applying fuzzy tests and finally test fixtures.

Test Coverage

We can measure the number of lines executed when running our tests. For example the following code snippet doesn't include the fourth line 'return False' in the test suite.

def is_odd(n: int) -> bool:
    if n % 2:
        return True
    return False

def test_is_odd():
    assert is_odd(7)

Using the ratio of lines tested to lines not tested gives us the test coverage metric.

Along with context of how important each line of our code is; test coverage is helpful, however as a performance metric or as a blind target it's quite useless.

If you've got a 20% coverage and the code is critical to your business then getting that to 80% is crucial. Trying to eek out the extra 0.1% coverage when you're at 98% is fruitless, or having a goal to go from 98% -> 99% is a poor man's KPI.

Unless they're easy to add you'll be trying to make very minor coverage gains for edge-cases that might be rarely hit. At this point there is likely something more impactful to focus on.

It might be an interesting exercise to understand if the code you're not covering is reachable during the programs lifetime. If not, the dead code should just be removed instead of tested.

Mutation Testing

We rely on tests to ensure our code is correct and works as expected but how do we ensure our tests are correct and they work as expected. What tests our tests? Mutation testing aims to fill this gap.

How often have you written a passing test and you purposely make it fail just to ensure that the test is catching the case you are intending it to catch? This forms the basis of mutation testing.

When applied, it goes through your test suite and makes subtle changes to your tests. If we use the test_is_odd method as an example, it might bump 7 to 8. Sometimes it might take a string and remove a character. It will also change operators in your test such as changing <= to <. The test will then run as normal but under the expectations that it should fail with the mutations. If our example still passes when we pass 8 instead of 7 then something is wrong, the test isn't working as expected which might indicate that we are mocking too many dependencies, we aren't being specific enough or nothing is really being tested.

Fuzzy Tests

Some software may receive user inputted or malformed data, in these cases you might not want the system to behave irregularly. A developer might not know of all the funky data that could be provided to the method upfront, in these cases they may rely on writing fuzzy tests.

As an example, if we had a method that expects a user provided string, we can define a fuzzy test which enumerates a data bank of known edge-cases for strings such as providing an emoji, an empty string or a large string of zero width characters.

Test Fixtures

As the code base grows you might notice that we are writing repeated lines of code in order to setup a user object or prepare data before passing it to the method we are testing.

Large projects get around this by defining test fixtures. These can be passed as parameters to our tests so that we know the setup the test requires before it runs. The benefit of keeping the fixture separate to the test is that it reducing the amount of code that is duplicated across tests and if the setup for the user changes then only the fixture should require changing.

Tests should be focused on asserting one thing and fewer lines in the test makes it easier to see what's going wrong when something breaks.

Finally

If you're printing it, maybe you should assert it.

Further Reading

• Test Coverage
• Mutation Tests
• Fuzzy Tests

Natural language is context dependant and ambiguous. Do you think you can one shot a solid business idea? It took twitch seven years to pivot into gaming. It wasn't seven years of accumulating stacks of code that helped them stick this landing.

We are prompting a machine using an ambiguous and context dependent natural language to create precise and detailed machine instructions. It is no wonder we are finding those with experience with coding are at an advantage when it comes to commanding the machine. The vibe coder is overlooking the techniques and the vocabulary the profession has developed over several decades.

We've learnt in order to generate the best response from an LLM we need more precision and less ambiguity from our prompts. If only we could develop a language that helps us achieve a precise way of creating machine instructions that eliminates ambiguity, perhaps we could call this a programming language?

Fingers crossed

Nothing is built on stone; all is built on sand, but we must build as if the sand were stone.

Jorge Luis Borges (From "Software Engineering at Google")

Most software is built on hope. I write a function that multiplies two integers together and hope that it works. We can also write a test to assert that these numbers will output the correct number, but are you going to write a test for every combination of all numbers?

Once we put lines of code into production the function may or may not be run with the exact input that we expected when we wrote the code.

We are required to create programs without knowledge of the concrete values that will be passed into it; to think of a result in terms of it's name.

double_n = add(n, n)

For every computation we rely on hope.

Program testing can be used to show the presence of bugs, but never to show their absence!

EWD-249 (1970)

Staying Organised

How we ensure our programs are correct also tends to relate to how we scale a project. We've recognised the limitations of a single mind to contain the details of an entire program

It's the core responsibility of a software engineer to watch and manage this complexity.

The art of programming is the art of organizing complexity, of mastering multitude and avoiding its bastard chaos as effectively as possible.

EWD-249 (1970)

Since then we've had multiple attempts at growing a project. There's a link between how we structure code and how we test it. Tests enables us to offload the checking of our functionality and well structured code tends to be easier to test.

This line of thinking led to the practice of Test Driven Development (TDD) where it's thought that writing out the tests as a first step leads the programmer to write more cohesive and well structured code.

Describing Tests

If our tests are determining how we structure the code, what's determining how we structure the tests?

First let's address one of the biggest issues in software engineering. The way we teach and introduce how-to-test is vague and ambiguous, using abstract examples of unrealistic classes and functions. The worse offending term being the "Unit Test" as the definitive boundary for a unit can always be argued.

We have a better understanding of what is not a unit test, than what a unit test is.

The second offender is the testing pyramid. Vehement advocates will disagree on the boundaries of each layer and these layers won't apply to all projects. Setting out to define these at the beginning of a project just waste our time. Often we can only determine where areas of a project will grow with hindsight and we are already building software on a foundation of hope so we should stick to just enough testing.

We shouldn't let the question "Where should we test it" get in the way of testing it.

Managing Tests

We should start thinking more about how we manage tests.

The first thing to address is test duplication. It is all too easy to see a test, make a copy and change it slightly. This can lead us to having the same thing tested across multiple tests. We can reduce the amount of code we are maintaining if we have tests that are targeted. If small changes lead to an unexpected amount of tests breaking we have too much assert duplication.

I compare testing to a climber scaling a mountain with a limited number of pegs. If you are too cautious and nail a peg after every metre you'll find it tougher to make changes when you change direction as the climber's rope is limited by the distance between each peg. Each peg also requires removing when a direction needs to change larger than a metre. However if you nail a peg every 10 metres the climber is flexible to direction changes at the risk of taking a battering when they fall.

Techniques that lead us to balance being defensive and flexible lead us to having a better test suite. Reducing test duplication is one example of this. If we are using 3 pegs in the same location we aren't providing a greater level of safety and we risk requiring unnecessary changes in the future.