Sleep on it

In my experience, the best way to get a high-quality software product is to take your time, not crunch to some deadline. On one project I led, after a couple of months we realised that the feature goals (ALL of them) and the performance goals (also ALL of them, in this case ALL of the iPads back to the first-gen, single core 2010 one) were incompatible.

We talked to the customer, worked out achievable goals with a new timeline, and set to work with a new idea based on what we had found was possible. We went from infrequently running Instruments (the Apple performance analysis tool) to daily, then every merge, then every proposed change. If a change led to a regression, find another change. Customers were disappointed that the thing came out later than they originally thought, but it was still well-received and well-reviewed.

At the last release candidate, we had two known problems. Very infrequently sound would stop playing back, which with Apple’s DTS we isolated to an OS bug. After around 12 hours of uptime (in an iPad app, remember!) there was a chance of a crash, which with DTS we found to be due to calling an animation API in a way consistent with the documentation but inconsistent with the API’s actual expectations. We managed to fix that one before going live on the App Store.

On the other hand, projects I’ve worked on that had crunch times, weekend/evening working, and increased pressure to deliver from management ended up in a worse state. They were still late, but they tended to be even later and lower quality as developers who were under pressure to fix their bugs introduced other bugs by cutting corners. And everybody got upset and angry with everybody else, desperate to find a reason why the late project wasn’t my fault alone.

In one death march project, my first software project which was planned as a three month waterfall and took two years to deliver, we spent a lot of time shovelling bugs in at a faster rate than we were fixing them. On another, an angry product manager demanded that I fix bugs live in a handover meeting to another developer, without access to any test devices, then gave the resulting build to a customer…who of course discovered another, worse bug that had been introduced.

If you want good software, you have to sleep on it, and let everybody else sleep on it.

Posted in process, team | Leave a comment

Episode 35: a bored man with a microphone

I explore the theme of community and the difficulty I have with feeling like a member of a community of technologists. This was motivated by Joy, or Not by Ron Jeffries.

The Descent of Man by Grayson Perry

Society of Research Software Engineering

Thinking Clearly about Corporations by John Sullivan

[objc retain];, the video stream for Objective-C programmers on GNUstep

Leave a comment

My proposal for scaling open source: don’t

I’ve had a number of conversations about what “we” in the “free software community” “need” to do to combat the growth in proprietary, user-hostile and customer-hostile business models like cloud user-generated content hosts, social media platforms, hosted payment platforms, videoconferencing services etc. Questions can often be summarised as “what can we do to get everyone off of Facebook groups”, “how do we get businesses to adopt Jitsi Meet instead of Teams” or “how do we convince everyone that Mattermost is better for community chat than Slack”.

My answer is “we don’t”, which is very different from “we do nothing about those things”. Scaled software platforms introduce all sorts of problems that are only caused by trying to operate the software at scale, and the reason the big Silicon Valley companies are that big is that they have to spend a load of resources just to tread water because they’ve made everything so complex for themselves.

This scale problem has two related effects: firstly the companies are hyper-concerned about “growth” because when you’ve got a billion users, your shareholders want to know where the next hundred million are coming from, not the next twenty. Secondly the companies are overly-focused on lowest common denominator solutions, because Jennifer Miggins from South Shields is a rounding error and anything that’s good enough for Scott Zablowski from Los Angeles will have to be good enough for her too, and the millions of people on the flight path between them.

Growth hacking and lowest common denominator experiences are their problems, so we should avoid making them our problems, too. We already have various tools for enabling growth: the freedom to use the software for any purpose being one of the most powerful. We can go the other way and provide deeply-specific experiences that solve a small collection of problems incredibly well for a small number of people. Then those people become super-committed fans because no other thing works as well for them as our thing, and they tell their small number of friends, who can not only use this great thing but have the freedom to study how the program works, and change it so it does their computing as they wish—or to get someone to change it for them. Thus the snowball turns into an avalanche.

Each of these massive corporations with their non-free platforms that we’re trying to displace started as a small corporation solving a small problem for a small number of people. Facebook was internal to one university. Apple sold 500 computers to a single reseller. Google was a research project for one supervisor. This is a view of the world that’s been heavily skewed by the seemingly ready access to millions of dollars in venture capital for disruptive platforms, but many endeavours don’t have access to that capital and many that do don’t succeed. It is ludicrous to try and compete on the same terms without the same resources, so throw Marc Andreessen’s rulebook away and write a different one.

We get freedom to a billion people a handful at a time. That reddit-killing distributed self-hosted tool you’re building probably won’t kill reddit, sorry. Design for that one farmer’s cooperative in Skåne, and other farmers and other cooperatives will notice. Design for that one town government in Nordrhein-Westfalen, and other towns and other governments will notice. Design for that one biochemistry research group in Brasilia, and other biochemists and other researchers will notice. Make something personal for a dozen people, because that’s the one thing those massive vendors will never do and never even understand that they could do.

Posted in whatevs | 10 Comments

Episode 34: actually it’s about ethics in software engineering

An episode about the philosophy of software engineering.

Leave a comment

There is no “us” in team

I’ve talked before about the non-team team dynamic that is “one person per task”. Where the management and engineers collude to push the organisation beyond a sustainable pace by making sure that at all times, each individual is kept busy and collaboration is minimised.

I talked about the deleterious effect on collaboration, particularly that code review becomes a burden resolved with a quick “LGTM”. People quickly develop specialisations and fiefdoms: oh there’s a CUDA story, better give it to Yevgeny as he worked on the last CUDA story.

The organisation quickly adapts to this balkanisation and optimises for it. Is there a CUDA story in the next sprint? We need something for Yevgeny to do. This is Conway’s Corrolary: the most efficient way to develop software is when the structure matches the org chart.

Basically all forms of collaboration become a slog when there’s no “us” in team. Unfortunately, the contradiction at the heart of this 19th century approach to division of labour is that, when applied to knowledge work, the value to each participant of being in meetings is minimised, while the necessity for each participant to be in a meeting is maximised.

The value is minimised because each person has their personal task within their personal fiefdom to work on. Attending a meeting takes away from the individual productivity that the process is optimising for. Additionally, it increases the likelihood that the meeting content will be mostly irrelevant: why should I want to discuss backend work when Sophie takes all the backend tasks?

The meetings are necessary, though, because nobody owns the whole widget. No-one can see the impact of any workflow change, or dependency adoption, or clean up task, because nobody understands more than a 1/N part of the whole system. Every little thing needs to be run by Sophie and Yevgeny and all the others because no-one is in a position to make a decision without their input.

This might sound radically democratic, and not the sort of thing you’d expect from a business: nobody can make a decision without consulting all the workers! Power to the people! In fact it’s just entirely progress-destroying: nobody can make a decision at all until they’ve got every single person on board, and that’s so much work that a lot of decisions will be defaulted. Nothing changes.

And there’s no way within this paradigm to avoid that. Have fewer meetings, and each individual is happier because they get to maximise progress time spent on their individual tasks. But the work will eventually grind to a halt, as the architecture reflects N different opinions, and the N! different interfaces (which have each fallen into the unowned gaps between the individual contributors) become harder to work with.

Have more meetings, and people will grumble that there are too many meetings. And that Piotr is trying to land-grab from other fiefdoms by pushing for decisions that cross into Sophie’s domain.

The answer is to reconstitute the team – preferably along self-organising principles – into a cybernetic organism that makes use of its constituent individuals as they can best be applied, but in pursuit of the team’s goals, not N individual goals. This means radical democracy for some issues, (agreed) tyranny for others, and collective ignorance of yet others.

It means in some cases giving anyone the autonomy to make some choices, but giving someone with more expertise the autonomy to override those choices. In some cases, all decisions get made locally, in others, they must be run past an agreed arbiter. In some cases, having one task per team, or even no tasks per team if the team needs to do something more important before it can take on another task.

Posted in agile, team | 1 Comment

Episode 33: Ask Me Anything

@sharplet asked me a few questions, and you can too!

The theme music is Blue-Eyed Stranger.

Evidence-Based Software Engineering Using R.

Median household income in the UK in January 2021 was £29,900 (I said “about £28,000”). Median software engineer salary (i.e. for the role “software engineer” with no seniority modifiers) is £37,733.

Leave a comment

It was requested on twitter that I start answering community questions on the podcast. I’ve got a few to get the ball rolling, but what would you like to ask? Comment here, or reach me wherever you know I hang out.

Posted on by Graham | Leave a comment

An Imagined History of Object-Oriented Programming

Having looked at hopefully modern views on Object-Oriented analysis and design, it’s time to look at what happened to Object-Oriented Programming. This is an opinionated, ideologically-motivated history, that in no way reflects reality: a real history of OOP would require time and skills that I lack, and would undoubtedly be almost as inaccurate. But in telling this version we get to learn more about the subject, and almost certainly more about the author too.
They always say that history is written by the victors, and it’s hard to argue that OOP was anything other than victorious. When people explain about how they prefer to write functional programs because it helps them to “reason about” code, all the reasoning that was done about the code on which they depend was done in terms of objects. The ramda or lodash or Elm-using functional programmer writes functions in JavaScript on an engine written in C++. Swift Combine uses a functional pipeline to glue UIKit objects to Core Data objects, all in an Objective-C IDE and – again – a C++ compiler.

Maybe there’s something in that. Maybe the functional thing works well if you’re transforming data from one system to another, and our current phase of VC-backed disruption needs that. Perhaps we’re at the expansion phase, applying the existing systems to broader domains, and a later consolidation or contraction phase will demand yet another paradigm.
Anyway, Object-Oriented Programming famously (and incorrectly, remember) grew out of the first phase of functional programming: the one that arose when it wasn’t even clear whether computers existed, or if they did whether they could be made fast enough or complex enough to evaluate a function. Smalltalk may have borrowed a few ideas from Simula, but it spoke with a distinct Lisp.

We’ll fast forward through that interesting bit of history when all the research happened, to that boring bit where all the development happened. The Xerox team famously diluted the purity of their vision in the hot-air balloon issue of Byte magazine, and a whole complex of consultants, trainers and methodologists jumped in to turn a system learnable by children into one that couldn’t be mastered by professional programmers.
Actually, that’s not fair: the system already couldn’t be mastered by professional programmers, a breed famous for assuming that they are correct and that any evidence to the contrary is flawed. It was designed to be learnable by children, not by those who think they already know better.

The result was the ramda/lodash/Elm/Clojure/F# of OOP: tools that let you tell your local user group that you’ve adopted this whole Objects thing without, y’know, having to do that. Languages called Object-*, Object*, Objective-*, O* added keywords like “class” to existing programming languages so you could carry on writing software as you already had been, but maybe change the word you used to declare modules.

Eventually, the jig was up, and people cottoned on to the observation that Object-Intercal is just Intercal no matter how you spell come.from(). So the next step was to change the naming scheme to make it a little more opaque. C++ is just C with Classes. So is Java, so is C#. Visual BASIC.net is little better.

Meanwhile, some people who had been using Smalltalk and getting on well with fast development of prototypes that they could edit while running into a deployable system had an idea. Why not tell everyone else how great it is to develop prototypes fast and edit them while running into the deployable system? The full story of that will have to wait for the Imagined History of Agile, but the TL;DR is that whatever they said, everybody heard “carry on doing what we’re already doing but plus Jira”.

Well, that’s what they heard about the practices. What they heard about the principles was “principles? We don’t need no stinking principles, that sounds like Big Thinking Up Front urgh” so decided to stop thinking about anything as long as the next two weeks of work would be paid for. Yes, iterative, incremental programming introduced the idea of a project the same length as the gap between pay checks, thus paving the way for fire and rehire.

And thus we arrive at the ideological void of today’s computering. A phase in which what you don’t do is more important than what you do: #NoEstimates, #NoSQL, #NoProject, #NoManagers…#NoAdvances.

Something will fill that void. It won’t be soon. Functional programming is a loose collection of academic ideas with negation principles – #NoSideEffects and #NoMutableState – but doesn’t encourage anything new. As I said earlier, it may be that we don’t need anything new at the moment: there’s enough money in applying the old things to new businesses and funnelling more money to the cloud providers.

But presumably that will end soon. The promised and perpetually interrupted parallel computing paradigm we were guaranteed upon the death of Moore’s Law in the (1990s, 2000s, 2010s) will eventually meet the observation that every object larger than a grain of rice has a decent ARM CPU in, leading to a revolution in distributed consensus computing. Watch out for the blockchain folks saying that means they were right all along: in a very specific way, they were.

Or maybe the impressive capability but limited applicability of AI will meet the limited capability but impressive applicability of intentional programming in some hybrid paradigm. Or maybe if we wait long enough, quantum computing will bring both its new ideas and some reason to use them.

But that’s the imagined future of programming, and we’re not in that article yet.

Posted in OOP | Leave a comment

A hopefully modern description of Object-Oriented Design

We left off in the last post with an idea of how Object-Oriented Analysis works: if you’re thinking that it used around a thousand words to depict the idea “turn nouns from the problem domain into objects and verbs into methods” then you’re right, and the only reason to go into so much detail is that the idea still seems to confuse people to this day.

Similarly, Object-Oriented Design – refining the overall problem description found during analysis into a network of objects that can simulate the problem and provide solutions – can be glibly summed up in a single sentence. Treat any object uncovered in the analysis as a whole, standalone computer program (this is called encapsulation), and do the object-oriented analysis stuff again at this level of abstraction.

You could treat it as turtles all the way down, but just as Physics becomes quantised once you zoom in far enough, the things you need an object to do become small handfuls of machine instructions and there’s no point going any further. Once again, the simulation has become the deployment: this time because the small standalone computer program you’re pretending is the heart of any object is a small standalone computer program.

I mean, that’s literally it. Just as the behaviour of the overall system could be specified in a script and used to test the implementation, so the behaviour of these units can be specified in a script and used to test the implementation. Indeed some people only do this at the unit level, even though the unit level is identical to the levels above and below it.

Though up and down are not the only directions we can move in, and it sometimes makes more sense to think about in and out. Given our idea of a User who can put things in a Cart, we might ask questions like “how does a User remember what they’ve bought in the past” to move in towards a Ledger or PurchaseHistory, from where we move out (of our problem) into the realm of data storage and retrieval.

Or we can move out directly from the User, asking “how do we show our real User out there in the world the activities of this simulated User” and again leave our simulation behind to enter the realm of the user interface or network API. In each case, we find a need to adapt from our simulation of our problem to someone’s (probably not ours, in 2021) simulation of what any problem in data storage or user interfaces is; this idea sits behind Cockburn’s Ports and Adapters.

Moving in either direction, we are likely to encounter problems that have been solved before. The trick is knowing that they have been solved before, which means being able to identify the commonalities between what we’re trying to achieve and previous solutions, which may be solutions to entirely different problems but which nonetheless have a common shape.

The trick object-oriented designers came up with to address this discovery is the Pattern Language (OK, it was architect Christopher Alexander’s idea: great artists steal and all that), in which shared solutions are given common names and descriptions so that you can explore whether your unique problem can be cast in terms of this shared description. In practise, the idea of a pattern language has fared incredibly well in software development: whenever someone says “we use container orchestration” or “my user interface is a functional reactive program” or “we deploy microservices brokered by a message queue” they are relying on the success of the patterns language idea introduced by object-oriented designers.

Meanwhile, in theory, the concept of patterns language failed, and if you ask a random programmer about design patterns they will list Singleton and maybe a couple of other 1990s-era implementations patterns before telling you that they don’t use patterns.

And that, pretty much, is all they wrote. You can ask your questions about what each object needs to do spatially (“who else does this object need to talk to in order to answer this question?”), temporally (“what will this object do when it has received this message?”), or physically (“what executable running on what computer will this object be stored in?”). But really, we’re done, at least for OOD.

Because the remaining thing to do (which isn’t to say the last thing in a sequence, just the last thing we have yet to talk about) is to build the methods that respond to those messages and pass those tests, and that finally is Object-Oriented Programming. If we start with OOP then we lose out, because we try to build software without an idea of what our software is trying to be. If we finish with OOP then we lose out, because we designed our software without using knowledge of what that software would turn into.

Posted in ooa/d | Leave a comment

Episode 32: freeing software from source code

I muse on a concept I’ve been thinking about for a long time: that software engineering is trapped within the confines of the fixed-width text editing paradigm. This was motivated by a discussion with Orta about his Shiki Twoslash project but it’s been rattling around in here for years.

I also talk about Pharo, the Lindisfarne Gospels, Code Bubbles, Scratch, RStudio and more!

Leave a comment