Dev meeting – Data intensive applications; and Rust

At our dev meeting this week, Rich talked about ‘Designing Data Intensive Applications‘, that he’d been reading, and Inigo talked about ‘Talking to your TV using Rust‘.

‘Designing Data Intensive Applications’, from O’Reilly, describes various tools and approaches for storing and processing data. It’s split into four sections. Section 1 discusses the basics of databases and how they store content, for SQL databases, NoSQL databases, flat file stores. It covers in-memory stands, b-trees, indexing structures, and the best approaches to use in various situations. Section 2 talks about distributed databases – clustering, transactions, serializing commits at a point in time so shared logs can work, and so on. Section 3 discusses distributed batch processing, using tools like Spark. The first three sections were all very interesting and informative. The fourth section was not as good – it covered the future of data, and was a lot woollier than the previous sections. Rich found the book on the whole good.

To talk to his TV using Rust, Inigo used a Google Home Mini, an RM 3 Mini IR blaster, an Open Source library to interface with the IR blaster, and some Rust on a Raspberry Pi. The Google Home part was setting up a Google Home App via – you configure a set of actions and intents online, and then assign a ‘fulfillment’ to it, which is a web service you make available via HTTPS. This part is fairly straightforwards – it requires filling in various web forms, and it’s easy to publish a test version of an app that’s only available via your own Google account.

To actually use this for a home application, you need to have SSL, so you need a domain name and a certificate. Inigo used Amazon Route 53 for the domain name, and Lets Encrypt for the certificate. There are various other services that will provide a domain name for a dynamic IP address for free, although the original dyndns is no longer free.

THe ‘fulfillment’ part is a web service that Inigo wrote that accepts JSON from Google, and then calls out to a Python library that interacts with the IR blaster. He wrote the first version in Scala, to get something simple working fast, and then rewrote it in Rust. He learned from O’Reilly’s Programming Rust book, and used IDEA’s Rust plugin. The experience was positive, although it was a bit tricky to get some of the memory management working correctly as a first-timer using Rust. The code is on Github at

Dev meeting – Effective Java (again)

Stephen recently read Effective Java, by Josh Bloch, and we discussed things that he picked up from it, particularly approaching Java from a perspective of much more experience in C# than in Java.

  • When you have multiple overloaded constructors, instead use a static factory method on the class. We discussed the advantages and disadvantages of this.
  • We discussed the preference for interfaces rather than concrete classes in Java vs C#. We discussed how this was probably influenced by type inference and type erasure. It may also be related to Java’s history with frameworks like Hibernate and Spring, that initially used lots of interfaces to support proxies before bytecode rewriting became common.
  • Overriding equals and hashcode – it’s common in Java, but less common in C#. In Scala, case classes remove a lot of this complexity. We discussed what identity means for classes and objects, and how you express that in various languages.
  • The difference between Exceptions and RuntimeExceptions is surprising from a C# background. We discussed the merits of checked exceptions – we broadly agreed that in an ideal world, checked exceptions were a good idea, but in practice it depends on library authors being competent and library users being competent. Sadly, this doesn’t happen. Checked exceptions also cause issues for lambda functions. There was some discussion that the way Scala does this is better.
  • The use of external libraries – it’s a lot more common in Java, whereas .NET developers tend to use the CLR content much more frequently. We discussed the issues of transitive dependencies in Java bulking out your build. We also discussed how the ‘small number of large libraries’ assumption is no longer true in .NET Core – there are many different packages within .NET Core.

We discussed Effective Java a while ago in a previous dev meeting, too.


Dev meeting – AIs, interplanetary file systems and Nix

We had several lighting talks in the dev meeting this week.

Rich talked about WitAI – a language parsing web service that helps with building chatbots. He’s written a “drinksbot” in it, which we used via Slack before the dev meeting to order drinks. You provide it with a corpus of example sentences, and it learns from that.

Chris spoke on the Interplanetary file system. In typical direct file request, you might retrieve it synchronously from a single location. IPFS allows you to retrieve content from a set of distributed stores – essentially a peer-to-peer, torrent-like filesystem, that can be more robust and potentially faster than a direct file retrieval; because content is stored can be stored in a number of places.

Reece talked about antagonistic attacks on neural networks. Because neural networks (e.g. for image recognition) will typically work by perturb the input image to make the neural network recognize it as something entirely different – even though the change to the human might be imperceptible. For example – researchers recently made a model of a turtle (to human eyes) look like a gun (to neural network recognition) to illustrate this issue. This is a problem with neural networks generally – that the neural network may be using very specific things for its recognition, and it may be hard to identify these.

Stephen talked about teaching (very) young developers – like his daughter – to program. There is a lot of implicit knowledge and setup and background in a typical dev environment, and you want to get to working code as quickly as possible. HTML 5 and JavaScript are good for this. You should optimize for small victories, and concentrate on fast feedback loops.

Tim talked about Vue, the JS framework. You create reusable components that have an HTML component and some connected JavaScript that can control . It’s somewhere in the middle in terms of ease of use – it’s more complicated and powerful than Knockout, but simpler and more usable than React and Angular (and it can be used initially just by including script tags, and then extended to manage things like transpilation). We’re using it in a number of projects at present.

Rodney talked about “Where I’ve been” – specifically about the Nix conference that he recently attended. There is a new Nix tool for handling packaging and building in the Nix system, which has a simpler (command-line) UI. Another talk was about the security developments to the Nix project, such as having specific security experts who can getting security notifications from other projects early on; about automated scanning tools for scanning deployed Nix implementations for security holes. Nix in production was being discussed – e.g. for a company that’s doing cycle dock management in the Netherlands. Using Nix throughout an entire environment, from development through to production, brings many advantages. At Tumblr, Nix is used for testing of live SQL instance replication. Then there was a Hackfest – with spontaneous collaboration, and good work done on Nix cross-compilation.

Tell me more about talking about chatbots

In our dev meeting this week, Alex talked about chatbots. Bots are a massive shift in how we interact with computers – but to some extent they’re still a solution looking for a problem. Bots are conversation driven – and developers typically have never dealt with conversation! There’s a lot of existing support for English language bots that comprehend language reasonably – but there’s far less support for non-English bots. There are libraries and services to help with this – for example, Microsoft Cognitive Services provides endpoints that pick out from the users input what the user is asking, and even what their emotions are. However, these tools are probabilistic – they will say “it’s 90% likely that the user wants this”, rather than “the user definitely wants this”. Bots can get confused – because the bot will typically maintain state so it can refer back to earlier elements of the conversation, it may sometimes lose track of where it is – and needs to be reset. However – this is adding complexity to what is supposedly an intuitive interface; and users may not understand the underlying complexity.

You need to think about the platform that you’re deploying to, in terms of how you develop your bot and what capabilities you expect it to have. Slack bots, bots responding to SMS, Facebook bots – all have different characteristics based on the capabilities of the platform they’re running against. This also shapes what users expect in terms of level of communication – no-one expects lengthy complex sentences from an SMS bot, but they might do from a Slack bot. Discord bots have an expectation that they’re used for gaming, because of the assumptions of the Discord culture. Facebook bots have more in the way of graphics and buttons – so they’re moving to “Monkey Island” style user interfaces, whereas other bots are more similar to an old-style text adventure user interface or a Twine (and Twine is good for developing bots).

For bots, uptime is critical – so the bot should always be available. However, it’s not essential that they respond immediately at machine speeds – in some ways its better to pause a few seconds, because it may feel more natural, as well as helping deal with load. Cloud hosting is useful for bot development to help with this availability and coping with load. Also tracking user responses is important.

Given these complexities – why bots? It offers a high engagement rate, because it’s the same way of communicating as people are used to using with their friends. Bots can be proactive – if a user looks up an event like a concert, then the bot can prompt that the user might want to buy a ticket for it. Bots can be more engaging for users who are unfamiliar with technology. Bots are new and exciting – it’s immature and in flux, and there are no great settled ways of doing things, so it’s good for experimentation. There’s lots of space for exploration – like creating a Twitter bot that offers to sell wine to people based on their tweets of wine bottles. They are a new space – and we should think about ‘how can a bot benefit you’ – like using a bot to collect orders to buy drinks at the bar. Go forth, and make the next Flappy Bird of bots!

Elm, ScalaJS and GHCJS – a “better JavaScript” shootout

At our dev meeting this week, we talked about three different alternatives to JavaScript – Elm, ScalaJS and GHCJS. All three allow you to write code in a modern, statically-typed functional language, and then produce output in JavaScript that can be run in a browser.

Chris had previously talked about Elm a few weeks ago, in “Elm and Punishment”. Elm is a Haskell-inspired language, designed specifically as a web language that compiles to JavaScript, HTML and CSS. Chris has written a sliding blocks puzzle (source) in Elm as a learning exercise.

Inigo was inspired by Chris’s work in Elm to look at ScalaJS. ScalaJS is a plugin for Scala/SBT that allows standard Scala code to be compiled to JavaScript. It covers the vast majority of the Scala standard library (although it excludes a few areas like reflection and reading files from disk). It doesn’t convert JVM bytecode, only Scala sourcecode, so Java code first needs to be converted to Scala, and it’s not possible to immediately use standard Java libraries. Inigo also wrote a sliding blocks puzzle (sourcecode) and a minesweeper clone (source).

Rodney took a different approach – he implemented Flattris – a flat version of Tetris (source). Rather than using Elm, which is inspired by Haskell, he used GHC JS – which actually is Haskell (the Glasgow Haskell Compiler’s JavaScript compiler).

All three of the languages turned out to be very similar in approach and in results. All of them worked very well. The core languages (Elm, Scala and Haskell) are all more powerful languages than JavaScript. The generated JavaScript is bulkier than hand-written JavaScript would be (from 150kb for ScalaJS, up to about 1MB for GHC JS), but this still isn’t particularly huge compared to the size of modern web pages. All three implementations used a pure functional approach (this is standard in Elm and Haskell, and recommended in Scala), and the sourcecode for the two different sliding blocks implementations ended up remarkably similar. ScalaJS and Elm both emit sourcemaps that allow the generated JavaScript to be debugged e.g. by stepping through the Scala source directly in the browser, or via IDEA. All three were enjoyable to program.

We’re not planning to embrace any of the three as a core language we use at 67 Bricks yet, but we were pleasantly surprised by how well they all worked, and we’re certainly considering them for the future.



Effective code reviews

At our last dev meeting, we discussed code reviews. We do code reviews for all code on all projects, before code is merged to the master branch, and everyone is familiar with doing them, but we can always improve how we do things. So, we had a general discussion on what good code reviews are, and how we can do them best.

We initially discussed what one should look for in a code review. We agreed on some questions a code reviewer should ask themselves:

  • Do you understand the change being made, and more importantly why it’s being made?
  • Is the change itself good? Is it clear, are things well-named? Is it testable; and does it have tests? Does it introduce any non-functional issues such as security or performance problems?
  • One should look at what’s been done, but also the context of the change. Is there code elsewhere that should be reused? Are there similar changes to make elsewhere? Is there now obsolete code that can be removed?
  • Even if the change itself is fine, has the codebase as a whole become less clearer? Has a file or method become too big?

We talked about some of the benefits we have had from good code reviews, such as:

  • Finding out better ways of doing things – such as being able to change a large chunk of procedural code into much more concise and clear functional code, and new ways of doing legacy things
  • Improving consistency throughout the system, using common approaches such as use of optionals, and putting code in the correct classes, clustering similar methods close to each other
  • Getting the benefit of other peoples’ wider experience – e.g. new people have new perspectives and approaches that they have from previous companies, that we can learn from
  • Sometimes it makes sense for the reviewer to make changes and improvements to the code, rather than just handing it straight back to the original developer
  • Identifying areas that are unclear or confusing in the existing system as a whole, and that need refactoring
  • Helping standardize our practices within and across teams, so we can improve generally

As a reviewer, you are aiming to do a good code review, but as a coder, you also have a responsibility to help make the code review useful as well. These include:

  • Asking for a design review before starting the task at all
  • Making useful, atomic commits, with meaningful commit messages, rather than a big bulk commit of everything followed by a few “fixing this” commits
  • Making tests clear and readable, and avoiding repetition in them, just as we aim to make the main code clear and readable
  • Review your code yourself first, before you commit – stepping back from the moment-by-moment coding often helps you spot issues. This incudes the basics – check that the code compiles and relevant tests pass
  • Asking for partial review of long tasks, rather than waiting until the whole task is complete. Frequent review and merge makes the process simpler
  • For a long branch, being ready to fix issues on that branch quickly if a reviewer find any, e.g. by reviewing other peoples code or by doing short tasks, rather than jumping into another long task

Ultimately, it is the reviewer that is doing the review, and we discussed the ways that a reviewer can improve the quality of their reviews:

  • Consider whether you are the right person to review the code – usually it’s good to spread reviews around the team, but sometimes you might want an expert in a particular area to review the code
  • Being able to start the review quickly – so prioritizing code reviewing over starting on new tasks
  • Providing all the feedback from a review at once, rather than in dribs and drabs
  • Pointing out what’s good, as well as what’s wrong in the code – this is good for morale, and also encourages good practices
  • Providing a rationale for feedback – never just “do this”, but “this is a better way because…”
  • There can be multiple issues with a piece of code – sometimes you’ll be initially distracted by minor points, but might need to revisit the review once those are resolved to spot fundamental problems of approach
  • Pointing out alternative approaches that might be useful in future, even though a change isn’t needed in the current review
  • Avoid giving feedback that is impossible to apply – as a reviewer, think about how you might solve a problem yourself, and discuss the issues with the reviewee

And after the review, for all the items raised, each should be discussed and either resolved, or agreed to be not needed now, or a ticket raised for them to be fixed soon.


Sorting, and impatience – our dev meeting for 16th June

Chris talked about things that he has learned from the new 3rd edition of the “Scala for the Impatient” book… that we received several months after Amazon had initially told us it would be delivered. He was impatient.

Chris discussed:

  • String interpolation – while he knew about the s”” format for interpolating variables, he learned about the f”” prefix to do printf style interpolation, and that you can define your own string interpolation operators
  • Scala ARM – provides resource management in a similar way to the try/resources construct in Java
  • Renaming classes on import – while you might normally import individual files under different names, for example to deal with the clash between “Files” in Guava, and “Files” in Java, you can also import the whole of a package but block out an import with -> _ to avoid it conflicting.
  • Values classes – you can use the AnyVal class when you’re wrapping a simple type – like wrapping an integer inside a class to make it strongly typed – and the compiler will not create a class
  • Stream creation via functions – you can use the the #:: operator on streams to create a lazily evaluated stream based on a function. This is useful for all the times when you need to, for example, create an infinite sequence of Fibonacci numbers.
  • The @Elidable annotation allows you to strip out methods at compile-time at particular levels – which potentially allows you to remove things like calls to logging in a production environment.

Then, Rodney talked about generic sorting and discrimination – the problem of sorting a list of things, such as a pack of cards. He talked about recent research on how to improve the traditional ways of sorting. Normally you might implement something like “Comparable” or “Equals” on your “Card” class – but that only gives you one way of ordering, whereas you might want to group your cards by number or by suit. Using just an equals method means that finding unique values in a set is hard – it takes O(N^2) comparisons, as you compare every item against every other item.

To improve this, “Discrimination” generalizes partitioning and sorting, and defines a language for doing this.

Rodney demonstrated sorting a pack of cards, using the bucket sort. If you have 52 buckets, then you can put each card in the right place on a table in linear time. If you have 13 buckets and don’t care about suit, you can sort the cards into buckets purely by value, as a linear operation. If you’re sorting an arbitrarily large list of integers, you just need a potentially infinite amount of buckets, stored in a potentially infinite amount of memory, and you can get linear time sorting – no problem…

The problem comes if you don’t have enough memory for all the buckets you want. However, if you have 2^32 numbers to sort, then you can split each into a pair of words, and then sort them all initially by the high word and then by the low word – using multiple bucket sorts where you only need 2^16 buckets not 2^32.

The language of ordering is a DSL that defines various ways of ordering – natural ordering puts things in their natural order; product ordering sorts by one feature such as the suit then another such as the number; sum ordering sorts by type such as whether the card is a joker or not.

You can use this generically by defining a discriminator for your type, such as a card, in terms of the language of ordering, and then you can apply the rules of sorting those terms to split the ordering into an appropriate sequence of buckets for your type, and this can be significantly more efficient than using a single bucket sort.

In conclusion, using a DSL is an effective technique for simplifying and solving problems, and we should use them more widely.

The price of Some(everything), the value of None

This week, Loic talked about how to measure the value of software, and Reece talked about category theory.

Loic based his talk on an academic study of software value from Gio Wiederhold, published in ACM. He talked about how the value of software changes over time. Typically, revenue from a software product will rise until it reaches a peak, but the price of software is usually not expected to increase significantly over time. Maintenance costs, particular for keeping old versions available, can rise over time. In practice, 5% of code is deleted per year, and a new version of a product consists of less than 30% more code – with most effort spent on keeping everything running smoothly. Loic discussed rational design decisions about where to focus the effort in software development, and making sure there is sufficient investment in maintenance.

Everyone else was generally interested in the topic, but found the research it was based on somewhat dated. It’s very much based on the “large software company producing a big product” model, like sales of MS Office. However, the current market is much more about web delivery of systems that can be updated very frequently in small ways, and about apps, and about Open Source. We discussed how these sales models impacted software value.

Then, Reece talked about category theory, and its relevance to software development – in a talk called “Monads are not (always) Optional”. He talked about the value of using standard mathematical terminology for the concepts of programming, and defined “category”, “object”, and “morphism”. Then we discussed composition of morphisms, and how classes of numbers (e.g. integers, rational numbers, natural numbers, complex numbers) could be transformed into each other via various operations, but other operations (such as addition and multiplication) changed the values of numbers but kept them within the same category. He discussed the relevance to programming, and how category theory and its concepts provided a more unified mechanism for structuring solutions than other approaches such as design patterns.

We then discussed monads, and monads in C# via Linq and in Scala via sequences and optional. We discussed that the “Advanced Scala with Cats” had recently been made available for free, and several people planned to read it. No-one was clear of the difference between Scalaz and Cats – both seem to cover similar ground (some subsequent research seems to show that they’re similar, and Cats might generally be preferred).

EMACS org mode, tea and sprints – dev meeting for 5th May

Rodney discussed EMACS org mode – after some discussion of the relative merits of EMACS vs Vim vs inferior other editors. It essentially allows you to edit bullet points, but it can be used for calendars, todo lists, estimates, and so on. Via a keypress, you can add templates for a todo item, including schedules and other metadata, and include it in your todo list. This metadata is used in the display – e.g. you can display an agenda for the day, which will include all of the tasks scheduled for today. You can mark those tasks as done once complete. Tasks can also be split into a hierarchy of tasks, and the top task can display a percentage of the remaining tasks under it. It also has time tracking – once you start a task, you can start time tracking, and then it will be logged against that task.

Advantages of it are – it’s simple, it’s fast, it’s free, and it’s text based so other tools will work with it.

To achieve similar results, Chris uses todo.txt – which is essentially a text file in Dropbox, but it’s supplemented by command-line tools, and a phone app. Because it’s backed by a text file, then you can always access your data anywhere without any other tools. Reece uses a text file in Vim. Inigo used to use a text file in Vim, but now uses Trello.

We then discussed tea making. There was no consensus reached.

Stephen then talked about sprint planning. In various of our projects, we are doing sprint planning in different ways. The default Agile approach is to start with a demo of the work done in the previous sprint, then a retrospective of the previous sprint, then the project owner provides the things that they would like to be done in the sprint, and the project team discusses which ones can be done in the sprint. In one of our projects, we have a looser sprint planning, that doesn’t involve prior discussion of the statistics from previous sprints, and not much backlog grooming. We discussed the merits of estimation – even though the results of the estimation are not always useful, the process of doing estimation and giving developers a good idea of how long tasks are expected to take is useful. We also discussed prioritization, and how new tasks should be dealt with – whether they should be completely rejected until the end of the sprint, or whether they should be prioritized against exising tasks. We agreed that we would look at more explicit prioritization in some projects. We also agreed to combine some brief explicit estimation of tasks in the sprint planning meetings, with announcing on Slack a brief description of the task and how we would approach it and how long it would take.‰

Elm, and Punishment

Chris talked about Elm, which he subtitled “JavaScript for people who like Haskell not JavaScript”.

However, he immediately explained that Elm is not like Haskell, and isn’t intended to be compared to Haskell, and that it is aimed at JavaScript developers not Haskell developers, and avoids the community avoids using Haskell terminology. This is intended to help increase the uptake of the language.

However, it is a pure functional language, and does look quite Haskell-ish: for example, it supports pattern matching, immutability, and so on. It is event-driven.

One thing it concentrates on is having good error messages. It will even do things like identify references to variables that don’t exist, and suggest what you might have intended. It’s also helpful that there’s an Elm plugin for IntelliJ IDEA.

Like most modern JavaScript frameworks, it uses a virtual DOM, to make it faster.

There is a time-travelling debugger, which is a very cool feature, but Chris didn’t use it in practice.

It has automatic semantic versioning, so it updates the major version automatically when existing signatures change. However, the Elm language itself has changed significantly in minor versions, so you can’t depend on code for older Elm versions being compatible. This was frustrating for working in the language – finding older libraries that hadn’t been updated, and older examples that were out of date.

You can incrementally rewrite an existing JavaScript codebase to Elm, on a file-by-file basis. Chris wasn’t sure how compatible Elm output is with older browsers.

There are currently no books (although there will be later this year). There is a “try elm” site that is good.

It’s primarily aimed at interactive UIs, and games.

Chris’s sample Elm code is at However, he hasn’t actually completed the game, because it’s too hard.

Then, Loic talked about Punishment Driven Development. He talked about reasons why companies punish people, and the effects of that punishment. He described the importance of understanding why people are behaving as they do, and that sometimes you may need to change your own behaviour to work with them in order to achieve your objectives. Then, he talked about the axioms of Punishment Driven Development, and contrasted them with People Driven Development.