Roben Kleene

Codespaces: GitHub's Play for a Remote Development Future

Codespaces Logo

When I first saw Codespaces, I immediately wanted it. With ubiquitous high-speed internet, why not offload more work to the cloud? What could our devices look like if most of their power came from the server? What would their battery life be like?

Seamlessly leveraging remote resources has always felt like an idea that’s just around the corner, but never arrives. Just having a big beefy machine on site usually ends up being the most practical solution (outside of some specialized use cases)1.

Codespaces is perhaps the biggest play ever to take remote development more mainstream. Development has always been a prime candidate for remote computing, because with time-sharing machines, it’s how the roots of programming itself began2.

Visual Studio Online to GitHub Codespaces

GitHub Codespaces began as a different product, called Visual Studio Online. Visual Studio Online was announced on the Visual Studio Blog in November 2019. Then, in April 2020, it was renamed to Visual Studio Codespaces, Nik Molnar described the motivation behind the name change on the same blog:

We learned that developers are finding Visual Studio Online to be much more than just an “editor in the browser”. They are saying that “the capabilities of this cloud-hosted dev environment make it the space where I want to write all my code“.

To better align with that sentiment, and the true value of the service, we’re renaming Visual Studio Online to Visual Studio Codespaces.

A few days later, a corresponding announcement appeared on the GitHub blog that Codespaces was coming to GitHub. Then, almost a year later in September 2020, it was announced on the Visual Studio Blog that Visual Studio Codespaces would be consolidated into GitHub Codespaces, and that Visual Studio Codespaces would be retired in February 2021.

Visual Studio Codespaces was similar to GitHub Codespaces, but it did have some key differences. Visual Studio Codespaces wore more of its implementation details on its sleeve, in particular, as being built on top of Microsoft Azure. When you setup a Visual Studio Codespaces, it was linked to an Azure subscription and location, and you chose a “Default Instance Type” for new codespaces3.

Create Environment

The decision to remove these details from GitHub Codespaces, and provide quick access to launch a codespace from a repository, was highlighted in the announcement letter about shutting down Visual Studio Codespaces in favor of GitHub Codespaces:

During the preview we’ve learned that transitioning from a repository to a codespace is the most critical piece of your workflow and the vast majority of you preferred a richly integrated, native, one-click experience.

This is a great example of iterative product design. From a practical perspective, Visual Studio Codespaces is essentially the same product as GitHub Codespaces (and GitHub Codespaces is presumably also running on Azure), but hiding the virtual machine implementation details makes GitHub Codespaces feel different, and a bit more revolutionary4.

Tour

Once you’re in the Codespaces beta, a “Codespaces” item appears in the navigation menu when you click your user icon in the upper right5. Click it, and you’re brought to a screen where you can manage the Codespaces you’ve already created, including removing them by clicking “Delete” under the three disclosure dots.

Codespaces in Navigation

Every repository also has an “Open with Codespaces” option, which can either create a new Codespace or open an existing one for that repository.

Open With Codespaces

After opening a codespace, you’re brought to a browser window running Visual Studio Code. It works similarly enough to the desktop version that it’s practically indistinguishable6.

Codespaces Running

Alternatively, you can connect to the codespace directly from the desktop version of VS Code by using the Visual Studio Codespaces extension. The extension adds a “Remote Explorer” icon to the Activity Bar where you can connect to, and manage, your codespaces.

Codespaces Running in VS Code

The About Codespaces section of the documentation explains a couple of details about the relationship between codespaces are repositories:

Each codespace is associated with a specific branch of a repository. You can create more than one codespace per repository or even per branch. However, each user account has a two-codespace limit during limited public beta.

Implementation Details

Codespaces uses Docker containers to setup development environments. GitHub and Microsoft calls a running codespace a “development container” presumably after Docker containers, emphasizing their close relationship.

Regarding what’s running locally, and what’s running in the development container, the Remote Development FAQ describes how the user-interface runs locally, i.e., in the browser or VS Code app, while a separate process running on the server (“VS Code Server”) handles the operations that need to happen on the server, such as file system access:

The VS Code Server is quickly installed by VS Code when you connect to a remote endpoint and can host extensions that interact directly with the remote workspace, machine, and file system.

The FAQ also includes this handy diagram illustrating what’s running on the server and what’s running locally:

Architecture

Whether extensions runs locally or on the development container depends on whether they “contribute to the VS Code user interface”. If they do, they’re called “UI Extensions” and run locally, if they don’t, they’re called “Workspace Extensions” and run on the server.

Whether extensions are UI Extensions or not, they’re all installed on the development container at the path ~/.vscode-remote/extensions/:

% ls ~/.vscode-remote/extensions/
castwide.solargraph-0.21.1
davidanson.vscode-markdownlint-0.37.0
dbaeumer.vscode-eslint-2.1.8
dbankier.vscode-quick-select-0.2.9
eamodio.gitlens-10.2.2
editorconfig.editorconfig-0.15.1
...

The Rise of Virtualization

The story of server-side infrastructure over the last couple of decades is the story of the rise of virtualization, and, its sibling, containerization. Both are ways of abstracting the hardware away from the software running on it, which has some powerful benefits. It makes it easier add or remove hardware at will, for example, which simplifies scaling. It also facilitates automating configuration, which eases deployment. Both of these qualities of virtualization are leveraged by Codespaces.

AWS, Azure, Docker, Heroku, and Kubernetes are all examples of services or technologies that leverage virtualization or containerization. It’s also the backbone of most CI/CD and serverless systems. While virtualization has revolutionized the server-side, it hasn’t had much impact on development environments outside of specialized use cases.

There are two, equally valid, ways of seeing the origins of Codespaces: one, is as a natural extension of an editor that began as a browse-based version of Visual Studio (formerly called “Visual Studio Online” now “Azure DevOps Services”), the other is as another step in the march of virtualization revolutionizing every aspect of development. These could even be considered the same story: Azure DevOps Services is of course also built on virtualization.

The Promise of Remote Development

Just being able to quickly spin up a remote development machine from git repo to make an open source contribution, or to get a quick development environment to spelunk into a dependency’s implementation details, is already enough benefit to make Codespaces popular. But the ceiling of Codespaces’ success hinges on how useful it is for day-to-day development.

On the VS Code blog, the vision is expressed with admirable restraint, focusing on the benefits for large code bases and data models requiring “massive storage and compute services”:

Because the code bases are so large, we see engineers at shops like Facebook (and Microsoft!) use editors like vim to work remotely against secure and powerful “developer VMs”, using alternative cloud-based search and navigation services that scale beyond what even the best laptop can handle.

Data Scientists building and training data models often need massive storage and compute services to analyze large datasets that can’t be stored or processed even on a robust desktop.

In Facebook’s later announcement of their partnership with Microsoft on remote development, the advantages are expressed in broader terms, suggesting that “any developer can gain” from remote development:

As Microsoft’s Visual Studio Code team stated when they first released the remote extensions, remote development is an emerging trend. While our use cases may be more advanced than most development teams given our scale, any developer can gain the benefits of remote development:

  • Work with larger, faster, or more specialized hardware than what’s available on your local machine
  • Create tailored, dedicated environments for each project’s specific dependencies, without worrying about errors due to mixed or conflicting configurations
  • Support the flexibility of being able to quickly switch between multiple running development environments without impacting local resources or tool performance

Those are compelling advantages that most developers could benefit from. So what are the chances of Codespaces supplanting local development, not just for specialized use cases, but developer’s day-to-day work on their main project?

Remote development isn’t new, it’s been around since the dawn of programming, and VS Code already has best-in-class support for it. But remote development in VS Code, while frequently praised, hasn’t moved the needle much on its own for day-to-day development. Which means we can look at the advantages of remote development that VS Code already had before Codespaces, and note that they probably won’t be enough on their own to make remove development more popular. Here are the often cited advantages of remote development before Codespaces:

  1. Developing in the same server environment that production code runs in.
  2. Using more powerful hardware.
  3. Accessing the same development environment from any machine.

In addition to those advantages, Codespaces has a new trick up it’s sleeve: Automatically setting up development environments when a new codespace is created, by installing dependencies via Docker7. In other words, Codespaces brings the same automated configuration advantages to the development side that virtualization and containerization have already brought to the deployment side. Configuring development environments is surprisingly complex, and subtle differences between manually-configured development machines creates its own problems.

It remains to be seen whether reproducible development environments is enough of a draw to move more developers over to remote development, but it’s certainly a compelling solution to a real problem.

Finally, there’s another important trait about Codespaces: It works with locked-down devices, like iPads, which normally can’t download and execute source code due to App Store Review Guideline 2.5.2. It also doesn’t require source code to be checked out locally, which many companies already consider a big security risk. These advantages will likely make some developers uncomfortable, those that see current computing trends as the gradual erosion of user freedoms, but the purpose of this piece is to predict the impact Codespaces will have on the development process, and that it aligns well with both the direction some devices are going, and many company’s security goals, are both important traits to consider.

Remote Development in Practice

Codespaces creates a fairly convincing illusion of working locally8. This is especially true when using the VS Code app with the Codespaces extension. Performing tasks like editing text, project-wide find-and-replace, or file management in the Explorer don’t exhibit any major differences from editing files locally.

One of VS Code’s best tricks is automatically forwarding ports for URLs printed in the console when connected to a remote machine. If, for example, a server process prints 127.0.0.1:3000 (because it’s running on port 3000) then port 3000 is automatically be forwarded to your local machine. You can then open that URL in a local browser window (or just -click the URL in the console), just like you would be able to if the server process were running locally9. This is another example of how VS Code creates the illusion of working locally.

But there are some situations where the illusion breaks down. Developing offline is obviously no longer an option. Another example is that when developing remotely, VS Code becomes the only easy way to edit files. If you want to edit a bitmap in Photoshop, or open a CSV file in Excel, you’ll have to figure out another way of doing so.

The vastness of VS Code’s ecosystem is an interesting tangent to explore from the limitation of not only being able to edit files with anything besides VS Code. There are extensions for tasks like editing raster graphics, a Draw.io editor for diagrams, and a tabular data viewer. If you squint, VS Code starts to look more like a general purpose platform, rather than just a text editor. The fact that this platform provides in many ways a better experience than say, VNC, is quite powerful10.

Draw.io Integration

The Draw.io Integration VS Code extension by Henning Dieterichs

Setting up and tearing down development environments at will, which Codespaces encourages, also has its downsides. If your development environment requires installing a lot of additional tools, such as compilers, linters, and other shell tools, then those tools will all need to be installed each time you create a new codespace. While Codespaces’ dotfiles support can automate this, having more dependencies will make it take longer to spin up a new codespace.

Finally, the last issue I observed while using Codespaces is that each project being in its own codespace makes it harder to make changes spanning multiple projects. This comes up when performing maintenance tasks (like updating continuous integration settings for several projects at once), making changes that span multiple projects (like many an API change and then updating consumers of that API), or even just trying to search through several projects to find a piece of code I know I’ve already written, but I don’t know where. These are all problems where organizing projects in the file-system hierarchy makes it easier to work on several related projects at once. But with Codespaces, every project is an island11.

It’s also worth mentioning that there are many types of development that Codespaces isn’t applicable for at all. Anything that needs access to local hardware, like mobile development, is obviously going to be out. The biggest audience for Codespaces is web developers (which not coincidentally, is the biggest audience of VS Code itself). Web development is a natural fit for remote development, since the deployment target of the project is also remote.

Conclusion

Codespaces provides enough utility that I suspect it will find its way into many, if not most, developers’ workflows. Just being able to open a Codespace to quickly explore, or make contributions to, a repository seems like enough to make it popular on its own. Not to mention being able to quickly edit a project from a machine that hasn’t been setup for development12. But the question I find the most interesting is whether Codespaces also has the potential to replace local development entirely, at least for some kinds of developer (those that aren’t deploying to local hardware).

I don’t expect Codespaces to win over many longtime developers, who already have sophisticated development environments setup, since Codespaces’ biggest gains come from initially setting up those environments13. The real benefit from Codespaces comes from never having to setup those local development environments in the first place, ever, over the course of a career. So what will be more interesting to watch is when new developers join projects. Without Codespaces, their first task would be to setup their development environment. With Codespaces, they can just click a button and start coding. Will developers who start working this way ever get around to setting up local development environments?14


  1. The continued relevance of the Mac Pro is an example of how relevant powerful, on premises, hardware still is. ↩︎

  2. Since Codespaces is still in beta, we’re not going to spend any time reviewing bugs or incomplete features, which might be fixed before release. This piece is about the full promise of Codespaces and remote development when it’s finished. ↩︎

  3. Setting an instance type will also come to GitHub Codespaces:

    Compute usage is billed per hour, at a rate that depends on your codespace’s instance type. During the beta, Codespaces offers a single, Linux instance type. At general availability, we’ll support three Linux instance types.

    It remains to be seen whether these features can be added without compromising the one-click experience. ↩︎

  4. At least one important feature was lost in the transition from Visual Studio Codespaces to GitHub Codespaces: self-hosted codespaces (which appears to be the most requested feature on the Codespaces Beta forum). In a way, it’s not surprising that it was removed, self-hosted codespaces fit more naturally into the Visual Studio Codespaces world (why not just let users swap the underlying Azure instance with their own hardware?), than they do into the GitHub Codespaces world (if a codespace is an extension of a repository on GitHub, how does using your own server make sense?). ↩︎

  5. Earlier in the beta, Codespaces was in the main GitHub navigation along the top (i.e., alongside “Pull Requests”, “Issues”, “Marketplace”, and “Explore”), I wonder why it was removed from there? ↩︎

  6. On macOS, the difference that jumps out between running Codespaces in the browser vs. the desktop app, is that some shortcuts that normally go straight to VS Code are instead interpreted by the browser. For example ⌘W, which closes the current split or tab, instead closes the entire browser tab. ↩︎

  7. In addition to installing a projects development dependencies, codespaces can also be personalized by installing dotfiles↩︎

  8. Emacs’ Tramp Mode is also known for creating the illusion of working locally when editing remote files. ↩︎

  9. VS Code’s port forwarding also works well with launch configurations. A launch configuration can be setup where hitting F5 (for Debug: Start Debugging) launches the server and navigates to it in your browser, and this launch configuration will work regardless of whether your project is running locally or on a remote server. ↩︎

  10. VNC works by sending a video feed from the server to the client (and forwarding keyboard and mouse events to the server), whereas with VS Code the client is actually running the front-end code. VS Code’s approach seems better to me, and it fixes the most glaring problem with VNC today: Video compression artifacts. ↩︎

  11. I’ve stopped using Codespaces for my own projects. My development environment is quite elaborate (e.g., I install many shell utilities), and I also like having all of my projects organized together on the same file system, so I can do searches or make edits across related projects. Neither of these are a good fit for Codespaces.

    But I have found I like some of the benefits of remote development. In particular, it’s nice to not have to use local hard drive space for things like npm modules, especially for smaller projects. So instead of Codespaces, I’ve been using Microsoft’s Remote SSH extension, with a VPS. This provides some of the benefits of Codespaces, while working more seamlessly with my workflow. This approach also forgoes some of Codespaces’ major selling points, like automatically setting up new development environments, and, perhaps most notably, web access via browser (it should be possible to add web access using code-server, if I ever decide I need it). ↩︎

  12. Codespaces can also be considered in terms of automation. This is my definition of automation:

    Software automation is the alignment of intention and action.

    You should be able to take one action to perform your intention.

    And ideally, that action is configurable, e.g., you can either select a menu item, press a button, or perform a keyboard shortcut.

    Codespaces takes what’s normally multi-step process, e.g., checking out the source code and then setting up a development environment, and turns it into a single action: Creating an environment for running and editing a project. Codespaces similarly optimizes finishing with a project. Normally, when you finish with a project, you might just delete the source, but this would still leave around any dependencies that were installed globally. When you remove a codespace, all of its dependencies are automatically removed with it.

    With Codespaces, intention and action are aligned. The single action of creating or removing a codespace accomplishes the intent of creating a working development environment or completely removing it. ↩︎

  13. Codespaces also presents a future for development that’s compatible with locked-down devices (e.g., iPads). I once thought creative professionals, like programmers, would eventually end up working on locked-down devices (defined here as a system that can only run sandboxed apps, but I no longer think that’s the case. ↩︎

  14. Replit is a start-up that’s also trying to remove the effort involved in setting up and maintaining development environments. See Replit co-founder Amjad Msaad discuss the original motivation behind it where he describes setting up a development environment as more difficult than development itself.

    The comparison of Replit to Codespaces is that Codespaces takes existing development workflows, and works backwards to figuring out how to make it as easy as possible for new developers to join projects. Whereas Replit asks what if development prioritized making it as easy as possible for new developers to start coding from the beginning? Both of these seem like valid approaches, and will likely end up serving different segments of the market. ↩︎


The Era of Visual Studio Code

VS Code Hero

The most important thing I look for when choosing which tools to use is longevity. Learning software is an investment, and if you have to switch to another application later, you lose some of that investment.

In most software categories, choosing the software with longevity is easy, the most popular tools are usually the ones that have been around the longest. Microsoft Excel and Adobe Illustrator were both released in 1987 and, for the most part, they’ve remained the most popular software in their categories since then.

Text editors, on the other hand, are a software category where the most popular options are not the oldest. According to the Stack Overflow Annual Developer Survey, Sublime Text was the most popular text editor available on the Mac from 2015–2017. Sublime Text was released in 2008, a sprightly youth compared to Excel and Illustrator. Text editors have been a category with a lot of movement: In the last 20 years, TextMate, Sublime Text, and Atom have all been the text editor with the most momentum1. For big complicated desktop software, has any other category ever had so much movement?

I believe the era of new text editors emerging and quickly becoming popular has now ended with Visual Studio Code. VS Code has reached unprecedented levels of popularity and refinement, laying a foundation that could mean decades of market dominance. If, like me, one of your priorities for your tools is longevity2, then that means VS Code might be a great text editor to invest in learning today.

The case for VS Code’s longevity comes from several points we’ll cover in this piece:

  1. Popularity: It’s crossed a popularity threshold that no earlier text editor in recent history has crossed.
  2. The Text Editor as Platform: It’s the endgame of a revolution that saw text editors be remade around extensions.
  3. Paradigm Transcendence: It’s transcended its paradigm as a desktop app by becoming a hosted web app, and even a reference implementation.
  4. Company Management: It’s managed by a powerful tech company, and it’s being developed aggressively.

Popularity

VS Code is the most popular text editor today. It’s so popular, that it could be the most popular GUI programming tool of all time. Since 2015, Stack Overflow has included questions about text editors in their survey3. Back then Notepad++ was the most popular text editor, with 34.7% of respondents saying they were “likely to use it”. In the following years, the popularities of different text editors moved around a bit, but nothing ever broke the 40% mark. That is, until its most recent poll in 2019, when VS Code jumped to 50.7%. This was the second year in a row that VS Code increased by ~45%, this time jumping from 34.9% in 2018, where it had already been the most popular.

Text Editor Popularity 2015–2019

Text Editor Popularity

(Note that Stackoverflow started allowing multiple answers between 2015 and 2016, so I’d take the changes between those two years in particular with a grain of salt.)

The Text Editor as Platform

So VS Code is objectively wildly popular; the next point we’re going to look at is more qualitative. For the past couple of decades, text editors have been on a trajectory that I believe VS Code is the final representation of. This is the progression of text editors becoming platforms in their own right by increasing the role and capabilities of extensions. What follows is the history of this progression4.

Pre-2004: BBEdit, Emacs, and Vim

BBEdit, Emacs, and Vim are all great text editors in their own right, but they all have idiosyncrasies that (while beloved by people like me) prevent them from ever being the most popular text editor.

Emacs, and Vim’s predecessor Vi, were both first released in 1976, before many of todays user-interface conventions were solidified. Predating conventions like using a modifier key with Z, X, C, and V for undo, cut, copy, and paste (keyboard shortcuts that were popularized by the original Macintosh and Windows 1.0, released in 1984 and 1985 respectively). Neither Emacs5 or Vim use these keys, and instead use their own terminology. They both use the term “yank” for example (although to mean different things, it’s copy in Vim, and paste in Emacs).

BBEdit was released in 1992, around the time that some of the first GUI tools emerged that would become dynasties. Note the proximity to Excel (1987), Illustrator (1987), and Photoshop (1990). And just like those apps, BBEdit is still relevant today. But unlike those apps, it’s not the most popular in its category, by a wide margin. The reason seems to be at least partially that it never fully adapted to a world where text editors put so much emphasis on package-driven ecosystems.

2004: TextMate

TextMate, released in 2004, is arguably the most influential text editor ever. Among the numerous features it popularized are abbreviation-based snippets, automatic paired characters, and fuzzy finding by file name. All of these features became staples in every popular text editor that followed. The implementations of Scope Selectors and theming that TextMate pioneered have also formed the basis for themes and syntax highlighting in every subsequent popular text editor.

That’s already a lot to originate from a single app, but it still doesn’t even include TextMate’s most significant innovation; the one that would go on to re-shape text editors, solidify niche status for every text editor that came before it, and pave the way for VS Code to become the most popular text editor in history a decade later. TextMate’s most important innovation was that it was the first popular text editor that was primarily built around extensions.

While TextMate popularized the concept of a text editor built around extensions, in hindsight, it didn’t go far enough. TextMate’s extensions had limitations that later text editors would thrive by removing.

2008: Sublime Text

Sublime Text, released in 2008, popularized the minimap and multiple cursors. And unlike TextMate and BBEdit, it’s cross-platform, running on Linux, MacOS, and Windows, which helped it reach a wider audience than those editors. But Sublime Text’s biggest impact was greatly expanding the capabilities of extensions.

Sublime Text’s extensions run in an embedded Python runtime with an extensive API, unlike TextMate which uses the scripting languages built-in to macOS, and rather than having a proper extension API, mainly centers on processing standard out.

Sublime Text greatly expanded what extensions could do, allowing more sophisticated integrations such as linters that included GUI components. And Package Control, the enormously popular package manager for Sublime Text built by Will Bond6, features a centralized source for package management, reducing the friction to browse, install, and update packages; a model that all subsequent popular text editors would also adopt.

Even with Sublime Text’s expanded extensions, it still didn’t go far enough. Package Control wasn’t built-in, and, while Sublime Text does have an API, its use of Python with custom calls for GUI components still left room for future text editors to make extensions more accessible to build.

2014: Atom

Atom, released by GitHub in 2014, finally brings extensions to their final form. Atom’s package manager is built in7, displays extension READMEs complete with inline images (and early extensions made by GitHub themselves popularized the convention of using animated GIFs to illustrate functionality), creating an extension experience reminiscent of an app store.

Then there’s the matter of HTML and CSS8. Atom is built on Electron9, which means the editor itself is written in JavaScript and runs on Node10. Compared to Sublime Text’s Python API; HTML, CSS, and JavaScript are some of most widely-known languages in existence, which greatly lowers the barrier of entry for creating extensions.

Atom had essentially perfected the extension-based editor, there was just one problem: It’s slow. Performance complaints have plagued Atom since its release, and market ended up split with Sublime Text, which is lightning fast by comparison.

2015: Visual Studio Code

VS Code was released in 2015, based on the Monaco Editor that Microsoft had first released in 2013 that could be embedded into websites. When GitHub released Electron along with Atom. Microsoft used it to create a desktop version of the Monaco Editor called Visual Studio Code.

VS Code takes the same formula as Atom11—a local web-based text editor written in Electron with an emphasis on extensions—and makes it more performant. VS Code makes extensions even more visible, by putting them in the sidebar, raising to the same level as file browsing, searching, source control, and debugging. VS Code extensions can have rich user-interfaces, being written in HTML, CSS, and JavaScript, and with full-access to Node, they can essentially do anything any other application can do. And indeed, some extensions start to look like apps in and of themselves.

With VS Code, the extension-based text editor has seemingly reached its final form. Ever since TextMate, extensions have increased in prominence and capabilities, and with VS Code, that progression appears to have culminated. There just isn’t anywhere else to go. Correspondingly, there isn’t a way a new text editor can leapfrog VS Code the same way previous text editors have been leapfrogging each other by improving extensions.

Paradigm Transcendence

So far we’ve looked at VS Code’s popularity, and its extensions implementation, as indicators of longevity. The third indicator we’ll look at is how VS Code has moved beyond the confines of the desktop. The code-server project runs VS Code as a regular web app, in other words, hosted on a server and accessed through the browser. GitHub’s Codespaces also run VS Code as a web app, this time by spinning up an ad hoc development environment.

Transcending a paradigm, like going from a desktop app to a web app, is a great indicator of longevity. For one, it means it’s more likely to be ported to more paradigms in the future. It takes herculean effort to port to a new paradigm, and expending that effort is a great indicator of value. Emacs and Vim were both ported from the terminal to GUI applications; they were too valuable not to have GUI versions. Photoshop and Excel both run on mobile12, and Illustrator is coming soon. Excel also has a web version13, and there’s a streaming version of Photoshop (although it’s been in closed beta for six years).

Not only has VS Code transcended the parameters of its initial implementation by becoming a web app, it’s also became something of a standard. Version 1.0 of the Theia IDE maintained by the Eclipse Foundation is a re-implementation of VS Code. VS Code is now not only a text editor, but also a model of how a text editor should behave.

Company Management

TextMate is largely the work of one developer, Allan Odgaard, the same with Sublime Text and Jon Skinner. Both of these applications eventually ran into trouble with frustrated users for perceived slow release schedules.

Here’s the history of major releases for these two applications:

  • 2004: TextMate 1
  • 2008: Sublime Text 1
  • 2011: Sublime Text 2 Alpha
  • 2012: Sublime Text 2
  • 2012: TextMate 2 Alpha
  • 2013: Sublime Text 3 Beta
  • 2017: Sublime Text 3
  • 2019: TextMate 2

Here’s a graph of the number of years between stable major releases (contrasted with the release dates for BBEdit 10–13 for comparison):

Text Editor Release Schedule

A couple things jump out from this graph immediately:

  1. TextMate 2 took a long time.
  2. Sublime Text has been consistent with their release schedule.

The complaints about Sublime Text seem to center around the gap between the Sublime Text 3 Beta being announced in 2013 and released in 2017, and a perceived lack of sufficient changes during that period. Sublime Text’s release schedule is slow when compared to BBEdit’s which has released three major versions (11, 12, and 13), while Sublime Text 3 has been in beta. Although Coda 2 was released in 2012, and hasn’t been updated since, so it’s unclear whether Sublime Text’s release schedule is really an anomaly for a commercial text editor.

The current version of VS Code is 1.49, but VS Code is open source, so it plays by different rules than commercial apps. Major versions exist at least partially as an opportunity for companies to charge for upgrades.

Since VS Code is developed out in the open, we can evaluate its pace of development directly by reviewing its commit history. VS Code’s commit graph on GitHub tells a story of aggressive development, out pacing Atom, and even other large open source project like Facebook’s React (note that these graphs have different scales on the Y-axis).

Visual Studio Code Commit Graph

Visual Studio Code Commit Graph

Atom Commit Graph

Atom Commit Graph

React Commit Graph

React Commit Graph

Aggressive development pulls platforms away from the pack because the combination of forward momentum, and third parties building on the platform, is difficult to compete with14. This is the same combination that makes it so hard for new entrants to compete with popular browsers or OSes.

Conclusion

The goal of this piece is to determine if VS Code is a good investment in learning if you value longevity. An implication of the Text Editor as Platform, is that since TextMate’s introduction in 2004, every few years the text editor with the most momentum has changed. These would be short reigns by any standard, but they’re positively miniscule compared to apps like Excel and Photoshop. Learning a new text editor is a questionable investment if you expect something new to come along every few years.

VS Code is giving indications that the era of short reigns for text editors is over. It has the potential to maintain its position as the most popular text editor for a much longer time, possibly for decades, if we use examples of popular software in other categories as a guides. As we’ve outlined in this piece, the case for this is following:

  1. It’s crossed a popularity threshold that’s eluded other text editors by being used by over 50% of developers.
  2. It’s the final form of progression towards maximizing the power and visibility of extensions, making it immune to being leapfrogged by a new text editor with a more capable extension model.
  3. It’s moved beyond its origins as a desktop app, it’s also a web app, and it’s even become a model of how a text editor should behave.
  4. It’s managed by a company, so it’s unlikely to run into the development stagnation that’s plagued other text editors.

Before VS Code, I expected to never learn another general-purpose text editor that wasn’t Emacs or Vim again, it was just too risky. I’ve found a good way to make predictions is to assume things will stay the same; with text editors, that means expecting a new text editor will emerge every few years that gains most of the momentum. Expecting anything else to happen requires extraordinary evidence.

I believe VS Code has presented extraordinary evidence. I’m hoping it moves into the category with apps like Excel, Illustrator, Photoshop, software that has held the most popular position in its category for decades. These applications are reliably time investments that repay their cost of learning over the course of a career. Emacs and Vim have always been like that, but it’s always good to have more choice.


  1. If you think about it, the fact that the most popular text editor is newer than popular software in other categories is pretty strange, since text editing predates almost every other computer task. I think there are a couple of reasons for this. The first is that, on a technical level, writing a text editor is easier than other categories. While I don’t want to downplay the difficulty, text files are the lingua franca of computers, and nearly every major software development framework has at least some built-in support for them. Modern hardware also gives you a lot of performance headroom to develop a text editor that you don’t have if you’re developing, say, a video editor.

    The second reason is that it’s easier for users to switch text editors. While learning new complex software packages is always difficult, at least with text editors you can just open your existing projects with a new one and start editing them, since development projects are just made up of plain text files. There’s almost no other software category where this is true, most applications use proprietary formats that only their application can read. Another reason text editors tend to be easier to pick up is that it’s usually, but not always, easy to figure out the basics: How to enter and edit text. The basics are usually easier to figure out than, say, Adobe Illustrator, which is almost impossible to use without reading the manual.

    These factors combine to make text editors a particularly competitive market, and competition is effective in driving innovation. For my money, it’s made text editors the best software there is: They have the best balance of form and function of any software category. The closest competition are browsers and terminals, which also combine power and flexibility into an elegant package, but I give the edge to text editors, because browsers and terminals achieve their power by simply wrapping powerful concepts, a protocol and standard streams respectively. With text editors in contrast, the user interface is the application in a way that just isn’t true for those other types of applications. (This is also why browsers and terminals all feel roughly the same, while text editors are wonderfully diverse.) ↩︎

  2. If longevity is my priority, then why not use Emacs or Vim? For Vim, the answer is easy, I do already use it. But I don’t like writing prose in Vim’s modal editing model, so I started seeking out a non-modal editor to complement Vim.

    I’ve also spent a fair amount of time with Emacs, but it started crashing for me with an issue similar to this one. The author of that post solved their problem by compiling Emacs locally to run it in lldb, which is farther than I was willing to go to solve my problem.

    Emacs has a difficult balancing act to walk: It’s incredibly customizable, but it’s also fragmented. For the Mac, there are several popular ports. And, macOS isn’t a high-priority platform for Emacs. There’s a history of blocking macOS-only features from Emacs, as well as removing features that are already working. All-in-all this makes Emacs a hard sell on macOS. Customizability and fragmentation aren’t a great combination to begin with, because customizations will often work in one version and not another. But when combined with relatively low market-share (4.5% in 2019), and being on a platform that’s a second-class citizen relative to GNU/Linux, it’s hard to recommend, despite its strong points. ↩︎

  3. For some reason Stack Overflow removed the questions about developer tools like text editors for the 2020 survey unfortunately. ↩︎

  4. The progression of text editors becoming a platform is adapted from a previous post, which is in turn adapted from a Twitter thread↩︎

  5. Emacs does include cua-mode, which when turned on, defines C-x, C-c, C-v, and C-z as “cut (kill), copy, paste (yank), and undo respectively”. (The name cua-mode, is a bit of a misnomer because IBM Common User Access never used these key bindings.) ↩︎

  6. Will Bond was hired by Sublime HQ in 2016↩︎

  7. TextMate 2, released in December 2011, also had the first built-in extension browser in a popular text editor. ↩︎

  8. Light Table, released in 2012, is another important milestone in the web-based text editor journey. Light Table uses NW.js (formerly node-webkit), a predecessor to Electron, and it had an integrated package manager—foreshadowing the same combination that Atom would use a couple of years later.

    What’s most interesting about Light Table that it focused on innovative new features first, like watching variables change as code executes, evaluating code inline, and embedding running projects in the editor itself (some of these features inspired by Bret Victor’s seminal “Inventing on Principle” talk). These are features that even now, eight years later, have been slow to make it into the text editors that followed.

    Light Table was about new features that weren’t available anywhere else, whereas Atom, its closet successor that used a similar web-based approach, was about incremental improvements over previous text editors. Atom’s main feature was that it was web-based, whereas Light Table was about new features that had never been done before. ↩︎

  9. Electron was originally called “Atom Shell”↩︎

  10. Atom was originally written in CoffeeScript↩︎

  11. VS Code is less “hackable” than other text editors. For example, it doesn’t have an “init” file in the tradition of .emacs.d and .vimrc (Atom does have one). This makes VS Code harder to customize, since the only way to do it is to create an extension. ↩︎

  12. Presumably, VS Code would already exist on iOS were it technically feasible to do so, since it’s open source and so popular. It makes an interesting case study for the future of iPadOS as a platform. Because, if it’s not technically possible to port VS Code to iPadOS, then, as VS Code becomes ubiquitous, that increasingly becomes synonymous with iPadOS not supporting programming at all.

    The point is probably moot, because an iOS native version of VS Code would probably work with the same client-server model described in Paradigm Transcendence. But it’s still an interesting thought experiment, because I often see the prediction that iPadOS will [disrupt] the industry from the bottom(https://en.wikipedia.org/wiki/Disruptive_innovation). I wonder how can that happen if a platform puts up so many technical barriers for creating a text editor? ↩︎

  13. Another nice thing about having a web version is that web apps don’t have to abide by the App Store Review Guidelines, so applications prohibited by Apple can still exist on the platform. ↩︎

  14. As I’m fond of saying, if you’re looking for areas that will be disrupted by new technology, look for areas of neglect↩︎


Remembering the O'Reilly Mac OS X Innovators Contest

From 2003 to 2004, O’Reilly Media ran the O’Reilly Mac OS X Innovators Contest, sponsored by Apple via the Apple Developer Connection (now just Apple Developer). I still think of these winners as some of the high watermarks of app innovation. Many concepts we take for granted today were either introduced, or popularized, by apps on this list. Here are a few of my favorite examples:

  • NetNewsWire, while not necessarily the first RSS feed reader, was one of the most popular early ones. RSS feed readers are the origin of consuming media through streams of content, now exemplified by Twitter and Facebook’s Newsfeed.
  • SubEthaEdit was one of the earliest practical implementations of multiple simultaneous live document editing, a concept later brought to a much wider audience by Google Docs in 2006.
  • LaunchBar popularized many features we take for granted in search user interfaces today, such as seeing search results live as you type, fuzzy string matching, and combining search results of various types, such as apps, bookmarks, contacts, events, and files all into one unified interface.

I’ve listed the winners below, and linked to all the ones that are still maintained, so you can see visually just how many these apps are still around. All of these apps are over fifteen years old now.

2003 Round One Winners

2003 Second Round Winners

2003 O’Reilly Mac OS X Conference Winners

  • First Place, US Category: OmniOutliner
  • Second Place, US Category: iBlog
  • First Place, International Category: iStopMotion
  • Second Place, International Category: ACSLogo
  • Honorable Mention: F-Script

2004 Winners

  • First Place, U.S.: Delicious Library
  • First Place, International: FotoMagico
  • Second Place, U.S.: Curio
  • Second Place, International: iDive
  • Honorable Mention, U.S.: Nicecast
  • Honorable Mention, International: Process

macOS Big Sur: Has the Dust Finally Settled on Limiting Third-Party Apps?

Apple’s strategy for years has been to trade desktop power for cross-device feature parity. As expected, macOS Big Sur continues this trend, emphasizing a consistent user interface across devices, and focusing on cross-device development technologies like SwiftUI and Catalyst.

Personally, I wish Apple had different priorities. I’d like to see more apps like Sketch, an industry-leading creative app that’s built top to bottom on Apple technologies. But Sketch was released in 2010, and Apple hasn’t created any new frameworks like Core Graphics and Core Image that support these kinds of apps in over a decade. So I wasn’t holding my breath for them to announce anything new for these kinds of apps at WWDC this year.

Since Apple isn’t prioritizing powerful desktop apps with their own technologies, that means supporting these use cases mostly falls on third parties. This is where companies like Adobe, Avid, Maxon, and Microsoft come in. While Apple’s priorities regarding their own technologies have been clear for awhile now, what hasn’t been clear is their priorities for third-party apps, in particular, ones that aren’t sandboxed. The trend for the last few years has been making it harder to develop these kinds of apps for macOS. AEpocalypse (2018), Catalina’s security features (2019), and Notarization (2018) are all examples of this trend.

The overarching reason behind the changes that make developing these kinds of apps harder is “Security”. And unlike cross-device feature parity, it’s unclear exactly where this all ends. Because after all, the most secure system there is is the one that doesn’t run any software at all. That’s why it’s such a pleasant surprise, that, as far as I can tell, Apple has done everything they can to make Big Sur, and the accompanying transition to Apple Silicon, as seamless as possible, even for apps that aren’t sandboxed.

Some were predicting that Macs on Apple Silicon wouldn’t even run apps from outside of the Mac App Store, that didn’t happen. It seemed more likely that Apple would drop OpenCL and OpenGL, but those are sticking along for the ride. No details were known about whether there would be an emulation layer like the original Rosetta from the 2006 Intel transition. Apple appears to have gone above in beyond with Rosetta 2, which even supports plug-ins like VSTis, giving lots of options for migration paths for powerful modular software.

I’m still frustrated that there probably won’t be another Sketch for the foreseeable future, but that ship sailed a long time ago. And no other platform has a Sketch either, an industry defining app that’s a platform exclusive, so while Apple has lost a unique advantage that only they had, they haven’t lost anything that other platforms already have. Other platforms can run powerful modular software that’s distributed outside the Mac App Store, but today, so can new Macs running Big Sur. Here’s to hoping that the dust has settled, and the last of the restrictions on third-party apps are behind us now.




Software to Die For

Before I switched to being a full-time developer in 2010, I worked as a user-interface designer for seven years. Something that always bothered me during that time is that so much of what I was learning was just how to use Photoshop really well. After I switched to development, I was hesitant to ever invest that much in just learning a big software package again. “What if I choose wrong? And waste all those years of learning by switching to another software package?” I asked myself. Recently, I’ve re-evaluated that decision, based on my analysis of the market share of major creative applications. It turns out if I’d just chosen which software I want to learn ten years ago, for most categories, it would still be the same today. For some categories, it would still be the same if I’d chosen twenty years ago, and it’s often the first software that was ever introduced to solve a problem, even if that application is over 30 years ago, that’s still the best choice today. So it turns out I was overcorrecting relative to the risk in learning big complex packages, so now I’m investing in doing it again.

This is the list of software I’ve chosen to learn:

Some of these I already know quite well (Vim, Photoshop), and some I’ve barely touched (Premiere Pro, Final Cut Pro). I’m not happy with the duplication, and frankly, this is probably just too much for one person. Learning any one of these applications is an lifetime of work, let alone all of them. But I can’t decide what to cut, so here we are.