Deno is a Browser for Code

I started contributing to Deno soon after Ry made the prototype visible in May 2018. The most frequent question that people have is "where is the package manager?" which often times isn't even in the form of a question. It is statements like "I thought Deno took security seriously, and just downloading resources off the internet is insecure." or "How can I possibly manage my dependencies?"

In my opinion, we need to shift our mental model. Lots of folks take the ubiquity of package managers and centralized code registries as a requirement to have a package manager and a centralized code registries. Because they exist doesn't mean they are required. They came into existence because they solved problems in a particular way, and we have just accepted them as the only way to solve that problem. I would argue that isn't true.

Browsers

In order to publish a website, we don't login to a central Google server, and upload our website to the registry. Then if someone wants to view our website, they use a command line tool, which adds an entry to our browser.json file on our local machine and goes and fetches the whole website, plus any other websites that the one website links to to our local websites directory before we then fire up our browser to actually look at the website. That would be insane, right? So why accept that model for running code?

The Deno CLI works like a browser, but for code. You import a URL in the code and Deno will go and fetch that code and cache it locally, just like a browser. Also, like a browser, your code runs in a sandbox, which has zero trust of the code you are running, irrespective of the source. You, the person invoking the code, get to tell that code what it can and can't do, externally. Also, like a browser, code can ask you permission to do things, which you can choose to grant or deny.

The HTTP protocol provides everything that is needed to provide information about the code, and Deno tries to fully leverage that protocol, without having to create a new protocol.

Discovering code

The first thing to think about is that, like a browser, the Deno CLI doesn't want to have any opinions about what code you run. It lays out the rules of how code is fetched, and how it sandboxes itself from the machine it runs on. In my opinion, that is as much of an opinion a runtime should have.

In the Node.js/npm ecosystem, we have conflated the management of code on our local machine, with a centralized registry of code to help facilitate discovery. In my opinion, both have really bad flaws.

Back in the early days of the internet, we experimented with npm type of discoverability. You would go add your website to Yahoo! under the right categorization and people would come along, maybe use the search function, but it was all structured based on the opinions of those providing the content, not really based on optimizing for the needs of the consumer. Eventually along came Google. Why did Google win? Because it was useful. It indexed websites in a way that matched simple expressions of need (search terms) with the most relevant web pages that met that need, looking at multiple factors, including meta data provided the content provider as one factor in the mix.

While we don't have that model quite yet for code for Deno, it is a model that works. In addition, we use Google because it solves problems for us, instead of being told "you must use Google", as well as there are also other viable alternatives to Google.

I got into a bit of a debate with Laurie Voss on twitter, someone who knows a fair deal about the npm ecosystem I would say. He argued that Deno needed a package manager, and this blog post is a longer winded version of the thoughts I wanted to express, but Laurie raised a very valid point.

{{< tweet 1261140647056076801 >}}

GitHub has become the home for open source code, because it was useful and solved problems, and built on top of the de facto source code versioning tool, git. From the Deno CLI perspective, there should be no technical restrictions to where you source code from, it is up to the wider eco-system to create and evolve ways to make code for Deno discoverable, probably in innovative ways that could never have been conceived by those of us creating the CLI.

Repeatable builds

In the npm eco-system, this became a problem. Because of the heavy reliance on semantic versioning, and the complex dependency graphs that tend to come from the Node.js/npm eco-system, having a repeatable build became a real problem. Yarn introduced the concept of lock files, of which npm followed suit.

My personal feeling is it was a bit of the tail wagging the dog, in that the behaviours of developers in the eco-system created a problem that then needed an imperfect solution to fix it. Any of us that have lived with the eco-system for a long time know that the fix to a lot of issues is rm -rf node_modules package-lock.json && npm install.

That being said, Deno has two solutions for that. First, is that Deno caches modules. That cache can be checked into your source control, and the --cached-only flag will ensure that there is not attempts to retrieve remote modules. The DENO_DIR environment variable can be used to specify where the cache is located to provide further flexibility.

Second, Deno supports lock files. --lock lock.json --lock-write would write out a lock file with hashes of all the dependencies for a given workload. This would be used to validate future runs when the --lock lock.json is used.

There are also a couple other commands that make managing repeatable builds. deno cache would resolve all the dependencies for a supplied module and populate the Deno cache. deno bundle can be used to generate a single file "build" of a workload which all the dependencies are resolved and included in that file, so only that single file is needed for future deno run commands.

Trusting code

This is another area where I think we have a skewed mental model. For whatever reason, we put trust in code that is in a centralized registry. We don't even think about it. Not only that, we trust that that code has fully vetted all of its dependencies and that those are to be trusted to. We do a quick search and type in npm install some-random-package and think "This is Fine!" I argue the rich npm package eco-system has lulled is into a sense of complacency.

To compensate for this laxness and complacency, we implement security monitoring software in our tool chains, to analyse our dependencies and the thousands upon thousand lines of code to let us know that maybe some of the code is exploitable. Corporations setup private registries to host packages that might be vetted slightly more than the single public registry.

It feels like there is an elephant in the room here. The best strategy is we shouldn't trust any code. Once we have that established, then opening it back up becomes a little be easier. But we are lying to ourselves if we think a package manager and a centralised registry solve this problem, or even substantially help with this problem. In fact, I argue they make use let our guards down. "Well it is on npm, if it were bad for me, surely someone would take it down."

Deno in this aspect isn't quite as done as I think it should be, but it is starting from a good position. It has zero trust at startup, and provides fairly fine grained permissions. One of the things I personally dislike is that there is the -A flag, which is basically saying "oh yeah allow everything" which is such an easy thing for a frustrated developer to do instead of figuring out what they really need.

It is also hard to break down those permissions, to say "this code can do this, but this other code over here can't" or when code prompts to escalate privileges where is that code coming from. Hopefully we can figure out an easy to use mechanism coupled with something that would be effective and performant at runtime to try to solve those challenges.

A recent change though, which is a good one, in my opinion, is that Deno no longer allows you to downgrade your imports. If something is imported from https:// then it can only import from other https:// locations. This follows the browser model of not being able to downgrade transport. I still think longer term it would be good to kill off any remote imports that aren't over https://, much like Service Workers require HTTPS, so we will see what the future holds.

Dependency management

I think we need to talk frankly about dependencies in the npm ecosystem. To be honest, it is broken. An ecosystem that enables 5 lines of code to be downloaded and installed 30 million times a week for code that has been in every browser for the last 9 years and never was needed in Node.js is a broken ecosystem. This one example, the actual code is 132 bytes, but the package size is 3.4kb. The runnable code is 3.8% of the package size. "This is Fine!"

My opinion is that there are several factors involved in this. A big part of it is that we have the model inverted, which I talked about Deno being a browser for code. The problem is that this backwards model has infected how we create websites. While we don't have a central registry, when we build a website, we download all the code we depend up and bake it into something that we load up on a server, and then each user downloads a bunch of code to their local machine. Some evidence is that only around 10% of that code that is downloaded is unique to that site or web application, the rest is all that code we are downloading to our development workstation and bundling up. This model being broken are some of the problems solutions like Snowpack are trying to solve.

Another significant problem is that our dependencies are not coupled with our code. We put dependencies in our package.json but if our code actually uses that code or not is totally decoupled. While our code expresses what we are using out of that other code, it is very loosely coupled to the version of that code. That is contained in the package.json, though it has the biggest impact on the code we write, because it is the code that is actually consuming the dependent code.

This leads us to the Deno model, which I like to call Deps-in-JS, since all the cool kids are doing *-in-JS things. Explicitly stating our external dependencies as URLs means that the code depends upon the other code is concise and clear, and our code and dependencies are tightly coupled together. If you want to see that dependency graph, you simply need to use deno info with a local or remote module:

$ deno info https://deno.land/x/oak/examples/server.ts
local: $deno/deps/https/deno.land/d355242ae8430f3116c34165bdae5c156dca21aeef521e45acb51fcd21c9f724
type: TypeScript
compiled: $deno/gen/https/deno.land/x/oak/examples/server.ts.js
map: $deno/gen/https/deno.land/x/oak/examples/server.ts.js.map
deps:
https://deno.land/x/oak/examples/server.ts
  ├── https://deno.land/std@0.53.0/fmt/colors.ts
  └─┬ https://deno.land/x/oak/mod.ts
    ├─┬ https://deno.land/x/oak/application.ts
    │ ├─┬ https://deno.land/x/oak/context.ts
    │ │ ├── https://deno.land/x/oak/cookies.ts
    │ │ ├─┬ https://deno.land/x/oak/httpError.ts
    │ │ │ └─┬ https://deno.land/x/oak/deps.ts
    │ │ │   ├── https://deno.land/std@0.53.0/hash/sha256.ts
    │ │ │   ├─┬ https://deno.land/std@0.53.0/http/server.ts
    │ │ │   │ ├── https://deno.land/std@0.53.0/encoding/utf8.ts
    │ │ │   │ ├─┬ https://deno.land/std@0.53.0/io/bufio.ts
    │ │ │   │ │ ├─┬ https://deno.land/std@0.53.0/io/util.ts
--snip--

Deno has no strong opinions around "versions" of code. A URL is a URL is a URL. While Deno requires an appropriate media type in order to understand how to treat code, all the "opinions" about what code to serve up is left up to the web server. A server can implement semantic versioning to its hearts content, or do any sort of "magical" mapping of URLs to resources it wants. Deno doesn't care. For example https://deno.land/x/ is effectively nothing but a URL redirect server, where it rewrites URLs to include a git commit-ish reference in the redirected URL. So https://deno.land/x/oak@v4.0.0/mod.ts becomes https://raw.githubusercontent.com/oakserver/oak/v4.0.0/mod.ts, which GitHub serves up a nice versioned module.

Of course spreading "versioned" remote URLs throughout your codebase doesn't make a lot of sense, so don't do that. The great thing about the dependencies just being code though is that you can structure them any way you want to. A common convention is to use a deps.ts which re-exports all the dependencies you might want. Take a look at the one for oak server:

// Copyright 2018-2020 the oak authors. All rights reserved. MIT license.

// This file contains the external dependencies that oak depends upon

// `std` dependencies

export { HmacSha256 } from "https://deno.land/std@0.51.0/hash/sha256.ts";
export {
  Response,
  serve,
  Server,
  ServerRequest,
  serveTLS,
} from "https://deno.land/std@0.51.0/http/server.ts";
export {
  Status,
  STATUS_TEXT,
} from "https://deno.land/std@0.51.0/http/http_status.ts";
export {
  Cookie,
  Cookies,
  delCookie,
  getCookies,
  setCookie,
} from "https://deno.land/std@0.51.0/http/cookie.ts";
export {
  basename,
  extname,
  isAbsolute,
  join,
  normalize,
  parse,
  resolve,
  sep,
} from "https://deno.land/std@0.51.0/path/mod.ts";
export { assert } from "https://deno.land/std@0.51.0/testing/asserts.ts";

// 3rd party dependencies

export {
  contentType,
  lookup,
} from "https://deno.land/x/media_types@v2.3.1/mod.ts";

I created oak server and maintained for 18 months through about 40 releases of Deno and the Deno std library, including moving of media_types from internal to oak, out to the std library, to only have it be "ejected" from the std library to be its own thing. Not once did I think to myself "hey, I need a package manager to manage this for me".

One of the benefits of TypeScript is that you can get comprehensive validation of compatibility of your code with other code. If your dependencies are "raw" TypeScript written for Deno, this is great, but let's say that you want to take advantage of pre-processing of the TypeScript to JavaScript, but still have the ability to consume that remote code safely. Deno supports a couple different ways to allow that to happen, but the most seamless is the support for the X-TypeScript-Types header. This header indicates to Deno where a types file is located which can be used when type checking the JavaScript file that you are depending upon. Pika CDN supports this. Any packages that are available on the CDN that have types associated with them will serve up that header and Deno will also fetch those types and use that when type checking the file.

All this being said, there may still be a need to "remap" a remote (or local) dependency to what is expressed in the code. In this case, the unstable implementation of import-maps can be used. It is a proposal specification that is part of the W3C incubator where browser standards come out of. It allows a map to be provided which will map a particular dependency in code to another resource, be it a local file or a remote module.

We had it implemented in Deno for an extended period of time, as we had really hoped that it would become adopted widely. Sadly, it was only an origin trial in Chrome and hasn't gotten wider adoption yet. This led us to putting it behind the --unstable flag for Deno 1.0. My personal opinion is that it is still a big risk of being a dead end, and should be avoided.

But, but, but...

I suspect a lot of people are still coming with a list of objections to the model that Deno has. I think the strategy Deno has tried to take, which I am very aligned to, is to deal with real problems when they arise. A lot of the objections I hear are from people who are new to Deno, who haven't worked with it, who haven't tried to understand that there might be a different way.

All that being said, if we collectively run into a problem and there is a compelling need to change something in the Deno CLI, I am confident that it will happen, but a lot of problems simply don't exist, or there are other ways to solve them that don't require your runtime to have strong opinions or be coupled to an external programme to manage your code.

So my challenge to you is, flirt a bit with not having a package manager or a centralised package repository and see how it goes. You might never go back!