Development

Excluding nested node_modules in Rollup.js

We’re often developing multiple Node packages at the same time, symlinking their trees around in order to test them in other projects prior to release.

And sometimes we hit some pretty confusing behavior. Crazy caching issues, confounding crashes, and all manner of chaos. All resulting from one cause: Duplicate modules appearing in our Rollup.js-bundled JavaScript.

For example, we may be developing Ink (our in-progress UI component library) over here with one copy of Spina (our modern Backbone.js successor), and bundling it in Review Board (our open source, code review/document review product) over there with a different copy of Spina. The versions of Spina should be compatible, but technically they’re two separate copies.

And it’s all because of nested node_modules.

The nonsense of nested node_modules

Normally, when Rollup.js bundles code, it looks for any and all node_modules directories in the tree, considering them for dependency resolution.

If a dependency provides its own node_modules, and needs to bundle something from it, Rollup will happily include that copy in the final bundle, even if it’s already including a different copy for another project (such as the top-level project).

This is wasteful at best, and a source of awful nightmare bugs at worst.

In our case, because we’re symlinking source trees around, we’re ending up with Ink’s node_modules sitting inside Review Board’s node_modules (found at node_modules/@beanbag/ink/node_modules.), and we’re getting a copy of Spina from both.

Easily eradicating extra node_modules

Fortunately, it’s easy to resolve in Rollup.js with a simple bit of configuration.

Assuming you’re using @rollup/plugin-node-resolve, tweak the plugin configuration to look like:

{
    plugins: [
        resolve({
            moduleDirectories: [],
            modulePaths: ['node_modules'],
        }),
    ],
}

What we’re doing here is telling Resolve and Rollup two things:

  1. Don’t look for node_modules recursively. moduleDirectories is responsible for looking for the named paths anywhere in the tree, and it defaults to ['node_modules']. This is why it’s even considering the nested copies to begin with.
  2. Explicitly look for a top-level node_modules. modulePaths is responsible for specifying absolute paths or paths relative to the root of the tree where modules should be found. Since we’re no longer looking recursively above, we need to tell it which one we do want.

These two configurations together avoid the dreaded duplicate modules in our situation.

And hopefully it will help you avoid yours, too.

Review Board: Between Then and Now

I just realized, before I know it, we’ll be hitting 20 years of Review Board.

Man, do I feel old.

It’s hard to imagine it now, but code review wasn’t really a thing when we built Review Board back in 2006. There were a couple expensive enterprise tools, but GitHub? Pull requests? They didn’t exist yet.

This meant we had to solve a lot of problems that didn’t have readily-made or readily-understood solutions, like:

🤔 What should a review even *be*? What’s involved in the review process, and what tools do you give the user?

We came up with tools like:

  • Resolvable Issue Tracking (a To Do list of what needs to be done in a change)
  • Comments spanning 1 or more lines of diffs
  • Image file attachment review
  • Previews of commented areas appearing above the comments.

Amongst others.

🤔 How should you discuss in a review? Message board style, with one box per reply? Everything embedded in top-level reviews? Comments scattered in a diff?

We decided on a box per review, and replies embedded within it, keeping discussion about a topic all in one place.

Explicitly not buried in a diff, because in complex projects, you also may be reviewing images, documents, or other files. Those comments are important, so we decided they should all live, threaded, under a review.

A lot of tools went the “scatter in a diff” route, and while that was standard for a while, it never sat right with me. For anything complex, it was a mess. I think we got this one right.

🤔 How do you let users keep track of what needs to be reviewed?

We came up with our Dashboard, which shows a sortable, filterable, customizable view of all review requests you may be interested in. This gave a bird’s-eye view across any number of source code repositories, teams, and projects.

Many tools didn’t go this route. You were limited to seeing review requests/pull requests on that repository, and that’s it. For larger organizations, this just wasn’t good enough.

🤔 How do you give organizations control over their processes? A policy editor? APIs? Fork the code?

We settled on:

  • A Python extension framework. This was capable of letting developers craft new policy, collect custom information during the review process, and even build whole new review UIs for files.
  • A full-blown REST API, which is quite capable.
  • Eventually, features like WebHooks, once those became a thing.

Our goal was to avoid people ever having to fork. But also, we kept Review Board MIT-licensed, so people were sure to have the control they needed.

—

I could probably go on for a while. A lot of these eventually worked their way into other code review tools on the market, and are standard now, but many started off as a lot of long nights doodling on a whiteboard and in notebooks.

We’ve had the opportunity to work for years with household names that young me would have never imagined. If you’ve been on the Internet at all in the past decade, you’ve regularly interacted with at least one thing built in Review Board.

But the passage of time and the changes in the development world make it hard these days. We’re an older tool now, and people like shiny new things. That’s okay. We’re still building some innovative shiny things. More on some of those soon 😉

This is a longer post than I planned for, but this stuff’s on my mind a lot lately.

I’ve largely been quiet lately about development, but I’m trying to change that. Develop in the open, as they say. Expect a barrage of behind-the-scenes posts coming soon.

Building Multi-Platform Docker Images Using Multiple Hosts

Here’s a very quick, not exactly comprehensive tutorial on building Docker images using multiple hosts (useful for building multiple architectures).

If you’re an expert on docker buildx, you may know all of this already, but if you’re not, hopefully you find this useful.

We’ll make some assumptions in this tutorial:

  1. We want to build a single Docker image with both linux/amd64 and linux/arm64 architectures.
  2. We’ll be building the linux/arm64 image on the local machine, and linux/amd64 on a remote machine (accessible via SSH).
  3. We’ll call this builder instance “my-builder”

We’re going to accomplish this by building a buildx builder instance for the local machine and architecture, then append a configuration for another machine. And then we’ll activate that instance.

This is easy.

Step 1: Create your builder instance for localhost and arm64

$ docker buildx create \
    --name my-builder \
    --platform linux/arm64

This will create our my-builder instance, defaulting it to using our local Docker setup for linux/arm64.

If we wanted, we could provide a comma-separated list of platforms that the local Docker should be handling (e.g., --platform linux/arm64,darwin/arm64).

(This doesn’t have to be arm64. I’m just using this as an example.)

Step 2: Add your amd64 builder

$ docker buildx create \
    --name my-builder \
    --append \
    --platform linux/amd64 \
    ssh://<user>@<remotehost>

This will update our my-builder, informing it that linux/amd64 builds are supported and must go through the Docker service over SSH.

Note that we could easily add additional builders if we wanted (whether for the same architectures or others) by repeating this command and choosing new --platform values and remote hosts

Step 3: Verify your builder instance

Let’s take a look and make sure we have the builder setup we expect:

$ docker buildx ls
NAME/NODE       DRIVER/ENDPOINT           STATUS    BUILDKIT  PLATFORMS
my-builder *    docker-container
  my-builder0   desktop-linux             inactive            linux/arm64*
  my-builder1   ssh://myuser@example.com  inactive            linux/amd64*

Yours may look different, but it should look something like that. You’ll also see default and any other builders you’ve set up.

Step 4: Activate your builder instance

Now we’re ready to use it:

$ docker buildx use my-builder

Just that easy.

Step 5: Build your image

If all went well, we can now safely build our image:

$ docker buildx build --platform linux/arm64,linux/amd64 .

You should see build output for each architecture stream by.

If we want to make sure the right builder is doing the right thing, you can re-run docker buildx ls in another terminal. You should see running as the status for each, along with an inferred list of other architectures that host can now build (pretty much anything it natively supports that you didn’t explicitly configure above).

Step 6: Load your image into Docker

You probably want to test your newly-built image locally, don’t you? When you run the build, you might notice this message:

WARNING: No output specified with docker-container driver. Build
result will only remain in the build cache. To push result image
into registry use --push or to load image into docker use --load

And if you try to start it up, you might notice it’s missing (or that you’re running a pre-buildx version of your image).

What you need to do is re-run docker buildx build with --load and a single platform, like so:

$ docker buildx build --load --platform linux/arm64 .

That’ll rebuild it (it’ll likely just reuse what it built before) and then make it available in your local Docker registry.

Hope that helps!

Re-typing Parent Class Attributes in TypeScript

I was recently working on converting some code away from Backbone.js and toward Spina, our TypeScript Backbone “successor” used in Review Board, and needed to override a type from a parent class.

(I’ll talk about why we still choose to use Backbone-based code another time.)

We basically had this situation:

class BaseClass {
    summary: string | (() => string) = 'BaseClass thing doer';
    description: string | (() => string);
}

class MySubclass extends BaseClass {
    get summary(): string {
        return 'MySubclass thing doer';
    }

    // We'll just make this a standard function, for demo purposes.
    description(): string {
        return 'MySubclass does a thing!';
    }
}

TypeScript doesn’t like that so much:

Class 'BaseClass' defines instance member property 'summary', but extended class 'MySubclass' defines it as an accessor.

Class 'BaseClass' defines instance member property 'description', but extended class 'MySubclass' defines it as instance member function.

Clearly it doesn’t want me to override these members, even though one of the allowed values is a callable returning a string! Which is what we wrote, darnit!!

So what’s going on here?

How ES6 class members work

If you’re coming from another language, you might expect members defined on the class to be class members. For example, you might think you could access BaseClass.summary directly, but you’d be wrong, because these are instance members.

Peer-Programming a Buggy World with ChatGPT AI

AI has been all the rage lately, with solutions like Stable Diffusion for image generation, GPT-3 for text generation, and CoPilot for code development becoming publicly available to the masses.

That excitement ramped up this week with the release of ChatGPT, an extremely impressive chat-based AI system leveraging the best GPT has to offer.

I decided last night to take ChatGPT for a spin, to test its code-generation capabilities. And I was astonished by the experience.

Together, we built a simulation of bugs foraging for food in a 100×100 grid world, tracking essentials like hunger and life, reproducing, and dealing with hardships involving seasonal changes, natural disasters, and predators. All graphically represented.

We’re going to explore this in detail, but I want to start off by showing you what we built:

Also, you can find out more on my GitHub repository

A Recap of my Experience

Before we dive into the collaborative sessions that resulted in a working simulation, let me share a few thoughts and tidbits about my experience:

Integration and Simulation Tests in Python

One of my (many) tasks lately has been to rework unit and integration tests for Review Bot, our automated code review add-on for Review Board.

The challenge was providing a test suite that could test against real-world tools, but not require them. An ever-increasing list of compatible tools has threatened to become an ever-increasing burden on contributors. We wanted to solve that.

So here’s how we’re doing it.

First off, unit test tooling

First off, this is all Python code, which you can find on the Review Bot repository on GitHub.

We make heavy use of kgb, a package we’ve written to add function spies to Python unit tests. This goes far beyond Mock, allowing nearly any function to be spied on without having to be replaced. This module is a key component to our solution, given our codebase and our needs, but it’s an implementation detail — it isn’t a requirement for the overall approach.

Still, if you’re writing complex Python test suites, check out kgb.

Deciding on the test strategy

Review Bot can talk to many command line tools, which are used to perform checks and audits on code. Some are harder than others to install, or at least annoying to install.

We decided there’s two types of tests we need:

  1. Integration tests — ran against real command line tools
  2. Simulation tests — ran against simulated output/results that would normally come from a command line tool

Being that our goal is to ease contribution, we have to keep in mind that we can’t err too far on that side at the expense of a reliable test suite.

We decided to make these the same tests.

The strategy, therefore, would be this:

  1. Each test would contain common logic for integration and simulation tests. A test would set up state, perform the tool run, and then check results.
  2. Integration tests would build upon this by checking dependencies and applying configuration before the test run.
  3. Simulation tests would be passed fake output or setup data needed to simulate that tool.

This would be done without any code duplication between integration or simulation tests. There would be only one test function per expectation (e.g., a successful result or the handling of an error). We don’t want to worry about tests getting out of sync.

Regression in our code? Both types of tests should catch it.

Regression or change in behavior in an integrated tool? Any fixes we apply would update or build upon the simulation.

Regression in the simulation? Something went wrong, and we caught it early without having to run the integration test.

Making this all happen

We introduced three core testing components:

  1. @integration_test() — a decorator that defines and provides dependencies and input for an integration test
  2. @simulation_test() — a decorator that defines and provides output and results for a simulation test
  3. ToolTestCaseMetaClass — a metaclass that ties it all together

Any test class that needs to run integration and simulation tests will use ToolTestCaseMetaClass and then apply either or both @integration_test/@simulation_test decorators to the necessary test functions.

When a decorator is applied, the test function is opted into that type of test. Data can be passed into the decorator, which is then passed into the parent test class’s setup_integration_test() or setup_simulation_test().

These can do whatever they need to set up that particular type of test. For example:

  • Integration test setup defaults to checking dependencies, skipping a test if not met.
  • Simulation test setup may write some files or spy on a subprocess.Popen() call to fake output.


For example:

class MyTests(kgb.SpyAgency, TestCase,
              metaclass=ToolTestCaseMetaClass):
    def setup_simulation_test(self, output):
        self.spy_on(execute, op=kgb.SpyOpReturn(output))

    def setup_integration_test(self, exe_deps):
        if not are_deps_found(exe_deps):
            raise SkipTest('Missing one or more dependencies')

    @integration_test(exe_deps=['mytool'])
    @simulation_test(output=(
        b'MyTool 1.2.3\n'
        b'Scanning code...\n'
        b'0 errors, 0 warnings, 1 file(s) checked\n'
    ))
    def test_execute(self):
        """Testing MyTool.execute"""
        ...

When applied, ToolTestCaseMetaClass will loop through each of the test_*() functions with these decorators applied and split them up:

  • Test functions with @integration_test will be split out into a test_integration_<name>() function, with a [integration test] suffix appended to the docstring.
  • Test functions with @simulation_test will be split out into test_simulation_<name>(), with a [simulation test] suffix appended.

The above code ends up being equivalent to:

class MyTests(kgb.SpyAgency, TestCase):
    def setup_simulation_test(self, output):
        self.spy_on(execute, op=kgb.SpyOpReturn(output))

    def setup_integration_test(self, exe_deps):
        if not are_deps_found(exe_deps):
            raise SkipTest('Missing one or more dependencies')

    def test_integration_execute(self):
        """Testing MyTool.execute [integration test]"""
        self.setup_integration_test(exe_deps=['mytool'])
        self._test_common_execute()

    def test_simulation_execute(self):
        """Testing MyTool.execute [simulation test]"""
        self.setup_simulation_test(output=(
            b'MyTool 1.2.3\n'
            b'Scanning code...\n'
            b'0 errors, 0 warnings, 1 file(s) checked\n'
        ))
        self._test_common_execute()

    def _test_common_execute(self):
        ...

Pretty similar, but less to maintain in the end, especially as tests pile up.

And when we run it, we get something like:

Testing MyTool.execute [integration test] ... ok
Testing MyTool.execute [simulation test] ... ok

...

Or, you know, with a horrible, messy error.

Iterating on tests

It’s become really easy to maintain and run these tests.

We can now start by writing the integration test, modify the code to log any data that might be produced by the command line tool, and then fake-fail the test to see that output.

class MyTests(kgb.SpyAgency, TestCase,
              metaclass=ToolTestCaseMetaClass):
    ...

    @integration_test(exe_deps=['mytool'])
    def test_process_results(self):
        """Testing MyTool.process_results"""
        self.setup_files({
            'filename': 'test.c',
            'content': b'int main() {return "test";}\n',
        })

        tool = MyTool()
        payload = tool.run(files=['test.c'])

        # XXX
        print(repr(payload))

        results = MyTool().process_results(payload)

        self.assertEqual(results, {
            ...
        })

        # XXX Fake-fail the test
        assert False

I can run that and get the results I’ve printed:

======================================================================
ERROR: Testing MyTool.process_results [integration test]
----------------------------------------------------------------------
Traceback (most recent call last):
    ...
-------------------- >> begin captured stdout << ---------------------
{"errors": [{"code": 123, "column": 13, "filename": "test.c", "line': 1, "message": "Expected return type: int"}]}

Now that I have that, and I know it’s all working right, I can feed that output into the simulation test and clean things up:

class MyTests(kgb.SpyAgency, TestCase,
              metaclass=ToolTestCaseMetaClass):
    ...

    @integration_test(exe_deps=['mytool'])
    @simulation_test(output=json.dumps(
        'errors': [
            {
                'filename': 'test.c',
                'code': 123,
                'line': 1,
                'column': 13,
                'message': 'Expected return type: int',
            },
        ]
    ).encode('utf-8'))
    def test_process_results(self):
        """Testing MyTool.process_results"""
        self.setup_files({
            'filename': 'test.c',
            'content': b'int main() {return "test";}\n',
        })

        tool = MyTool()
        payload = tool.run(files=['test.c'])
        results = MyTool().process_results(payload)

        self.assertEqual(results, {
            ...
        })

Once it’s running correctly in both tests, our job is done.

From then on, anyone working on this code can just simply run the test suite and make sure their change hasn’t broken any simulation tests. If it has, and it wasn’t intentional, they’ll have a great starting point in diagnosing their issue, without having to install anything.

Anything that passes simulation tests can be considered a valid contribution. We can then test against the real tools ourselves before landing a change.

Development is made simpler, and there’s no worry about regressions.

Going forward

We’re planning to apply this same approach to both Review Board and RBTools. Both currently require contributors to install a handful of command line tools or optional Python modules to make sure they haven’t broken anything, and that’s a bottleneck.

In the future, we’re looking at making use of python-nose‘s attrib plugin, tagging integration and simulation tests and making it trivially easy to run just the suites you want.

We’re also considering pulling out the metaclass and decorators into a small, reusable Python packaging, making it easy for others to make use of this pattern.

Weird bugs: Django, timezones, and importing from eggs

Every so often you hit a bug that makes you question your sanity. The past several days have been spent chasing one of the more confusing ones I’ve seen in a long time.

Review Board 1.7 added the ability to set the server-wide timezone. During development, we found problems using SSH with a non-default timezone. This only happened when updating os.environ[‘TZ’] to something other than our default of UTC. We’d see the SSH process (rbssh, our wrapper for SSH communication) break due to an EOF on stdin and stdout, and then we’d see the development server reload itself.

Odd.

Since this originated with a Subversion repository, I first suspected libsvn. I spent some time going through their code to see if a timezone update would break something. Perhaps timeout logic. That didn’t turn up anything interesting, but I couldn’t rule it out.

Other candidates for suspicion were rbssh itself, paramiko (the SSH library), Django, and the trickster god Loki. We just had too many moving pieces to know for sure.

So I wrote a little script to get in-between a calling process and another process and log all communication between them. I tested this with rbssh and with plain ol’ ssh. rbssh was the only one that broke. Strange, since it wasn’t doing anything obviously wrong, and it worked with the default timezone. Unless it was Paramiko somehow…

For the heck of it, I tried copying some of rbssh’s imports into this new script. Ah-ha! It dropped its streams when importing Paramiko, same as rbssh. Interesting. Time to dig into that code.

The base paramiko module imports a couple dozen other modules, so I started by narrowing it down and reducing imports until I found the common one that breaks things. Well that turned out to be a module that imported Crypto.Random. Replacing the paramiko import in my wrapper with Crypto.Random verified that that was the culprit.

Getting closer…

I rinsed and repeated with Crypto.Random, digging through the code and seeing what could have broken. Hmm, that code’s pretty straight-forward, but there are some native libraries in there. Well, all this is in a .egg file (not an extracted .egg directory), making it hard to look through, so I extracted it and replaced it with a .egg directory.

Woah! The problem went away!

I glance at the clock. 3AM. I’m not sure I can trust what I’m seeing anymore. Crypto.Random breaks rbssh, but only when installed as a .egg file and not a .egg directory. That made no sense, but I figured I’d deal with it in the morning.

My dreams that night were filled with people wearing “stdin” and “stdout” labels on their foreheads, not at all getting along.

Today, I considered just ripping out timezone support. I didn’t know what else to do. Though, since I’m apparently a bit of a masochist, I decided to look into this just a little bit more. And finally struck gold.

With my Django development server running, I opened up a separate, plain Python shell. In it, I typed “import Crypto.Random”. And suddenly saw my development server reload.

How could that happen, I wondered. I tried it again. Same result. And then… lightbulb!

Django reloads the dev server when modules change. Crypto is a self-contained .egg file with native files that must be extracted and added to the module path. Causing Django to reload. Causing it to drop the spawned rbssh process. Causing the streams to disconnect. Ah-ha. This had to be it.

One last piece of the puzzle. The timezone change.

I quickly located their autoreload code and pulled it up. Yep, it’s comparing modified timestamps. We have two processes with two different ideas of what the current timezone is (one UTC, one US/Pacific, in my case), meaning when rbssh launched and imported Crypto, we’d get a bunch of files extracted with US/Pacific-based timestamps and not UTC, triggering the autoreload.

Now that the world makes sense again, I can finally fix the problem!

All told, that was about 4 or 5 days of debugging. Certainly not the longest debugging session I’ve had, but easily one of the more confusing ones in a while. Yet in the end, it’s almost obvious.

Scroll to Top