Danny Guo | 郭亚东

Development Friction

I decided to start documenting all the things that cause me to lose focus on what I want to be doing. I’ll probably include issues with tooling, bugs that take me longer to figure out than they should, and times when figuring out what the “right” thing to do takes more than a couple minutes of research. In that sense, this is also a log of things that I have learned.

Apparently, Google uses friction logs as well.

2018-12-12

Safari’s developer console truncates output

While working on a Safari App Extension, I was using console.log to debug, but some of my output in the developer console was getting truncated with ellipses. It appears that at least in Safari 12.0.2, truncation can occur when a string is provided after the first argument to console.log or if it’s a value in an object that is provided as any argument.

Hovering over the output doesn’t show the full string, and using the console’s “Copy Selected” feature copies the truncated version. So there doesn’t seem to be any way to view the full string.

From trial and error, it seems that truncation occurs once the string hits 140 characters (fun fact: that is the old limit for Tweets).

Safari truncating developer console output

I submitted a report on the Apple Bug Reporter.

2018-11-12

Xcode doesn’t have native live reloading

I’m used to setting up live reloading for all my projects, but for a Safari App Extension, that doesn’t seem to be possible. Even for iOS development, it seems like you have to use a plugin.

Xcode’s settings are stored in binary format

This is my first time using Xcode. I usually try to store as much of my system configuration in my dotfiles repo, but Xcode doesn’t make it easy. Its settings are stored in a binary property list format. While I could still check this in, it would be easier to work with in a plain text format, like JSON.

Safari doesn’t detect extension changes during development

While working on a Safari App Extension, it took some trial and error to figure out how to get Safari to actually pick up a change to the code. The only reliable way I found was to clean the build folder every time. Simply restarting the project wasn’t good enough. Sometimes, the extension would even stop showing up in Safari preferences.

2018-08-26

I tried creating a symbolic link with a command that looked like:

$ ln -s ../foo.css bar/foo.css

But it didn’t work. I learned from this Stack Exchange answer that the target in a symbolic link is used as-is. If the target is an absolute path, the symbolic link will work as long as the file is not moved. If the target is a relative path, the relative path is determined from the location of the symbolic link, not from where you issue the ln command, which was my wrong assumption. The benefit of this approach is that the symbolic link will remain valid as long as the relative path remains valid.

2018-08-13

TypeScript’s watch mode has to do an initial compilation

My start script for development cleans the dist folder that holds the compiled files, runs tsc in watch mode in the background, and then starts the server using pm2, which watches the dist folder. This presents a small issue where initially running the script causes a lot of trace because pm2 tries to start the server before the files are done compiling. pm2 prints an error and then repeatedly tries to start it, printing an error each time until tsc is done.

There is an open GitHub issue, and I added the workaround of doing a normal tsc compile first. While this does get rid of a decent amount of the trace, there is another issue of pm2 restarting each time a file is recompiled by the second invocation of tsc. This wouldn’t be an issue if there was a way to disable the initial compilation when running in watch mode, as suggested by the GitHub issue.

This problem is exacerbated by the need to clean the dist folder in the first place because TypeScript doesn’t do it. See this GitHub issue. Without manually cleaning it, old files would get left in the folder.

Another related issue is that pm2 immediately tries to restart the server, which causes all the trace because each new file triggers a restart. It would be nice if the program could be smarter during development. pm2 also repeatedly tries to restart when there is a tsc error. Ideally, it would wait for me to make a change before trying to do so. I haven’t looked into the root cause for this particular issue yet.

2018-08-09

Yarn doesn’t handle @types packages well

When I tried to update React, I also tried to update the @types/react TypeScript definitions. Unfortunately, the resulting node_modules structure caused a TypeScript compilation error. I got a lot of trace that looked like:

TS2320: Interface 'Element' cannot simultaneously extend types 'ReactElement<any>' and 'ReactElement<any>'.
  Named property 'type' of types 'ReactElement<any>' and 'ReactElement<any>' are not identical.

I found this GitHub issue, and this StackOverflow question. The probelm is that nested instances of the React types are created. The issue apparently doesn’t happen with npm. It doesn’t seem that there is an obvious workaround with Yarn. I have run into this issue before and was able to fix it by blowing away my node_modules and Yarn lockfile. But that didn’t work this time, and it’s a suboptimal solution anyway because I’ll inevitably pick up many non-React related subdependency updates, which could break something.

I assume React + TypeScript + Yarn isn’t that uncommon of a combination, so I’m surprised that the GitHub issue hasn’t received more attention and thumbs ups. I may have to switch back to npm just because of it.

As a side note, TypeScript in general is great, but one of the biggest downsides is that the type definitions are frequently developed in DefinitelyTyped rather than being bundled with the package itself. It’s a huge kludge that usually works well enough, but I hope that in the near future, more and more developers choose to write their projects in TypeScript to begin with.

Hashids.js doesn’t work with ES6 imports

I tried using Hashids, but I got an error when I tried to include the JavaScript version in a TypeScript project. I got this error when I tried to import it:

Uncaught TypeError: HashIds.default is not a constructor

I found this GitHub issue, which links to a more in-depth issue. The issue is that the project exports its functionality in a strange way. As a workaround, I changed the import to a require.

2018-07-12

Deprecation warning in Vim

When I start Vim on my Mac, I get this error message: /must>not&exist/foo:1: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses. I found this GitHub issue. The problem occurs with Python 3.7. I’m just going to wait for it to be fixed upstream since it doesn’t actually seem to affect my Vim usage.

2018-05-04

Files created by Docker containers in mounted volumes are owned by root

For browser extensions, Mozilla requires developers to upload source code for review when the extension code is obfuscated or minified. To make it easier, I have a Python script in my project that first packages the extension using web-ext and then creates a ZIP of the source code using the git archive command. This was working fine on my Mac, but when I tried it on my Linux laptop, the git archive command failed because it lacked permission to create a file in the web-ext-artifacts directory that was created by web-ext. The issue was that the script ran web-ext inside a Docker container, which created web-ext-artifacts with root as the owner.

I found this GitHub issue, and one of the comments suggested setting the user with the --user flag when running the container. I tried it, and it fixed the issue. So the Python script grabs the user id and group id of the current user and passes them to Docker. Now all the files and directories are owned by me rather than root, and the Python script can successfully create the ZIP.

I was curious why the issue didn’t occur on my Mac. I read the documentation for osxfs, but I found this Stack Overflow answer easier to understand. The gist of it is:

Inside the container, the osxfs driver pretends that whatever uid/gid is running in the container is also the uid/gid that owns the mounted files.

If you chown files (in the container) to something else, no chown is performed on the real files; this owner information is stored in an extended file attribute instead, and that value is used by the container (and ignored by macOS).

The real files are owned by whoever owns them in macOS, outside the container. Access control is determined using those real file ownerships, and the uid/gid of the user running the Docker applications (which is probably your login user on the Mac).

This GitHub issue comment explains that the point of this is 1) to not require root on the host if the container creates files that are owned by root from its perspective and 2) to make sure that the container can access any files that the host user can access and has shared with the container through a volume mount.

2018-04-29

Can’t empty an AWS S3 bucket through the web interface

I have a bucket leftover from an old project. I have tried to empty/delete it before, but the operation never succeeded. It incurs less than ten cents in costs, and AWS doesn’t seem to charge for tiny amounts (presumably because of payment processing fees), so I never bothered to try emptying it from the CLI. It’s just interesting that a couple years later, I still can’t empty the bucket from the web interface. When I press “Empty bucket,” a progress bar immediately appears saying “100% Sucessful.” But when I open the bucket, it’s still filled with files. Furthermore, clicking on the progress bar opens a popup that lists how many objects are being deleted. It goes up by 1000 every few seconds, which seems to contradict the progress bar. The title also has “Completed” in it. Eventually, the total objects number stops going up. I refresh and open the bucket. It’s still filled with files.

Deleting the bucket is different. It actually says “In progress” when it’s..in process. But it eventually hangs too. Maybe part of the problem is that there are hundreds of thousands of files in the bucket. I just expected better from the UI.

2018-04-14

TypeScript’s watch mode watches too many files

I ran into a baffling issue where the TypeScript compiler’s watch mode would trigger recompilation whenever I changed seemingly irrelevant files like package.json or a JS file in node_modules. Watch mode is documented to only watch “input files”. After some debugging, I realized that removing almost all of my dependencies fixed the problem. I tried doing binary search to figure out if a particular package was causing the issue, but I couldn’t narrow it down to just one package. I ended up with a set of dependencies that reproduces the problem. Removing any of them seemed to fix the problem. What was particularly strange was the fact that the input files didn’t even reference any of the dependencies.

I made a repo demonstrating the problem and submitted an issue.

2018-03-14

The “prepublish” npm lifecycle event is confusing

I want to make sure that I run certain commands before publishing to the npm registry. I know npm supports lifecycle events in package.json. The prepublish one seems to be what I’m looking for, but apparently it’s problematic. The issue is that it was implemented to run not only before publishing but also during development after the user does npm install. This is incredibly unintuitive and seems to have caused a ton of issues and discussion. There is even an in-publish package that is used as a workaround and has been downloaded millions of times.

The npm team finally decided to resolve the issue, but the solution is pretty painful too. They added a prepare event that acts like prepublish (running after npm install too) and a prepublishOnly event that actually only runs before publishing. The plan is to issue a deprecation warning for prepublish, change it to act like prepublishOnly in a new major version, and then eventually deprecate prepublishOnly. I assume they put a lot of thought into it, but I don’t understand why they couldn’t skip the prepublishOnly part and just directly change the prepublish behavior in a new major version.

Another related issue is that the Yarn team has to keep up with the changes. It’s unfortunate how much effort has been spent on this issue when it seems like it could have been avoided by just not making something act in a way that isn’t implied by its name at all.

2018-03-13

JavaScript modules with TypeScript

I’m following this Twilio blog post on how to write a TypeScript package than can also be used by regular JavaScript. It’s smooth for the most part, but figuring out how to correctly import and export code takes some research.

I have to pull in another package (ms), and I want to use the now official ES6 import statement, but the generated code throws an error when ms is called. It turns out I need to use this strange, TypeScript-specific syntax to interop with CommonJS modules: import ms = require('ms');. It fixes the issue though.

I only have a single function to export, so I want to use export default. TypeScript happily compiles it, but when I try to require the output in regular JS, it fails. I’m running into this issue, which is that export default foo; becomes exports.default = foo;. So in Node, the user would have to do const foo = require('foo').default;. I decide to drop the default and just export the function. So an ES6 consumer would do import {foo} from 'foo';, and a Node user would do const foo = require('foo').foo;. It still looks clunky, but it’s better than default.

Then I learn that for importing at least, TypeScript 2.7 added a new esModuleInterop flag that allows for importing CommonJS in the way that I originally wanted to: import ms from 'ms';. I add the flag, and it works.

As I browse numerous GitHub issues and StackOverflow questions, it makes me wonder how much collective time we have spent on figuring out how to import and export code in JavaScript. I’m really glad ES6 was able to provide an official solution, but it seems like we’re going to be dealing with compatibility issues for a long time. Node still doesn’t support import natively because it’s such a hard problem. I’m thankful to the people who are working on making the whole situation as painless for end users as possible.

The package that I published is sleep-ts.

2018-02-15

“RESTful” Partial Updates

I want to follow convention when it comes to partially updating resources through a REST API. I know that POST is used to create resources, and PUT is used to update resources at a known location, but I don’t want to send the entire object for updates. Instead, I only want to send the values that should be changed. I learn about PATCH, but it’s not as simple as I imagined it would be. I read this blog post, which tells me that PATCH is supposed to send instructions (i.e. {"op": "replace", "path": "/email", "value": "newemail@example.com") for the update rather than just values (i.e. {"email": "newemail@example.com"}).

I’ve never actually seen the former. In API documentation, I’m used to seeing the pattern of “send the values that you want to be updated, and those are the only ones that will change.” That’s how the Stripe API works. It uses POST for updates. The Twilio and Lob APIs also use POST for updates, which seems to go against this guide, though Roy Fielding apparently doesn’t have a problem with it. I can’t find an API that actually uses PATCH, except for Google APIs.

I find another blog post that summarizes the topic. I think I will just use PATCH but not bother with the “official” approach of using a description of changes. My website is not complicated enough to justify that. I am following the informal approach that is detailed here. It is the Stripe/Twilio/Lob approach but uses PATCH instead of POST.

PUT is not in the conversation because it is supposed to be idempotent. And while PATCH might seem idempotent, it is not because the rest of the resource might change in the meantime, or the changeset might depend on the resource already having certain values.

Time spent: ~1 hour, 30 minutes

2018-02-08

Localhost stops working

All of a sudden, I can’t access my website that is running locally. The Docker Compose output looks good, and the port mapping looks fine when I run docker-compose ps. I try restarting the service, then restarting Docker (sudo systemctl restart docker), then restarting my computer. But it’s still not working. I try to access the site in Chrome, and it does work! I try curling, but that hangs, just like Firefox. I try starting another project in Docker on a different port, and the results are the same. I try serving up a folder with Python (python -m http.server for Python 3), and it’s still only accessible from Chrome. I decide to inspect my /etc/hosts file, and it’s empty. I add 127.0.0.1 localhost to it, and everything works again. I’m not sure how that happened though since I didn’t update anything that I know of. I also don’t know why Chrome was able to access localhost regardless.

Looking back, I should have immediately tried the Python server to verify that it wasn’t a Docker issue. That should have led me to checking out /etc/hosts more quickly rather than spending time restarting things.

On the bright side, I spent a bit of time learning the basic difference between 127.0.0.1 and 0.0.0.0. Essentially, 127.0.0.1 will only serve clients on the same host, whereas 0.0.0.0 will serve clients on any host (like another computer if they know your IP address).

Time spent: ~30 minutes

2018-02-06

Colorized output with ls

On Antergos, I’m not getting colorized ls output (different colors for directories, executables, etc.). I find a StackExchange answer. I need to add --color=auto to my ls alias, but that flag doesn’t exist on Mac (where the -G flag is used instead to turn on colors, which I assume is because OS X is based on BSD, not Linux). I find another StackExchange answer that tells me I can do export CLICOLOR=1 to achieve the same effect on Mac. I update my dotfiles to set that when I’m on Mac and add the color flag otherwise.

Time spent: ~10 minutes

No type checks with pm2 and ts-node

I’m not getting type errors for the server during development, and a type error made me spend five minutes debugging when it should have been caught immediately. It looks like it’s because ts-node changed to “fast” mode by default, which bypasses type checks. pm2 has a way to pass flags to the interpreter (I need to pass –type-check). But when I add it to the config, the server doesn’t start anymore, and I don’t see any pm2 trace in the Docker output. I open a shell and run pm2 manually. That’s when I see [PM2][ERROR] Interpreter /app/node_modules/pm2/node_modules/.bin/ts-node does not seem to be available. I get that error regardless of whether or not I add interpreter_args to the config, but the server still starts if I don’t add it. Very weird. I’m going to try updating pm2 to the latest version, though I don’t see aything in the changelog that looks like it would help. And it doesn’t.

I find a GitHub issue with that pm2 error message, but it suggests to run pm2 install typescript, which I don’t want to do because I want it to just use the versions of TypeScript and ts-node that I already have installed as dependencies, and I can’t find any documentation that details what pm2 install typescript actually does. The help output says install or update a module and run it forever, but how is TypeScript a pm2 module?

I’ll try specifying the interpreter with "interpreter": "./node_modules/.bin/ts-node". That doesn’t work. The error message is /usr/local/bin/node: bad option: --type-check. But why is node getting the option if I overrode the interpreter?

It’s a shot in the dark, but I’ll try passing the flag as args rather than interpreter_args. The server starts, but there’s still no type checking.

Taking another look at the ts-node documentation, I learn that I can also turn on type checking with an environment variable, so I add "TS_NODE_TYPE_CHECK": "1" to the pm2 config. And that works! Type errors are being reported again.

Time spent: ~1 hour

2017-10-12

Jest is very slow in Docker

I added Jest to a codebase to test React components, but even a single, trivial test took over a minute to complete. I had the :cached volume mount option set, but Jest was still very slow. I found this GitHub issue. Unfortunately, it was closed due to being perceived as a Docker issue. Nevertheless, it seems like a Jest issue to me because Mocha doesn’t have nearly as bad of a performance issue in Docker.