labelflow VS Snowpack

It is launch day for us at LabelFlow, the open platform for image labeling, would be great to get your feedback on this major update for us.

We are hosting image datasets on our platform and until recently the stored datasets were relatively small (several hundreds of images, few GB) so we only offered the possibility to export zip files containing images and labels in the COCO or YOLO format. As the average size of the datasets is growing, it's not convenient anymore to export a zip.

SWC in NextJS is still in canary with experimental settings, but it took me 3 lines of code yesterday to make it work on a fairly large app ( https://labelflow.ai ). Hot reload times instantly went from 10s to 1s. Twitter discussion here https://twitter.com/vlecrubier/status/1448371633673187329?s=...

Overall I’m pretty bullish on Rust tooling and integration within the JS/ Wasm ecosystem !

Check labelflow.ai. It's free, the code is published, web UI is super simple and the images do not need to be uploaded on remote servers so you get started in no time. For classification you would press the 1 key if image has hotdog else right key to go to the next image. Not gonna lie, you're going to need a bit of time for 10k images but definitely doable alone on a simple use case like that. To be fully transparent, I work there! Classification features are still in beta they will be released in 2 weeks. Happy labeling!

I’ve used storybook for 4 years in teams of 1-15 devs and I’d say it’s a must have for any serious react app with 3+ full time developers. It has its rough edges sure but the ROI is 10x nonetheless in my experiences.

Advantages

- Testing components in isolation forces some good practices and allows to keep the codebase in check by encouraging good practices (limited coupling of unrelated parts of the codebase

- It’s super productive because it is both a form of unit tests, useful during development of UX in « TDD mode », and a very good documentation of your UI components. It greatly reduces the effort needed for both these aspects.

- For DX, the hot reload is generally faster in storybook than in the App (except if you use vite/snowpack in your app, so far..) because reloading a single component is faster than reloading the whole app and its state. In a large CRA our hot reload could sometimes take up 1min in complex cases, while storybook was taking 3s.

- Coupled with Chromatic (their hosted platform) and its GitHub integration it makes QA and visual regression testing a joy, 10x faster than alternatives, I really recommend that.

- It allows to share/iterate easily your ongoing developments with non-tech people in your organisation at early stage. A very good bridge between Figma and the final UI. A good support during Daily meetings about UI, just shared the deployed story url to ask for feedback.

Drawbacks

- It has its own Webpack config. So if you have a custom Webpack config in your app (don’t do that anyway, unless absolutely necessary) then be prepared to duplicate the customizations in your storybook config

- Global React Contexts needs to be duplicated in your storybook config and, if necessary, configured for individual stories. For example if your signup button changes based on an Auth status stored in a global context, then you will have to use Story.parameters to customize the content of the Auth context.

- We had a couple instances where storybook was the limiting factor for us to embrace some new/fancy tech, like yarn v2 or service worker. However maybe that’s a good litmus test: things that storybook support are state of the art JS and generally safe to use. Things that storybook does not support out of the box will cause you problems with other tools anyway: if it’s not storybook, some other tool like Cypress, Jest, Next, or some browsers will cause you trouble with your “shiny new tech”

- It can be slow to startup. We had a storybook with 300+ complex stories and it took 5min to startup and 10min to build in the CI

- It had some API changes/ migration pains a couple years back. However I think the new API is very good and will last a long time so this is behind.

Overall I definitely advocate to use storybook, especially with Chromatic, the ROI is 10x. If you find yourself limited by it in 2021 despite configuring it, maybe question your own tech stack.

Don’t try to implement your own storybook copycat (we had a colleague develop an alternative https://github.com/remorses/vitro , but i think it was not worth the effort)

If you want to see a state of the art repo in NextJS that uses storybook extensively with some customizations, check https://github.com/Labelflow/labelflow/

As a matter of fact, LabelFlow uses a service worker exactly to avoid sending your data to a server (your data is stored in the local service worker instead). The code of this service worker is there: https://github.com/labelflow/labelflow/blob/main/typescript/web/src/worker/index.ts . You won't find any privacy-defeating stuff in there. It's super simple.

What was then just a landing page is now a product that you can try for free with no login required, the code is also publicly available on GitHub. (https://github.com/Labelflow/labelflow/).

4 months ago we announced Labelflow (https://www.labelflow.ai/), the open image annotation and dataset cleaning platform.

What was then just a landing page is now a product that you can try for free with no login required, the code is also publicly available on GitHub. (https://github.com/Labelflow/labelflow/).

In this first version, we are releasing your most wanted features: a straightforward online image annotation tool. For privacy concerns, your images are never uploaded to our server! You can create bounding boxes, polygons, export labels to COCO format and we added plenty of keyboard shortcuts for productivity!

We’re excited to hear your feedback, tell us what features would make your life easier (https://labelflow.canny.io/feature-requests) and upvote what you would like us to build. Stay tuned, It’s just the beginning of a long story.

Then there was Webpack which seemed like it would be around for a while. Even after things like Parcel and Snowpack came on the scene people still recommended Webpack. I mean, it's still the backbone of create-react-app. An then, Vite was released.

Unbundled development utilizes native ESM support in browsers to offer an ultra-fast development experience. Unlike a traditional bundler which bundles everything in development, unbundled development transforms the code and rewrites import paths to the ESM compliant file path without bundling your code. Additionally, most bundlers that do this pre-bundle dependencies because that decreases the number of imports needed, and dependencies are unlikely to change often. The two most prominent bundlers that utilize unbundled development are Vite and Snowpack. Snowpack, created in 2019, was the first bundler to have an unbundled development experience. However, while Snowpack was popular for some time, this did not last forever. In 2020, the team behind Vue created Vite. Vite has many advantages over Snowpack, like the ease of use, speed, better optimization, and more. Additionally, popular projects like SvelteKit adopted Vite instead of Snowpack. All of this helped Vite pass Snowpack in downloads, and it now has more than 10x downloads compared to Snowpack. In fact, even Astro, a project created by the team behind Snowpack (be on the lookout for an article about Astro), is now using Vite. Overall, if you want fast, unbundled development, I recommend Vite.

I've been hearing good things about snowpack and have been wanting to give it a go myself as an alternative to webpack/babel

So how did we get here? Well, it's a good thing to describe alternatives considered when you add issues to a repo! Months ago, @0vortex described in Issue #1131 some opportunities for dependency updates that would require version 5 of webpack with our webpack configuration, and warned that the dependency management would probably be tricky. I fixated on an alternative that he mentioned about converting the project to use Snowpack. I had wanted to learn more about bundling tools, so I took a few days here and there after Thanksgiving and got Open Sauced mostly working with it (see PR #1320).

I have written a post about setting up a React project using Parcel as a bundler which can give a more detailed walkthrough about getting a React application up and running from scratch. There are multiple ways to do this but some of the more common build tools include Webpack, Parcel, and Snowpack.

For more information about this issue, check these issue tickets: Issue 2998, Issue 3219, Issue 3243

The number of transforms that Babel is doing with an "evergreen" config ("last 2 browser versions") at this point is effectively miniscule. It's a massive toolchain for what increasing turns out to be a minimal amount of actual work. "Last 2 Browser Versions" is effectively everything through ES2019 at this point which covers almost all of the "modern JS syntax". If you aren't using custom transforms you might not be transforming anything that matters in Babel in 2021. I've seen a bunch of projects with huge Babel pipelines where the only actual transform was Typescript's type removal and at that point, if your codebase is entirely Typescript, Typescript has all the downlevel transforms you need "baked in" (and arguably a little bit cleaner and simpler to Babel's kitchen sink but also still somehow millions of plugins approach) and it's just setting Typescript's compile option to the ES level you are most comfortable with. (In 2021 that may even be as high as { "target": "es2019" } or higher in your tsconfig.json and even there Typescript's not going to even need to downlevel much.) Typescript can also transform TSX/JSX to JS without the need of Babel, if you are using React.

Even ES2015 modules which some people still think is the big reason to keep Babel around: a) has full Browser support if you use type="module", but most people still want to pack their JS because just about no one is assuming HTTP/2 or HTTP/3 yet, and b) Babel has never done module format transforms, that's always been the domain of your packer (webpack, parcel, rollup, snowpack, what have you).

If you are updating your project stack in 2021 right now my personal top recommendation is that I really like the approach of snowpack (https://www.snowpack.dev/): ES2105 modules with dev experience (which is great), great Typescript support, and a simpler overall config experience than most other options right now. (It uses esbuild under the hood rather than babel for dev and basic transforms/bundling. It can optionally piggy back webpack and parcel for Production bundling that needs more "power".) Especially that <script type='module"> dev experience feels great now (with Hot Module Reloading too) versus waiting for a full bundle even for dev builds.

SvelteKit uses Vite under the hood, which is quite surprising, as Sapper and most tools are developed using Snowpack. Vite 2 is framework-agnostic and designed with SSR at its core.

I've had some success with npm, snowpack, mocha, typescript as that sort of stack for more "vanilla" efforts that feel rather more "modern". I think mocha is easier and cleaner than jest. I like keeping all of my transpilation to just Typescript without needing a massive Babel install/pipeline. snowpack (https://www.snowpack.dev/) right now I think is in a sweet spot of a better "ES Module native" developer experience than webpack and has better defaults when left unconfigured. (So much so that while there are snowpack templates/generators provided by the project I mostly don't use them other than for reference.)