aurora-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Kai Huang <>
Subject Re: Redesign of the Aurora UI
Date Wed, 19 Jul 2017 20:56:03 GMT
Just a few thoughts as an aurora developer and operator:

>From my experiences with Aurora users, some persisting complaints are:

  1.  The current UI is not very intuitive for the users to understand the task lifecycle,
resource utilization of their job.
  2.  Often times users are unaware of the new features/changes in Aurora Scheduler/Executor,
which leads to a lot of misuse of the system.
  3.  Users have preferences on the appearance of the scheduler/thermos UI due to special
use cases, and ask us to customize for them(or start to write their own UI, which is often
not recommended).

The other major issue I see in the current UI is that it's built on an obsolete tech stack(AngularJS)
that has all the binaries and dependencies in the repo. From a developer's perspective, it's
a big burden to maintain/test the code, and make fast iterations on it.

Currently the scheduler UI is readonly, and mainly designed for debugging purposes. We could
have done much better to make the UI more friendly to the end user, empower them to discover,
understand and use all the Aurora features, and give them more insights into the their jobs,
or even the entire cluster. I love the idea that redesign the UI with a modern stack, and
more importantly, every single part of the application is a module that you could take your
own customization.

As for the development strategy, I'm in favor of the incremental approach that posts one page
at a time. The main benefit is that we are educating the developers while iterating on it,
and this will improve the adoption rate in the long term.

Overall I'm very interested in this work, and would like to collaborate with David to redesign
and improve the UI.

From: Joshua Cohen <>
Sent: Wednesday, July 19, 2017 11:39
Subject: Re: Redesign of the Aurora UI

I think this looks great overall! I'm super excited to see the UI get some
love (and to set us up for better iteration on the UI going forward). My
biggest concern, of course, is the current hubbub vis-a-vis Apache and the
BSD-3+Patents license[1]. Have you tried running this against Preact to
confirm compatibility/performance? Also note that other Facebook libraries
have the same license problem (e.g. Immutable, Jest), so unless FB changes
their patent grant clause, I imagine we'd have to find alternatives to
those as well. If only we had landed this a week ago, we could've been
grandfathered in on the license front :(.

As far as the options for landing this, I'll leave that up to more active
reviewers, but my gut says that smaller reviews will make this easier to
parse, especially for those unfamiliar with React. That said, perhaps we
could go with an alternate method for reviewing here, where people review
against your fork directly and only when they're comfortable do you post
the whole patch to reviewboard for what should, by that point, be a rubber
stamp review?

In general, fantastic work, David!


On Wed, Jul 19, 2017 at 12:55 AM, David McLaughlin <>

> Hey all,
> At Twitter we have had a long-standing desire to be able to put custom
> widgets and other UX enhancements into the Aurora UI. Recent prototype work
> to do this in a clean way has proved fruitful and I'd like to present this
> approach to the community and get feedback on the overall approach.
> The basic approach is simple:
> 1) Use the node plugin for gradle to bootstrap a modern web development
> build system using webpack and npm.
> 2) Use a modern JS view library like ReactJS (or Preact depending on the
> Facebook+Patents license issues for Apache projects) to have all UI
> components defined as ES6 modules.
> 3) Use webpack to set up a custom plugin directory that modules should be
> loaded from first. This would allow anyone with custom Scheduler builds to
> also drop in custom JS files to replace selected OSS UI components.
> Example, you could drop a "Navigation.js" into the plugin directory, and
> the build system will use that module instead whenever it sees a
> <Navigation /> used by parent components. Basically: it's a hybrid
> plugin/dependency injection (similar to Guice) approach to customization of
> the UI.
> The nice thing here is that since every single part of the application is a
> module you could take your customizations to any level of precision that
> you need. You could replace (or add) key pages or components, or just keep
> it as simple as adding some helper text to one of the pages.
> Some use cases this could enable:
> * A <ConfigurationViewer /> widget that you can replace with configuration
> parsing that understands special processes and/or custom executors.
> * Integration with your telemetry stacks to add resource utilization
> feedback or other performance indicators into the job page.
> * Adding custom stats tracking widgets for internal analytics.
> Some points from my prototype work:
> 1) Switching to npm for JS asset management allows us to delete the vast
> majority of our 3rdparty assets that we had to copy into the repo. We can
> most likely extend this to remove all of them, including our CSS frameworks
> like Bootstrap.
> 2) ReactJS was proposed before (by Joshua Cohen) and ironically I was one
> of the main voices against. The argument I used at the time was lack of
> familiarity with the stack within the community, as well as concerns about
> the complexity. So what changed? Almost the entire Aurora team at Twitter
> has ReactJS experience now - from working on our internal Continuous
> Delivery system. I also believe that as a view layer, ReactJS is
> conceptually very simple and that most of the complexity in these modern JS
> applications is in a state layer like Redux. The Aurora UI is currently
> read-only and relatively simple, and we can avoid much of the complexity by
> just using component state instead.
> 3) Unit tests are trivial to add as part of this work.
> 4) I'd like to move to modern, maintainable CSS tooling - SASS - as part of
> this.
> My prototype is (hastily) published here:

aurora - Mirror of Apache Aurora

> Some highlighted changes:
> 1) Bootstrapping the modern UI toolchain from gradle:

aurora - Mirror of Apache Aurora

> L459
> 2) The webpack configuration to enable loading from a plugin directory (and
> SASS):

aurora - Mirror of Apache Aurora

> 3) The Hello, World of the plugin mechanism:

aurora - Mirror of Apache Aurora

> components/Hello.js
> Is replaced by:
> js-build/plugin/js/components/Hello.js
> 4) Unit test example:

aurora - Mirror of Apache Aurora

> components/pages/__tests__/HomePage-test.js
> My goal now is to take the prototype to completion and upstream to Aurora
> OSS. If the community is in favor, we should also discuss how to handle the
> rollout.
> I would propose two options:
> * a massive code-dump with the entire UI rewritten in ReactJS.
> * an incremental approach where I (for example) post one page at a time,
> allowing those unfamiliar with JS to see the step by step process that can
> demystify the tooling. The new UI would not be enabled incrementally due to
> the issues with "greedy" one-page frameworks.. instead we'd just enable it
> at the very end. The major downside here is just maintaining all the weight
> of two UI frameworks, etc. in that period of development.
> Please let me know any questions, concerns or feedback.
> Cheers,
> David

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message