Here is a gif from my latest aframe experiments for Lightbeam.
For Mozfest 2017, I had submitted the following proposal – ‘Lightbeam, an immersive experience‘. While the proposal is still being reviewed, I have been experimenting with Aframe and the above gif is an initial proof of concept 🙂
Here is the excerpt from the proposal:
What will happen in your session?
Lightbeam is a key tool for Mozilla to educate the public about privacy. Using interactive visualisations, Lightbeam’s main goal is to show web tracking, aka, show the first and third party sites you interact with on the Web.
In this session, the participants will get to interact with the trackers in the VR world thus creating an immersive Lightbeam experience. With animated transitions and well-crafted interfaces, this unique Lightbeam experience can make exploring trackers feel more like playing a game. This can be a great medium for engaging an audience who might not otherwise care about web privacy & security.
What is the goal or outcome of your session?
The ultimate goal of this session is for the audience to know and understand about web tracking.
While web tracking isn’t 100% evil (cookies can help your favourite websites stay in business), its workings remain poorly understood. Your personal information is valuable and it’s your right to know what data is being collected about you. The trick is in taking this data and shacking up with third parties to help them come up with new ways to convince you to spend money and give up more information. It would be fine if you decided to give up this information for a tangible benefit, but no one is including you in the decision.
Read this blog post to understand Lightbeam’s migration from SVG to Canvas.
Ignore the transforms and the inversions (this.transform.invert) in this post. Those are part of d3-zoom and explaining the math of this and d3-force is beyond the scope of this blog post.
mousemove event is registered on the canvas element itself.
The mouse <clientX, clientY> positions are re-calculated w.r.t the canvas’s bounding rectangles. This ensures the mouse coordinates are confined to the canvas’s area.
getNodeAtCoordinates(x, y) returns a node, if a node is present at the given <x, y> values.
D3’s force layout has simulation.find(x, y[, radius] which returns the node closest to the position <x, y> with the given search radius. I chose to write isPointInsideCircle() to find out if a node exists at the given <x, y> values. The intention here is to isolate the logic from D3 specific as much as possible.
When you hover over the canvas, and if the mouse coordinates are inside any circle, then there is a node present at these coordinates.
The point <x, y> is
inside the circle if d < r
on the circle if d = r
outside the circle if d > r
Square roots are expensive. Hence d is compared with r*r!
The tooltip has position: absolute.
Because of this property, there is a need to check the tooltips’ left property doesn’t exceed the canvas’s right property, else there will be horizontal scrollbar on the parent container because of overflow-x.
x+tooltipWidth >= canvasRight takes care of the overflow and sets left to x-tooltipWidth.
Setting left to x-tooltipWidth/2 ensures the tooltip arrow is centre aligned to the node.
If a favicon exists for a given node, then it is drawn.
The favicon is drawn at the centre of the circle (firstParty) or triangle (thirdParty).
A square that fits exactly in a circle has a side length of sqrt(2) * radius.
firstParty & thirdParty nodes
Given that we are drawing on a canvas, firstParty is a circle on the canvas. thirdParty is an equilateral triangle.
Given the centre of the circle is at <x, y>, r is the radius of the circumcircle and dr is the radius of the incircle.
zoom and drag
d3-zoom and d3-drag are used to achieve the zoom and drag behaviours respectively. It is quite complex when the two are combined. If you click and drag on the background, the view pans; if you click and drag on a circle, it moves.
d3-drag requires a dragSubject. I am using the same getNodeAtCoordinates(x, y) function which is used to show the tooltips and the logic remains same. This is how drag and zoom are combined for Lightbeam. If there is a node, (dragSubject) then it drags, else it pans.
Here is the d3-zoom implementation.
The tricky part here is the need to distinguish between two coordinate spaces: the world coordinates used to position the nodes and links, and the pointer coordinates representing the mouse or touches. The drag behaviour doesn’t know the view is being transformed by the zoom behaviour, so we must convert between the two coordinate spaces.
This is where transform.invert or transform.apply come into play.
I hope I have done justice to the math in this post!
Here are two quick updates about Lightbeam because I can’t contain the excitement to myself.
Lightbeam goes responsive, yayyyyy!!!
I am extremely happy for achieving this today. Making the UI responsive was there in our to-dos, but this one got done today accidentally in an attempt to answer few of the comments on one of my PR. CSS grid is used and I must say the fr unit is so handy.
Mozfest proposal submission
Proposed a session for the Mozilla Festival 2017, London. Here is the proposal. The coolest part here is that the google-form submission automatically gets created as a GH issue.
SVG is the preferred choice for D3. But when you expect a lot of DOM nodes (yes, in the case of Lightbeam) you need to start worrying about the DOM performance and have to take the call to step out of the SVG comfort zone. We chose Canvas over SVG for Lightbeam 2.0!
In this post, I would like to highlight the key points of drawing and interactivity on the HTML5 Canvas element.
In our case, we followed the first approach – using D3 solely for its functional purpose (D3’s force layout algorithm) and then drawing onto the canvas.
Normally, when you follow this approach, you are trading off D3’s super rich data binding and joining functionality. It means you’re only drawing your graph once – you’re not expecting new data that would require a graph redraw or update. This is why D3 with some dummy HTML nodes is a preferred way in order to retain the data binds and joins. D3’s joins are a way to dynamically update the graph without having to redraw the whole thing over and over again.
The dummy HTML element approach
Here is an example of using custom aka dummy HTML elements to render D3 on canvas.
custom is definitely not a standard DOM element type, so it will not be rendered in any way and will live only in memory (virtual DOM). It is used as a container for other dummy nodes.
How does Lightbeam update dynamically without D3 joins?
Even though we followed the first approach, Lightbeam has dynamic updates. Luckily, D3’s force layout algorithm takes care of the new node and link updates/additions and we managed to use D3 only for its functionality without the dummy element approach. Here is the PR.
Interactivity on the canvas
Canvas is a single DOM element and mouse interactions on the canvas can be sometimes tricky because you don’t have independent access to the nodes and links (or any graph element). In our case, we have the following interactions:
On hover over the nodes (websites) show tooltips with the name of the website
Drag the graph
Zoom in and zoom out the graph
Panning – shifting the view of the graph
At the time of writing this blog post, only tooltip based interactivity is achieved. PR
I shall explain tooltip based canvas interactivity in the next post.
I have never done performance testing using browser devtools. When we migrated from SVG to Canvas I was curious to know the performance results of using canvas. For a given small sample of websites, here are the test results. Canvas has a rendering score of 4.3 milliseconds 🙂
In this blog post I want to share my lessons learned from creating tooltips for the Lightbeam visualisations.
Visualisations are the major part of this Outreachy – Lightbeam internship. We use D3-force to create a simulation for an array of nodes (websites), compose the desired forces, and then listen for tick events to render the nodes as they update. When we started, the preferred graphics renderer was SVG. We now have a Canvas based implementation because this is more efficient than SVG. Here is an article why canvas is better than svg and here are the initial results.
Having said that, lets talk about tooltips.
SVG based solution
I had first implemented tooltips(text labels) on the SVG version. Here is the PR.
In SVG, you can render every node as a DOM element. This is why SVG isn’t the right solution for graphs with too many nodes, because it can easily slow down the browser’s page loading time due to too many nodes.
Since this is an SVG based solution, I chose text-labels over tooltips.
Tooltips vs text-labels
Tooltips come with an additional overhead of dynamically updating the x,y coordinates and the title. Text labels are rendered like ordinary DOM elements, the SVG <text> element, one for each node(website).
But using tooltips, you can cut down the number of text nodes. As opposed to having one text node for each node(website), there is one tooltip (DOM element) whose position and content is dynamically updated.
I shall discuss the tooltip based solution in the canvas implementation.
1) Text labels in SVG
By default, the visibility of text-labels is set to hidden.
All text-labels have a common class textLabel.
On hover, over the circular nodes, the corresponding text-label’s visibility is changed from hidden to visible.
For each node(website) there is a corresponding <text> element.
In order to identify and set the visibility of the right <text> corresponding to the hover on node(website), every <text> element is assigned an additional class text-i where i is the index corresponding to each node(website).
When node(website) with index i is hovered, the corresponding <text> with class text-i is set to visible.
2) Hover on the nodes(websites)
In order to achieve the hover effect, mouseenter and mouseleave event handlers are registered on each node(website).
After the initial implementation, there was a lot of flicker when the text labels were made visible on node hover. Although it isn’t very clear from the above gif, notice the time taken to display http://www.google.com on the third node. Looks like gifs aren’t good at capturing flickers (I need to find out a better solution), but the time delay to show the third label in the above gif === flicker effect.
1) mouseenter vs mouseover
At first, I thought the flickering was due to the mouse events. Later, I found out that these mouse events had nothing to do here. However, it is worth making a note of the above two mouse events. Link
Below is the DOM for the SVG version:
Since the nodes/circles are independent of the text labels, and have no direct descendant nodes, the mouse events had nothing to do with the flicker.
2) SVG and z-index
After wracking my brains for a while, I figured out the cause for the flickering of labels. If you carefully examine the above gif, you will notice that the lines are on top of the circles. On mouse hover over the nodes, it was mouseenter for circles and labels were made visible. But, when the mouse accidentally appeared on top of the lines (because you don’t realise you are doing so) it caused the mouseleave for circles and the visibility was getting toggled.
When I figured out this cause, the obvious solution to run in my head was to set z-index for circles to be greater than the lines. Despite this, the flickering was still there. On further investigation, this is what I found from the SVG specification:
z-index in SVG is defined by the order the elements appear in the document. You will have to change the element order if you want to bring a specific shape to the top. Drawing lines first, followed by drawing circles fixed the flickering bug.
It’s more than 2 weeks now. I haven’t had the time to write anything down. So here’s a quick post on what’s happening!
San Fransisco – Mozilla All Hands 2017
The week of June 26 – 30 was Mozilla All Hands 2017 at San Fransisco. It was my first time into the United States and I was pretty excited about this trip. After a tiring 11 + 3 hour long haul flight, when I presented my Mozilla invitation letter at the immigration, the border security officer at the SFO airport was excited to find out that I too was attending the All Hands. It appeared to me that he already processed other fellow attendees and he knew about it. His cheerfulness made me forget my tiredness and I had a warm welcome into the city. With the kind of ongoing news and rules about US airport security, I was too paranoid. But everything went smooth, from gaining B1 visa, carrying laptop in hand baggage and the security clearance itself.
The All Hands itself was a great experience. I and Bianca came up with a timeline of the next tasks that we could work on for until the internship period. Our main focus is to improve the graph performance. The idea is to use canvas instead of SVG. Here is an article I have written why canvas is a better choice.
I did a bit of sight-seeing at SFO via the big-bus tour.
One week at SFO was great. I wish to go back and explore more of the city at some other time.
Berlin – Home, away from home
store is an integral component of Lightbeam. We use asynchronous message passing for communications between the background and page scripts. Our mentor Jonathan helped us a lot with code improvements on this store.
I have used D3 in the past, but D3’s force layout algorithm is new to me. We are able to draw the Lightbeam graph using this algorithm, but there are so many more things to learn. I am sure, over the coming weeks our graph would see significant improvements.
This week I am working on to fix the node overlap/colliding issue and to get a canvas prototype for D3’s force layout.
Thanks to Bianca, my fellow Outreachy-Lightbeam project partner for coining the above title in her blog and asking me questions to explain about the bug(PR) I recently fixed for Lightbeam. I started explaining her last evening via quick diagrams and I already had an idea to make my next blog post colourful 😉
I have been inspired by Lin Clark’s code cartoons and Mariko Kosaka who always draw and make technical reading fun! Here is an attempt to write a post and explain Lightbeam’s store architecture via diagrams 🙂
To revamp Lightbeam into a web extension we decided to follow the following architecture:
A web extension consists of a collection of files and we employed the following directory structure to achieve the above architecture:
We were able to chalk out an MVP soon, everything behaved as expected, until we found this bug.
Clear the storage manually through the console: browser.storage.local.remove('websites').then(()=>console.log('removed')).catch(err=>console.log(err))
Also set store._websites=null manually through the console
Refresh the page – graph disappears
Refresh the page again, graph appears with old values
Thanks to Jonathan for pointing out that there were two store instances created and thus began – ‘tale of 2 stores’.
Following was the code implementation to achieve the above architecture:
The problem with this approach was that store.js file was loaded from two different places and that’s how we ended up having two store instances.
In other words, this was the problem:
Now to debug the above specified bug:
capture.js had one store instance because of store.js loaded via background script
lightbeam.js had another store instance because of store.js loaded via the script tag
index.html page was loaded and lightbeam.js got the required websites to display via getAll() from it’s store instance
Next, web extension storage was manually cleared via browser.storage.local.remove and store._websites was explicitly set to null
in the first page load, there was nothing to display because the getAll() returned nothing from it’s store instance (because of the above manual deletion)
on second page load, capture.js had done a setFirstParty() call to it’s store instance which in turn wrote to the browser.storage.local.set
capture.js is triggered when a tab (re)loads
capture.js‘s store instance had its copy of _websites already populated (when the manual store._websites=null was done, it was done on lightbeam’s store instance) and setFirstParty() used this already populated _websites value and wrote to the browser.storage.local.set
This is how the visualisations appeared back on the second page load
To fix this bug, we now have storeChild.js which acts as an interface, or is the only point to talk to the background store.js. With the child and parent architecture, the parent will always be in charge of when the write happens and so will serialise the writes which will behave like a locking database.
The page script storeChild.js talks to the background script store.js via message passing.