For Protovis Users

D3 and Protovis have the same primary author (Mike Bostock), so it’s not surprising that these two systems take a similar approach to visualization. However, there are plenty of important differences, too; enough that it made sense to start anew, rather than patching the design of Protovis. Many of these changes were influenced by observing users’ successes and struggles with past approaches.

D3 and Protovis also share the same goal: to enable you, the web developer (however experienced you may be), to build custom visualizations in the browser with a minimum amount of effort. Not “zero” effort, certainly—but at least to alleviate the repetive burden of common tasks, while retaining the expressiveness needed for custom designs. D3 is not a charting library!

Where D3 and Protovis differ is the type of visualizations they enable (the what), and the method of implementation (the how). While Protovis excels at concise, declarative representations of static scenes, D3 focuses on efficient transformations: scene changes. This makes animation, interaction, complex and dynamic visualizations much easier to implement in D3. Also, by adopting the browser’s native representation (HTML & SVG), D3 better integrates with other web technologies, such as CSS3 and developer tools. Our hope is that this makes D3 not only more powerful, but easier to use.

So, What’s Different?

Native Representation

The first difference you’ll notice is that you specify elements as HTML or SVG directly, rather than using an intermediate (proprietary) representation. For example, a Protovis bar is specified using the pv.Bar mark type:

1 vis.add(pv.Bar)
2     .left(0)
3     .top(25)
4     .width(80)
5     .height(20);

In D3, the equivalent representation is just SVG’s rect element. In fact, this is what Protovis uses under the hood!

1 vis.append("svg:rect")
2     .attr("x", 0)
3     .attr("y", 25)
4     .attr("width", 80)
5     .attr("height", 20);

D3 can create any type of element the browser supports, and set any attribute or style property. This offers more expressiveness in terms of graphical output: for example, you can use dashed strokes, coordinate transforms, gradients, clipping and Bézier curves, oh my! D3 also offers more flexibility in how you represent your scene, which can dramatically improve performance for larger datasets. The web abounds with reference materials on HTML5, CSS3 and SVG that you can learn and apply using D3. Your knowledge of these standards will be relevant even if you decide to use an alternative to D3 down the road.

The native representation improves the development process by complementing your browser’s other technologies. One big convenience is that static styles can be moved out of code and into a stylesheet. For example, if you set attr("class", "snazzy") on your elements, you can then apply a gradient:

1 .snazzy {
2   background-image: -webkit-linear-gradient(#aad, #556);
3 }

Unlike Protovis’ intermediate representation, D3’s output can be inspected directly using developer tools, making debugging much easier. (Paul Irish has some helpful tips on getting the most out of the inspector.) You can even use the browser’s console to execute D3 commands interactively, watching how they transform the document. D3 selections are simply arrays of DOM elements!

A subtler advantage of native representation is that selections can be requeried from the document, across hierarchies, at any time. In Protovis, in order to modify the visualization, you need to modify the underlying data or property definitions, and then render. This requires a reference (e.g., var) to every mark in the scene. Furthermore, re-rendering is expensive in Protovis because the entire mark hierarchy (or subtree thereof) is first evaluated and then rebuilt as SVG. In contrast, D3 lets you touch only the elements and attributes that need changing. This not only improves performance, but makes arbitrary modifications easier. (Local variables were introduced to Protovis to ameliorate this problem, but ultimately proved too confusing and magical to be widely useful.)

Of course, Protovis’ specialized representation has advantages, too! Anchors and wedges have no trivial equivalent in SVG. Thus, it can require more code to accomplish the same task in D3 as compared to Protovis. However, helper classes, such as the d3.svg module, mitigate this additional effort. These helpers work with SVG path elements, setting the “d” attribute:

The SVG g (grouping) element serves a similar to purpose to the Protovis panel; the “transform” attribute can translate, scale and rotate child elements. SVG’s default coordinate system places the origin ⟨0,0⟩ in the top-left corner. To emulate the right and bottom positional properties of Protovis, use a transform to flip axes, or invert the attribute defintion (e.g., right = width - x).

Transformation

The next big difference is that D3 code describes transformations of scenes (scene changes), whereas Protovis describes representations (the scenes themselves). For example, to create a set of labels in Protovis:

1 var l = vis.add(pv.Label)
2     .data([4, 8, 15, 16, 23, 42])
3     .text(String);

If the data changed, Protovis would automatically re-evaluate all properties, and add or remove labels as necessary, on re-render. D3, in contrast, requires more explicit instruction. However, by specifying transformations explicitly, you can control which elements are added or removed, and what happens to them:

 1 // Update…
 2 var l = vis.selectAll("svg:text")
 3     .data([4, 8, 15, 16, 23, 42])
 4     .text(String);
 5 
 6 // Enter…
 7 l.enter().append("svg:text")
 8     .text(String);
 9 
10 // Exit…
11 l.exit().remove();

This approach eliminates the substantial overhead discussed in the previous section, and offers greater expressiveness. Performance is not just a matter of implementation—Protovis does not have enough information, in terms of dependencies, to update the scene efficiently. D3 can even bind data to existing documents, decoupling transformation from generation. For more on transformations, see the enter & exit section of the overview, and part 2 of the bar chart tutorial.

Perhaps the sexiest change is that D3 supports automatic transitions, where attributes or styles are smoothly interpolated over time. Transitions were experimented with Protovis, but without direct control over transformations, the Protovis model became unwieldy for dynamic scenes.

Immediate Evaluation

To improve debugging, D3 applies operators immediately, rather than deferring evaluation to a render call as with Protovis. By shortening the delay between when property functions are defined and when they are evaluated, D3 reduces the likelihood of external state changing (or evaporating), which can break evaluation. Reference capture, although convenient for re-evaluation, was a frequent source of confusion by users. This was exacerbated by a common misunderstanding of JavaScript’s var keyword, which is function- rather than block-scoped.

Immediate evaluation also minimizes internal control flow, moving it up to the user’s code; for example, the attr operator immediately sets attributes on the current selection and then returns. The Protovis model, in contrast, has elaborate hidden control flow that is only revealed when the code crashes—a confusing consequence of embedded declarative languages. Immediate evaluation offers better compatibility with standard JavaScript organizational constructs, such as functions and loops; D3 can even generate recursive structures through the each operator.

Another benefit of eliminating control flow is that layouts are decoupled from property evaluation. Protovis layouts were difficult to implement because they required understanding the internal rendering machinery. In D3, layouts are simply helper classes that create or modify data structures, such as a description of chords, or positions of nodes in a force-directed graph. The user then binds the layout’s data to attributes and elements as needed.

Anything Else?

Despite all the differences above, there remain many similarities between D3 and Protovis, too. To name a few:

D3 is also younger than Protovis, and very much a work-in-progress. We are planning on porting additional features from Protovis to D3, such as hierarchical layouts. If there’s a feature that’s particularly important to you, please let us know!

Copyright © 2011 Mike Bostock
Fork me on GitHub