# Changes in D3 4.0

D3 4.0 is modular. Instead of one library, D3 is now many small libraries that are designed to work together. You can pick and choose which parts to use as you see fit. Each library is maintained in its own repository, allowing decentralized ownership and independent release cycles. The default bundle combines about thirty of these microlibraries.

<script src="https://d3js.org/d3.v4.js"></script>


As before, you can load optional plugins on top of the default bundle, such as ColorBrewer scales:

<script src="https://d3js.org/d3.v4.js"></script>
<script src="https://d3js.org/d3-scale-chromatic.v0.3.js"></script>


You are not required to use the default bundle! If you’re just using d3-selection, use it as a standalone library. Like the default bundle, you can load D3 microlibraries using vanilla script tags or RequireJS (great for HTTP/2!):

<script src="https://d3js.org/d3-selection.v1.js"></script>


You can also cat D3 microlibraries into a custom bundle, or use tools such as Webpack and Rollup to create optimized bundles. Custom bundles are great for applications that use a subset of D3’s features; for example, a React chart library might use D3 for scales and shapes, and React to manipulate the DOM. The D3 microlibraries are written as ES6 modules, and Rollup lets you pick at the symbol level to produce smaller bundles.

Small files are nice, but modularity is also about making D3 more fun. Microlibraries are easier to understand, develop and test. They make it easier for new people to get involved and contribute. They reduce the distinction between a “core module” and a “plugin”, and increase the pace of development in D3 features.

If you don’t care about modularity, you can mostly ignore this change and keep using the default bundle. However, there is one unavoidable consequence of adopting ES6 modules: every symbol in D3 4.0 now shares a flat namespace rather than the nested one of D3 3.x. For example, d3.scale.linear is now d3.scaleLinear, and d3.layout.treemap is now d3.treemap. The adoption of ES6 modules also means that D3 is now written exclusively in strict mode and has better readability. And there have been many other significant improvements to D3’s features! (Nearly all of the code from D3 3.x has been rewritten.) These changes are covered below.

### Other Global Changes

The default UMD bundle is now anonymous. No d3 global is exported if AMD or CommonJS is detected. In a vanilla environment, the D3 microlibraries share the d3 global, even if you load them independently; thus, code you write is the same whether or not you use the default bundle. (See Let’s Make a (D3) Plugin for more.) The generated bundle is no longer stored in the Git repository; Bower has been repointed to d3-bower, and you can find the generated files on npm or attached to the latest release. The non-minified default bundle is no longer mangled, making it more readable and preserving inline comments.

To the consternation of some users, 3.x employed Unicode variable names such as λ, φ, τ and π for a concise representation of mathematical operations. A downside of this approach was that a SyntaxError would occur if you loaded the non-minified D3 using ISO-8859-1 instead of UTF-8. 3.x also used Unicode string literals, such as the SI-prefix µ for 1e-6. 4.0 uses only ASCII variable names and ASCII string literals (see rollup-plugin-ascii), avoiding encoding problems.

## Arrays (d3-array)

The new d3.scan method performs a linear scan of an array, returning the index of the least element according to the specified comparator. This is similar to d3.min and d3.max, except you can use it to find the position of an extreme element, rather than just calculate an extreme value.

var data = [
{name: "Alice", value: 2},
{name: "Bob", value: 3},
{name: "Carol", value: 1},
{name: "Dwayne", value: 5}
];

var i = d3.scan(data, function(a, b) { return a.value - b.value; }); // 2
data[i]; // {name: "Carol", value: 1}


The new d3.ticks and d3.tickStep methods are useful for generating human-readable numeric ticks. These methods are a low-level alternative to continuous.ticks from d3-scale. The new implementation is also more accurate, returning the optimal number of ticks as measured by relative error.

var ticks = d3.ticks(0, 10, 5); // [0, 2, 4, 6, 8, 10]


The d3.range method no longer makes an elaborate attempt to avoid floating-point error when step is not an integer. The returned values are strictly defined as start + i * step, where i is an integer. (Learn more about floating point math.) d3.range returns the empty array for infinite ranges, rather than throwing an error.

The method signature for optional accessors has been changed to be more consistent with array methods such as array.forEach: the accessor is passed the current element (d), the index (i), and the array (data), with this as undefined. This affects d3.min, d3.max, d3.extent, d3.sum, d3.mean, d3.median, d3.quantile, d3.variance and d3.deviation. The d3.quantile method previously did not take an accessor. Some methods with optional arguments now treat those arguments as missing if they are null or undefined, rather than strictly checking arguments.length.

The new d3.histogram API replaces d3.layout.histogram. Rather than exposing bin.x and bin.dx on each returned bin, the histogram exposes bin.x0 and bin.x1, guaranteeing that bin.x0 is exactly equal to bin.x1 on the preceeding bin. The “frequency” and “probability” modes are no longer supported; each bin is simply an array of elements from the input data, so bin.length is equal to D3 3.x’s bin.y in frequency mode. To compute a probability distribution, divide the number of elements in each bin by the total number of elements.

The histogram.range method has been renamed histogram.domain for consistency with scales. The histogram.bins method has been renamed histogram.thresholds, and no longer accepts an upper value: n thresholds will produce n + 1 bins. If you specify a desired number of bins rather than thresholds, d3.histogram now uses d3.ticks to compute nice bin thresholds. In addition to the default Sturges’ formula, D3 now implements the Freedman-Diaconis rule and Scott’s normal reference rule.

## Axes (d3-axis)

To render axes properly in D3 3.x, you needed to style them:

<style>

.axis path,
.axis line {
fill: none;
stroke: #000;
shape-rendering: crispEdges;
}

.axis text {
font: 10px sans-serif;
}

</style>
<script>

d3.select(".axis")
.call(d3.svg.axis()
.scale(x)
.orient("bottom"));

</script>


If you didn’t, you saw this:

D3 4.0 provides default styles and shorter syntax. In place of d3.svg.axis and axis.orient, D3 4.0 now provides four constructors for each orientation: d3.axisTop, d3.axisRight, d3.axisBottom, d3.axisLeft. These constructors accept a scale, so you can reduce all of the above to:

<script>

d3.select(".axis")
.call(d3.axisBottom(x));

</script>


And get this:

As before, you can customize the axis appearance either by applying stylesheets or by modifying the axis elements. The default appearance has been changed slightly to offset the axis by a half-pixel; this fixes a crisp-edges rendering issue on Safari where the axis would be drawn two-pixels thick.

There’s now an axis.tickArguments method, as an alternative to axis.ticks that also allows the axis tick arguments to be inspected. The axis.tickSize method has been changed to only allow a single argument when setting the tick size. The axis.innerTickSize and axis.outerTickSize methods have been renamed axis.tickSizeInner and axis.tickSizeOuter, respectively.

## Brushes (d3-brush)

Replacing d3.svg.brush, there are now three classes of brush for brushing along the x-dimension, the y-dimension, or both: d3.brushX, d3.brushY, d3.brush. Brushes are no longer dependent on scales; instead, each brush defines a selection in screen coordinates. This selection can be inverted if you want to compute the corresponding data domain. And rather than rely on the scales’ ranges to determine the brushable area, there is now a brush.extent method for setting it. If you do not set the brush extent, it defaults to the full extent of the owner SVG element. The brush.clamp method has also been eliminated; brushing is always restricted to the brushable area defined by the brush extent.

Brushes no longer store the active brush selection (i.e., the highlighted region; the brush’s position) internally. The brush’s position is now stored on any elements to which the brush has been applied. The brush’s position is available as event.selection within a brush event or by calling d3.brushSelection on a given element. To move the brush programmatically, use brush.move with a given selection or transition; see the brush snapping example. The brush.event method has been removed.

Brush interaction has been improved. By default, brushes now ignore right-clicks intended for the context menu; you can change this behavior using brush.filter. Brushes also ignore emulated mouse events on iOS. Holding down SHIFT (⇧) while brushing locks the x- or y-position of the brush. Holding down META (⌘) while clicking and dragging starts a new selection, rather than translating the existing selection.

The default appearance of the brush has also been improved and slightly simplified. Previously it was necessary to apply styles to the brush to give it a reasonable appearance, such as:

.brush .extent {
stroke: #fff;
fill-opacity: .125;
shape-rendering: crispEdges;
}


These styles are now applied by default as attributes; if you want to customize the brush appearance, you can still apply external styles or modify the brush elements. (D3 4.0 features a similar improvement to axes.) A new brush.handleSize method lets you override the brush handle size; it defaults to six pixels.

The brush now consumes handled events, making it easier to combine with other interactive behaviors such as dragging and zooming. The brushstart and brushend events have been renamed to start and end, respectively. The brush event no longer reports a event.mode to distinguish between resizing and dragging the brush.

## Chords (d3-chord)

Pursuant to the great namespace flattening:

## Collections (d3-collection)

The d3.set constructor now accepts an existing set for making a copy. If you pass an array to d3.set, you can also pass a value accessor. This accessor takes the standard arguments: the current element (d), the index (i), and the array (data), with this undefined. For example:

var yields = [
{yield: 22.13333, variety: "Manchuria",        year: 1932, site: "Grand Rapids"},
{yield: 26.76667, variety: "Peatland",         year: 1932, site: "Grand Rapids"},
{yield: 28.10000, variety: "No. 462",          year: 1931, site: "Duluth"},
{yield: 38.50000, variety: "Svansota",         year: 1932, site: "Waseca"},
{yield: 40.46667, variety: "Svansota",         year: 1931, site: "Crookston"},
{yield: 36.03333, variety: "Peatland",         year: 1932, site: "Waseca"},
{yield: 34.46667, variety: "Wisconsin No. 38", year: 1931, site: "Grand Rapids"}
];

var sites = d3.set(yields, function(d) { return d.site; }); // Grand Rapids, Duluth, Waseca, Crookston


The d3.map constructor also follows the standard array accessor argument pattern.

The map.forEach and set.forEach methods have been renamed to map.each and set.each respectively. The order of arguments for map.each has also been changed to value, key and map, while the order of arguments for set.each is now value, value and set. This is closer to ES6 map.forEach and set.forEach. Also like ES6 Map and Set, map.set and set.add now return the current collection (rather than the added value) to facilitate method chaining. New map.clear and set.clear methods can be used to empty collections.

The nest.map method now always returns a d3.map instance. For a plain object, use nest.object instead. When used in conjunction with nest.rollup, nest.entries now returns {key, value} objects for the leaf entries, instead of {key, values}. This makes nest.rollup easier to use in conjunction with hierarchies, as in this Nest Treemap example.

## Colors (d3-color)

All colors now have opacity exposed as color.opacity, which is a number in [0, 1]. You can pass an optional opacity argument to the color space constructors d3.rgb, d3.hsl, d3.lab, d3.hcl or d3.cubehelix.

You can now parse rgba(…) and hsla(…) CSS color specifiers or the string “transparent” using d3.color. The “transparent” color is defined as an RGB color with zero opacity and undefined red, green and blue channels; this differs slightly from CSS which defines it as transparent black, but is useful for simplifying color interpolation logic where either the starting or ending color has undefined channels. The color.toString method now likewise returns an rgb(…) or rgba(…) string with integer channel values, not the hexadecimal RGB format, consistent with CSS computed values. This improves performance by short-circuiting transitions when the element’s starting style matches its ending style.

The new d3.color method is the primary method for parsing colors: it returns a d3.color instance in the appropriate color space, or null if the CSS color specifier is invalid. For example:

var red = d3.color("hsl(0, 80%, 50%)"); // {h: 0, l: 0.5, s: 0.8, opacity: 1}


The parsing implementation is now more robust. For example, you can no longer mix integers and percentages in rgb(…), and it correctly handles whitespace, decimal points, number signs, and other edge cases. The color space constructors d3.rgb, d3.hsl, d3.lab, d3.hcl and d3.cubehelix now always return a copy of the input color, converted to the corresponding color space. While color.rgb remains, rgb.hsl has been removed; use d3.hsl to convert a color to the RGB color space.

The RGB color space no longer greedily quantizes and clamps channel values when creating colors, improving accuracy in color space conversion. Quantization and clamping now occurs in color.toString when formatting a color for display. You can use the new color.displayable to test whether a color is out-of-gamut.

The rgb.brighter method no longer special-cases black. This is a multiplicative operator, defining a new color r′, g′, b′ where r′ = r × pow(0.7, k), g′ = g × pow(0.7, k) and b′ = b × pow(0.7, k); a brighter black is still black.

There’s a new d3.cubehelix color space, generalizing Dave Green’s color scheme! (See also d3.interpolateCubehelixDefault from d3-scale.) You can continue to define your own custom color spaces, too; see d3-hsv for an example.

## Dispatches (d3-dispatch)

Rather than decorating the dispatch object with each event type, the dispatch object now exposes generic dispatch.call and dispatch.apply methods which take the type string as the first argument. For example, in D3 3.x, you might say:

dispatcher.foo.call(that, "Hello, Foo!");


To dispatch a foo event in D3 4.0, you’d say:

dispatcher.call("foo", that, "Hello, Foo!");


The dispatch.on method now accepts multiple typenames, allowing you to add or remove listeners for multiple events simultaneously. For example, to send both foo and bar events to the same listener:

dispatcher.on("foo bar", function(message) {
console.log(message);
});


This matches the new behavior of selection.on in d3-selection. The dispatch.on method now validates that the specifier listener is a function, rather than throwing an error in the future.

The new implementation d3.dispatch is faster, using fewer closures to improve performance. There’s also a new dispatch.copy method for making a copy of a dispatcher; this is used by d3-transition to improve the performance of transitions in the common case where all elements in a transition have the same transition event listeners.

## Dragging (d3-drag)

The drag behavior d3.behavior.drag has been renamed to d3.drag. The drag.origin method has been replaced by drag.subject, which allows you to define the thing being dragged at the start of a drag gesture. This is particularly useful with Canvas, where draggable objects typically share a Canvas element (as opposed to SVG, where draggable objects typically have distinct DOM elements); see the circle dragging example.

A new drag.container method lets you override the parent element that defines the drag gesture coordinate system. This defaults to the parent node of the element to which the drag behavior was applied. For dragging on Canvas elements, you probably want to use the Canvas element as the container.

Drag events now expose an event.on method for registering temporary listeners for duration of the current drag gesture; these listeners can capture state for the current gesture, such as the thing being dragged. A new event.active property lets you detect whether multiple (multitouch) drag gestures are active concurrently. The dragstart and dragend events have been renamed to start and end. By default, drag behaviors now ignore right-clicks intended for the context menu; use drag.filter to control which events are ignored. The drag behavior also ignores emulated mouse events on iOS. The drag behavior now consumes handled events, making it easier to combine with other interactive behaviors such as zooming.

The new d3.dragEnable and d3.dragDisable methods provide a low-level API for implementing drag gestures across browsers and devices. These methods are also used by other D3 components, such as the brush.

## Delimiter-Separated Values (d3-dsv)

Pursuant to the great namespace flattening, various CSV and TSV methods have new names:

The d3.csv and d3.tsv methods for loading files of the corresponding formats have not been renamed, however! Those are defined in d3-request.There’s no longer a d3.dsv method, which served the triple purpose of defining a DSV formatter, a DSV parser and a DSV requestor; instead, there’s just d3.dsvFormat which you can use to define a DSV formatter and parser. You can use request.response to make a request and then parse the response body, or just use d3.text.

The dsv.parse method now exposes the column names and their input order as data.columns. For example:

d3.csv("cars.csv", function(error, data) {
if (error) throw error;
console.log(data.columns); // ["Year", "Make", "Model", "Length"]
});


You can likewise pass an optional array of column names to dsv.format to format only a subset of columns, or to specify the column order explicitly:

var string = d3.csvFormat(data, ["Year", "Model", "Length"]);


The parser is a bit faster and the formatter is a bit more robust: inputs are coerced to strings before formatting, fixing an obscure crash, and deprecated support for falling back to dsv.formatRows when the input data is an array of arrays has been removed.

## Easings (d3-ease)

D3 3.x used strings, such as “cubic-in-out”, to identify easing methods; these strings could be passed to d3.ease or transition.ease. D3 4.0 uses symbols instead, such as d3.easeCubicInOut. Symbols are simpler and cleaner. They work well with Rollup to produce smaller custom bundles. You can still define your own custom easing function, too, if desired. Here’s the full list of equivalents:

¹ The -in, -out and -in-out variants of linear easing are identical, so there’s just d3.easeLinear.
² Elastic and bounce easing were inadvertently reversed in 3.x, so 4.0 eliminates -out-in easing!

For convenience, there are also default aliases for each easing method. For example, d3.easeCubic is an alias for d3.easeCubicInOut. Most default to -in-out; the exceptions are d3.easeBounce and d3.easeElastic, which default to -out.

Rather than pass optional arguments to d3.ease or transition.ease, parameterizable easing functions now have named parameters: poly.exponent, elastic.amplitude, elastic.period and back.overshoot. For example, in D3 3.x you might say:

var e = d3.ease("elastic-out-in", 1.2);


The equivalent in D3 4.0 is:

var e = d3.easeElastic.amplitude(1.2);


Many of the easing functions have been optimized for performance and accuracy. Several bugs have been fixed, as well, such as the interpretation of the overshoot parameter for back easing, and the period parameter for elastic easing. Also, d3-transition now explicitly guarantees that the last tick of the transition happens at exactly t = 1, avoiding floating point errors in some easing functions.

There’s now a nice visual reference and an animated reference to the new easing functions, too!

## Forces (d3-force)

The force layout d3.layout.force has been renamed to d3.forceSimulation. The force simulation now uses velocity Verlet integration rather than position Verlet, tracking the nodes’ positions (node.x, node.y) and velocities (node.vx, node.vy) rather than their previous positions (node.px, node.py).

Rather than hard-coding a set of built-in forces, the force simulation is now extensible: you specify which forces you want! The approach affords greater flexibility through composition. The new forces are more flexible, too: force parameters can typically be configured per-node or per-link. There are separate positioning forces for x and y that replace force.gravity; x.x and y.y replace force.size. The new link force replaces force.linkStrength and employs better default heuristics to improve stability. The new many-body force replaces force.charge and supports a new minimum-distance parameter and performance improvements thanks to 4.0’s new quadtrees. There are also brand-new forces for centering nodes and collision resolution.

The new forces and simulation have been carefully crafted to avoid nondeterminism. Rather than initializing nodes randomly, if the nodes do not have preset positions, they are placed in a phyllotaxis pattern:

Random jitter is still needed to resolve link, collision and many-body forces if there are coincident nodes, but at least in the common case, the force simulation (and the resulting force-directed graph layout) is now consistent across browsers and reloads. D3 no longer plays dice!

The force simulation has several new methods for greater control over heating, such as simulation.alphaMin and simulation.alphaDecay, and the internal timer. Calling simulation.alpha now has no effect on the internal timer, which is controlled independently via simulation.stop and simulation.restart. The force layout’s internal timer now starts automatically on creation, removing force.start. As in 3.x, you can advance the simulation manually using simulation.tick. The force.friction parameter is replaced by simulation.velocityDecay. A new simulation.alphaTarget method allows you to set the desired alpha (temperature) of the simulation, such that the simulation can be smoothly reheated during interaction, and then smoothly cooled again. This improves the stability of the graph during interaction.

The force layout no longer depends on the drag behavior, though you can certainly create draggable force-directed graphs! Set node.fx and node.fy to fix a node’s position. As an alternative to a Voronoi SVG overlay, you can now use simulation.find to find the closest node to a pointer.

## Number Formats (d3-format)

If a precision is not specified, the formatting behavior has changed: there is now a default precision of 6 for all directives except none, which defaults to 12. In 3.x, if you did not specify a precision, the number was formatted using its shortest unique representation (per number.toString); this could lead to unexpected digits due to floating point math. The new default precision in 4.0 produces more consistent results:

var f = d3.format("e");
f(42);        // "4.200000e+1"
f(0.1 + 0.2); // "3.000000e-1"


To trim insignificant trailing zeroes, use the none directive, which is similar g. For example:

var f = d3.format(".3");
f(0.12345);   // "0.123"
f(0.10000);   // "0.1"
f(0.1 + 0.2); // "0.3"


Under the hood, number formatting has improved accuracy with very large and very small numbers by using number.toExponential rather than Math.log to extract the mantissa and exponent. Negative zero (-0, an IEEE 754 construct) and very small numbers that round to zero are now formatted as unsigned zero. The inherently unsafe d3.round method has been removed, along with d3.requote.

The d3.formatPrefix method has been changed. Rather than returning an SI-prefix string, it returns an SI-prefix format function for a given specifier and reference value. For example, to format thousands:

var f = d3.formatPrefix(",.0", 1e3);
f(1e3); // "1k"
f(1e4); // "10k"
f(1e5); // "100k"
f(1e6); // "1,000k"


Unlike the s format directive, d3.formatPrefix always employs the same SI-prefix, producing consistent results:

var f = d3.format(".0s");
f(1e3); // "1k"
f(1e4); // "10k"
f(1e5); // "100k"
f(1e6); // "1M"


The new ( sign option uses parentheses for negative values. This is particularly useful in conjunction with $. For example: d3.format("+.0f")(-42); // "-42" d3.format("(.0f")(-42); // "(42)" d3.format("+$.0f")(-42); // "-$42" d3.format("($.0f")(-42); // "(\$42)"


The new = align option places any sign and symbol to the left of any padding:

d3.format(">6d")(-42);  // "   -42"
d3.format("=6d")(-42);  // "-   42"
d3.format(">(6d")(-42); // "  (42)"
d3.format("=(6d")(-42); // "(  42)"


The b, o, d and x directives now round to the nearest integer, rather than returning the empty string for non-integers:

d3.format("b")(41.9); // "101010"
d3.format("o")(41.9); // "52"
d3.format("d")(41.9); // "42"
d3.format("x")(41.9); // "2a"


The c directive is now for character data (i.e., literal strings), not for character codes. The is useful if you just want to apply padding and alignment and don’t care about formatting numbers. For example, the infamous left-pad (as well as center- and right-pad!) can be conveniently implemented as:

d3.format(">10c")("foo"); // "       foo"
d3.format("^10c")("foo"); // "   foo    "
d3.format("<10c")("foo"); // "foo       "


There are several new methods for computing suggested decimal precisions; these are used by d3-scale for tick formatting, and are helpful for implementing custom number formats: d3.precisionFixed, d3.precisionPrefix and d3.precisionRound. There’s also a new d3.formatSpecifier method for parsing, validating and debugging format specifiers; it’s also good for deriving related format specifiers, such as when you want to substitute the precision automatically.

You can now set the default locale using d3.formatDefaultLocale! The locales are published as JSON to npm.

## Geographies (d3-geo)

Pursuant to the great namespace flattening, various methods have new names:

Also renamed for consistency:

Projections now have more appropriate defaults. For example, d3.geoOrthographic has a 90° clip angle by default, showing only the front hemisphere, and d3.geoGnomonic has a default 60° clip angle. The default projection for d3.geoPath is now null rather than d3.geoAlbersUsa; a null projection is used with pre-projected geometry and is typically faster to render.

“Fallback projections”—when you pass a function rather than a projection to path.projection—are no longer supported. For geographic projections, use d3.geoProjection or d3.geoProjectionMutator to define a custom projection. For arbitrary geometry transformations, implement the stream interface; see also d3.geoTransform. The “raw” projections (e.g., d3.geo.equirectangular.raw) are no longer exported.

## Hierarchies (d3-hierarchy)

Pursuant to the great namespace flattening:

As an alternative to using JSON to represent hierarchical data (such as the “flare.json format” used by many D3 examples), the new d3.stratify operator simplifies the conversion of tabular data to hierarchical data! This is convenient if you already have data in a tabular format, such as the result of a SQL query or a CSV file:

name,parent
Eve,
Cain,Eve
Seth,Eve
Enos,Seth
Noam,Seth
Abel,Eve
Awan,Eve
Enoch,Awan
Azura,Eve


To convert this to a root node:

var root = d3.stratify()
.id(function(d) { return d.name; })
.parentId(function(d) { return d.parent; })
(nodes);


The resulting root can be passed to d3.tree to produce a tree diagram like this:

Root nodes can also be created from JSON data using d3.hierarchy. The hierarchy layouts now take these root nodes as input rather than operating directly on JSON data, which helps to provide a cleaner separation between the input data and the computed layout. (For example, use node.copy to isolate layout changes.) It also simplifies the API: rather than each hierarchy layout needing to implement value and sorting accessors, there are now generic node.sum and node.sort methods that work with any hierarchy layout.

The new d3.hierarchy API also provides a richer set of methods for manipulating hierarchical data. For example, to generate an array of all nodes in topological order, use node.descendants; for just leaf nodes, use node.leaves. To highlight the ancestors of a given node on mouseover, use node.ancestors. To generate an array of {source, target} links for a given hierarchy, use node.links; this replaces treemap.links and similar methods on the other layouts. The new node.path method replaces d3.layout.bundle; see also d3.curveBundle for hierarchical edge bundling.

The hierarchy layouts have been rewritten using new, non-recursive traversal methods (node.each, node.eachAfter and node.eachBefore), improving performance on large datasets. The d3.tree layout no longer uses a node._ field to store temporary state during layout.

Treemap tiling is now extensible via treemap.tile! The default squarified tiling algorithm, d3.treemapSquarify, has been completely rewritten, improving performance and fixing bugs in padding and rounding. The treemap.sticky method has been replaced with the d3.treemapResquarify, which is identical to d3.treemapSquarify except it performs stable neighbor-preserving updates. The treemap.ratio method has been replaced with squarify.ratio. And there’s a new d3.treemapBinary for binary treemaps!

Treemap padding has also been improved. The treemap now distinguishes between outer padding that separates a parent from its children, and inner padding that separates adjacent siblings. You can set the top-, right-, bottom- and left-outer padding separately. There are new examples for the traditional nested treemap and for Lü and Fogarty’s cascaded treemap. And there’s a new example demonstrating d3.nest with d3.treemap.

The space-filling layouts d3.treemap and d3.partition now output x0, x1, y0, y1 on each node instead of x0, dx, y0, dy. This improves accuracy by ensuring that the edges of adjacent cells are exactly equal, rather than sometimes being slightly off due to floating point math. The partition layout now supports rounding and padding.

The circle-packing layout, d3.pack, has been completely rewritten to better implement Wang et al.’s algorithm, fixing major bugs and improving results! Welzl’s algorithm is now used to compute the exact smallest enclosing circle for each parent, rather than the approximate answer used by Wang et al. The 3.x output is shown on the left; 4.0 is shown on the right:

A non-hierarchical implementation is also available as d3.packSiblings, and the smallest enclosing circle implementation is available as d3.packEnclose. Pack padding now applies between a parent and its children, as well as between adjacent siblings. In addition, you can now specify padding as a function that is computed dynamically for each parent.

## Internals

The d3.rebind method has been removed. (See the 3.x source.) If you want to wrap a getter-setter method, the recommend pattern is to implement a wrapper method and check the return value. For example, given a component that uses an internal dispatch, component.on can rebind dispatch.on as follows:

component.on = function() {
var value = dispatch.on.apply(dispatch, arguments);
return value === dispatch ? component : value;
};


The d3.functor method has been removed. (See the 3.x source.) If you want to promote a constant value to a function, the recommended pattern is to implement a closure that returns the constant value. If desired, you can use a helper method as follows:

function constant(x) {
return function() {
return x;
};
}


Given a value x, to promote x to a function if it is not already:

var fx = typeof x === "function" ? x : constant(x);


## Interpolators (d3-interpolate)

The d3.interpolate method no longer delegates to d3.interpolators, which has been removed; its behavior is now defined by the library. It is now slightly faster in the common case that b is a number. It only uses d3.interpolateRgb if b is a valid CSS color specifier (and not approximately one). And if the end value b is null, undefined, true or false, d3.interpolate now returns a constant function which always returns b.

The behavior of d3.interpolateObject and d3.interpolateArray has changed slightly with respect to properties or elements in the start value a that do not exist in the end value b: these properties and elements are now ignored, such that the ending value of the interpolator at t = 1 is now precisely equal to b. So, in 3.x:

d3.interpolateObject({foo: 2, bar: 1}, {foo: 3})(0.5); // {bar: 1, foo: 2.5} in 3.x


Whereas in 4.0, a.bar is ignored:

d3.interpolateObject({foo: 2, bar: 1}, {foo: 3})(0.5); // {foo: 2.5} in 4.0


If a or b are undefined or not an object, they are now implicitly converted to the empty object or empty array as appropriate, rather than throwing a TypeError.

The d3.interpolateTransform interpolator has been renamed to d3.interpolateTransformSvg, and there is a new d3.interpolateTransformCss to interpolate CSS transforms! This allows d3-transition to automatically interpolate both the SVG transform attribute and the CSS transform style property. (Note, however, that only 2D CSS transforms are supported.) The d3.transform method has been removed.

Color space interpolators now interpolate opacity (see d3-color) and return rgb(…) or rgba(…) CSS color specifier strings rather than using the RGB hexadecimal format. This is necessary to support opacity interpolation, but is also beneficial because it matches CSS computed values. When a channel in the start color a is undefined, color interpolators now use the corresponding channel value from the end color b, or vice versa. This logic previously applied to some channels (such as saturation in HSL), but now applies to all channels in all color spaces, and is especially useful when interpolating to or from transparent.

There are now “long” versions of cylindrical color space interpolators: d3.interpolateHslLong, d3.interpolateHclLong and d3.interpolateCubehelixLong. These interpolators use linear interpolation of hue, rather than using the shortest path around the 360° hue circle. See d3.interpolateRainbow for an example. The Cubehelix color space is now supported by d3-color, and so there are now d3.interpolateCubehelix and d3.interpolateCubehelixLong interpolators.

Gamma-corrected color interpolation is now supported for both RGB and Cubehelix color spaces as interpolate.gamma. For example, to interpolate from purple to orange with a gamma of 2.2 in RGB space:

var interpolate = d3.interpolateRgb.gamma(2.2)("purple", "orange");


There are new interpolators for uniform non-rational B-splines! These are useful for smoothly interpolating between an arbitrary sequence of values from t = 0 to t = 1, such as to generate a smooth color gradient from a discrete set of colors. The d3.interpolateBasis and d3.interpolateBasisClosed interpolators generate one-dimensional B-splines, while d3.interpolateRgbBasis and d3.interpolateRgbBasisClosed generate three-dimensional B-splines through RGB color space. These are used by d3-scale-chromatic to generate continuous color scales from ColorBrewer’s discrete color schemes, such as PiYG.

There’s also now a d3.quantize method for generating uniformly-spaced discrete samples from a continuous interpolator. This is useful for taking one of the built-in color scales (such as d3.interpolateViridis) and quantizing it for use with d3.scaleQuantize, d3.scaleQuantile or d3.scaleThreshold.

## Paths (d3-path)

The d3.path serializer implements the CanvasPathMethods API, allowing you to write code that can render to either Canvas or SVG. For example, given some code that draws to a canvas:

function drawCircle(context, radius) {
context.arc(0, 0, radius, 0, 2 * Math.PI);
}


You can render to SVG as follows:

var context = d3.path();
drawCircle(context, 40);
pathElement.setAttribute("d", context.toString());


The path serializer enables d3-shape to support both Canvas and SVG; see line.context and area.context, for example.

## Polygons (d3-polygon)

There’s no longer a d3.geom.polygon constructor; instead you just pass an array of vertices to the polygon methods. So instead of polygon.area and polygon.centroid, there’s d3.polygonArea and d3.polygonCentroid. There are also new d3.polygonContains and d3.polygonLength methods. There’s no longer an equivalent to polygon.clip, but if Sutherland–Hodgman clipping is needed, please file a feature request.

The d3.geom.hull operator has been simplified: instead of an operator with hull.x and hull.y accessors, there’s just the d3.polygonHull method which takes an array of points and returns the convex hull.

var quadtree = d3.geom.quadtree()
.extent([[0, 0], [width, height]])
(data);


Can be rewritten in 4.0 as:

var quadtree = d3.quadtree()
.extent([[0, 0], [width, height]])


The new quadtree implementation is vastly improved! It is no longer recursive, avoiding stack overflows when there are large numbers of coincident points. The internal storage is now more efficient, and the implementation is also faster; constructing a quadtree of 1M normally-distributed points takes about one second in 4.0, as compared to three seconds in 3.x.

The change in internal node structure affects quadtree.visit: use node.length to distinguish leaf nodes from internal nodes. For example, to iterate over all data in a quadtree:

quadtree.visit(function(node) {
if (!node.length) {
do {
console.log(node.data);
} while (node = node.next)
}
});


There’s a new quadtree.visitAfter method for visiting nodes in post-order traversal. This feature is used in d3-force to implement the Barnes–Hut approximation.

You can now remove data from a quadtree using quadtree.remove and quadtree.removeAll. When adding data to a quadtree, the quadtree will now expand its extent by repeated doubling if the new point is outside the existing extent of the quadtree. There are also quadtree.extent and quadtree.cover methods for explicitly expanding the extent of the quadtree after creation.

## Queues (d3-queue)

Formerly known as Queue.js and queue-async, d3.queue is now included in the default bundle, making it easy to load data files in parallel. It has been rewritten with fewer closures to improve performance, and there are now stricter checks in place to guarantee well-defined behavior. You can now use instanceof d3.queue and inspect the queue’s internal private state.

## Random Numbers (d3-random)

Pursuant to the great namespace flattening, the random number generators have new names:

There are also new random number generators for exponential and uniform distributions. The normal and log-normal random generators have been optimized.

## Requests (d3-request)

The d3.xhr method has been renamed to d3.request. Basic authentication is now supported using request.user and request.password. You can now configure a timeout using request.timeout.

If an error occurs, the corresponding ProgressEvent of type “error” is now passed to the error listener, rather than the XMLHttpRequest. Likewise, the ProgressEvent is passed to progress event listeners, rather than using d3.event. If d3.xml encounters an error parsing XML, this error is now reported to error listeners rather than returning a null response.

The d3.request, d3.text and d3.xml methods no longer take an optional mime type as the second argument; use request.mimeType instead. For example:

d3.xml("file.svg").mimeType("image/svg+xml").get(function(error, svg) {
…
});


With the exception of d3.html and d3.xml, Node is now supported via node-XMLHttpRequest.

## Scales (d3-scale)

Pursuant to the great namespace flattening:

Scales now generate ticks in the same order as the domain: if you have a descending domain, you now get descending ticks. This change affects the order of tick elements generated by axes. For example:

d3.scaleLinear().domain([10, 0]).ticks(5); // [10, 8, 6, 4, 2, 0]


Log tick formatting now assumes a default count of ten, not Infinity, if not specified. Log scales with domains that span many powers (such as from 1e+3 to 1e+29) now return only one tick per power rather than returning base ticks per power. Non-linear quantitative scales are slightly more accurate.

You can now control whether an ordinal scale’s domain is implicitly extended when the scale is passed a value that is not already in its domain. By default, ordinal.unknown is d3.scaleImplicit, causing unknown values to be added to the domain:

var x = d3.scaleOrdinal()
.domain([0, 1])
.range(["red", "green", "blue"]);

x.domain(); // [0, 1]
x(2); // "blue"
x.domain(); // [0, 1, 2]


By setting ordinal.unknown, you instead define the output value for unknown inputs. This is particularly useful for choropleth maps where you want to assign a color to missing data.

var x = d3.scaleOrdinal()
.domain([0, 1])
.range(["red", "green", "blue"])
.unknown(undefined);

x.domain(); // [0, 1]
x(2); // undefined
x.domain(); // [0, 1]


The ordinal.rangeBands and ordinal.rangeRoundBands methods have been replaced with a new subclass of ordinal scale: band scales. The following code in 3.x:

var x = d3.scale.ordinal()
.domain(["a", "b", "c"])
.rangeBands([0, width]);


Is equivalent to this in 4.0:

var x = d3.scaleBand()
.domain(["a", "b", "c"])
.range([0, width]);


The new band.padding, band.paddingInner and band.paddingOuter methods replace the optional arguments to ordinal.rangeBands. The new band.bandwidth and band.step methods replace ordinal.rangeBand. There’s also a new band.align method which you can use to control how the extra space outside the bands is distributed, say to shift columns closer to the y-axis.

Similarly, the ordinal.rangePoints and ordinal.rangeRoundPoints methods have been replaced with a new subclass of ordinal scale: point scales. The following code in 3.x:

var x = d3.scale.ordinal()
.domain(["a", "b", "c"])
.rangePoints([0, width]);


Is equivalent to this in 4.0:

var x = d3.scalePoint()
.domain(["a", "b", "c"])
.range([0, width]);


The new point.padding method replaces the optional padding argument to ordinal.rangePoints. Like ordinal.rangeBand with ordinal.rangePoints, the point.bandwidth method always returns zero; a new point.step method returns the interval between adjacent points.

The ordinal scale constructor now takes an optional range for a shorter alternative to ordinal.range. This is especially useful now that the categorical color scales have been changed to simple arrays of colors rather than specialized ordinal scale constructors:

The following code in 3.x:

var color = d3.scale.category10();


Is equivalent to this in 4.0:

var color = d3.scaleOrdinal(d3.schemeCategory10);


Sequential scales, are a new class of scales with a fixed output interpolator instead of a range. Typically these scales are used to implement continuous sequential or diverging color schemes. Inspired by Matplotlib’s new perceptually-motived colormaps, 4.0 now features viridis, inferno, magma, plasma interpolators for use with sequential scales. Using d3.quantize, these interpolators can also be applied to quantile, quantize and threshold scales.

4.0 also ships new Cubehelix schemes, including Dave Green’s default and a cyclical rainbow inspired by Matteo Niccoli:

For even more sequential and categorical color schemes, see d3-scale-chromatic.

For an introduction to scales, see Introducing d3-scale.

## Selections (d3-selection)

Selections no longer subclass Array using prototype chain injection; they are now plain objects, improving performance. The internal fields (selection._groups, selection._parents) are private; please use the documented public API to manipulate selections. The new selection.nodes method generates an array of all nodes in a selection.

Selections are now immutable: the elements and parents in a selection never change. (The elements’ attributes and content will of course still be modified!) The selection.sort and selection.data methods now return new selections rather than modifying the selection in-place. In addition, selection.append no longer merges entering nodes into the update selection; use selection.merge to combine enter and update after a data join. For example, the following general update pattern in 3.x:

var circle = svg.selectAll("circle").data(data) // UPDATE
.style("fill", "blue");

circle.exit().remove(); // EXIT

circle.enter().append("circle") // ENTER; modifies UPDATE! 🌶
.style("fill", "green");

circle // ENTER + UPDATE
.style("stroke", "black");


Would be rewritten in 4.0 as:

var circle = svg.selectAll("circle").data(data) // UPDATE
.style("fill", "blue");

circle.exit().remove(); // EXIT

circle.enter().append("circle") // ENTER
.style("fill", "green")
.merge(circle) // ENTER + UPDATE
.style("stroke", "black");


This change is discussed further in What Makes Software Good.

In 3.x, the selection.enter and selection.exit methods were undefined until you called selection.data, resulting in a TypeError if you attempted to access them. In 4.0, now they simply return the empty selection if the selection has not been joined to data.

In 3.x, selection.append would always append the new element as the last child of its parent. A little-known trick was to use selection.insert without specifying a before selector when entering nodes, causing the entering nodes to be inserted before the following element in the update selection. In 4.0, this is now the default behavior of selection.append; if you do not specify a before selector to selection.insert, the inserted element is appended as the last child. This change makes the general update pattern preserve the relative order of elements and data. For example, given the following DOM:

<div>a</div>
<div>b</div>
<div>f</div>


And the following code:

var div = d3.select("body").selectAll("div")
.data(["a", "b", "c", "d", "e", "f"], function(d) { return d || this.textContent; });

div.enter().append("div")
.text(function(d) { return d; });


The resulting DOM will be:

<div>a</div>
<div>b</div>
<div>c</div>
<div>d</div>
<div>e</div>
<div>f</div>


Thus, the entering c, d and e are inserted before f, since f is the following element in the update selection. Although this behavior is sufficient to preserve order if the new data’s order is stable, if the data changes order, you must still use selection.order to reorder elements.

There is now only one class of selection. 3.x implemented enter selections using a special class with different behavior for enter.append and enter.select; a consequence of this design was that enter selections in 3.x lacked certain methods. In 4.0, enter selections are simply normal selections; they have the same methods and the same behavior. Placeholder enter nodes now implement node.appendChild, node.insertBefore, node.querySelector, and node.querySelectorAll.

The selection.data method has been changed slightly with respect to duplicate keys. In 3.x, if multiple data had the same key, the duplicate data would be ignored and not included in enter, update or exit; in 4.0 the duplicate data is always put in the enter selection. In both 3.x and 4.0, if multiple elements have the same key, the duplicate elements are put in the exit selection. Thus, 4.0’s behavior is now symmetric for enter and exit, and the general update pattern will now produce a DOM that matches the data even if there are duplicate keys.

Selections have several new methods! Use selection.raise to move the selected elements to the front of their siblings, so that they are drawn on top; use selection.lower to move them to the back. Use selection.dispatch to dispatch a custom event to event listeners.

When called in getter mode, selection.data now returns the data for all elements in the selection, rather than just the data for the first group of elements. The selection.call method no longer sets the this context when invoking the specified function; the selection is passed as the first argument to the function, so use that. The selection.on method now accepts multiple whitespace-separated typenames, so you can add or remove multiple listeners simultaneously. For example:

selection.on("mousedown touchstart", function() {
console.log(d3.event.type);
});


The arguments passed to callback functions has changed slightly in 4.0 to be more consistent. The standard arguments are the element’s datum (d), the element’s index (i), and the element’s group (nodes), with this as the element. The slight exception to this convention is selection.data, which is evaluated for each group rather than each element; it is passed the group’s parent datum (d), the group index (i), and the selection’s parents (parents), with this as the group’s parent.

The new d3.local provides a mechanism for defining local variables: state that is bound to DOM elements, and available to any descendant element. This can be a convenient alternative to using selection.each or storing local state in data.

The d3.ns.prefix namespace prefix map has been renamed to d3.namespaces, and the d3.ns.qualify method has been renamed to d3.namespace. Several new low-level methods are now available, as well. d3.matcher is used internally by selection.filter; d3.selector is used by selection.select; d3.selectorAll is used by selection.selectAll; d3.creator is used by selection.append and selection.insert. The new d3.window returns the owner window for a given element, window or document. The new d3.customEvent temporarily sets d3.event while invoking a function, allowing you to implement controls which dispatch custom events; this is used by d3-drag, d3-zoom and d3-brush.

For the sake of parsimony, the multi-value methods—where you pass an object to set multiple attributes, styles or properties simultaneously—have been extracted to d3-selection-multi and are no longer part of the default bundle. The multi-value map methods have also been renamed to plural form to reduce overload: selection.attrs, selection.styles and selection.properties.

## Shapes (d3-shape)

Pursuant to the great namespace flattening:

Shapes are no longer limited to SVG; they can now render to Canvas! Shape generators now support an optional context: given a CanvasRenderingContext2D, you can render a shape as a canvas path to be filled or stroked. For example, a canvas pie chart might use an arc generator:

var arc = d3.arc()
.context(context);


To render an arc for a given datum d:

context.beginPath();
arc(d);
context.fill();


See line.context, area.context and arc.context for more. Under the hood, shapes use d3-path to serialize canvas path methods to SVG path data when the context is null; thus, shapes are optimized for rendering to canvas. You can also now derive lines from areas. The line shares most of the same accessors, such as line.defined and line.curve, with the area from which it is derived. For example, to render the topline of an area, use area.lineY1; for the baseline, use area.lineY0.

4.0 introduces a new curve API for specifying how line and area shapes interpolate between data points. The line.interpolate and area.interpolate methods have been replaced with line.curve and area.curve. Curves are implemented using the curve interface rather than as a function that returns an SVG path data string; this allows curves to render to either SVG or Canvas. In addition, line.curve and area.curve now take a function which instantiates a curve for a given context, rather than a string. The full list of equivalents:

But that’s not all! 4.0 now provides parameterized Catmull–Rom splines as proposed by Yuksel et al.. These are available as d3.curveCatmullRom, d3.curveCatmullRomClosed and d3.curveCatmullRomOpen.

Each curve type can define its own named parameters, replacing line.tension and area.tension. For example, Catmull–Rom splines are parameterized using catmullRom.alpha and defaults to 0.5, which corresponds to a centripetal spline that avoids self-intersections and overshoot. For a uniform Catmull–Rom spline instead:

var line = d3.line()
.curve(d3.curveCatmullRom.alpha(0));


4.0 fixes the interpretation of the cardinal spline tension parameter, which is now specified as cardinal.tension and defaults to zero for a uniform Catmull–Rom spline; a tension of one produces a linear curve. The first and last segments of basis and cardinal curves have also been fixed! The undocumented interpolate.reverse field has been removed. Curves can define different behavior for toplines and baselines by counting the sequence of curve.lineStart within curve.areaStart. See the d3.curveStep implementation for an example.

4.0 fixes numerous bugs in the monotone curve implementation, and introduces d3.curveMonotoneY; this is like d3.curveMonotoneX, except it requires that the input points are monotone in y rather than x, such as for a vertically-oriented line chart. The new d3.curveNatural produces a natural cubic spline. The default β for d3.curveBundle is now 0.85, rather than 0.7, matching the values used by Holten. 4.0 also has a more robust implementation of arc padding; see arc.padAngle and arc.padRadius.

4.0 introduces a new symbol type API. Symbol types are passed to symbol.type in place of strings. The equivalents are:

The full set of symbol types is now:

Lastly, 4.0 overhauls the stack layout API, replacing d3.layout.stack with d3.stack. The stack generator no longer needs an x-accessor. In addition, the API has been simplified: the stack generator now accepts tabular input, such as this array of objects:

var data = [
{month: new Date(2015, 0, 1), apples: 3840, bananas: 1920, cherries: 960, dates: 400},
{month: new Date(2015, 1, 1), apples: 1600, bananas: 1440, cherries: 960, dates: 400},
{month: new Date(2015, 2, 1), apples:  640, bananas:  960, cherries: 640, dates: 400},
{month: new Date(2015, 3, 1), apples:  320, bananas:  480, cherries: 640, dates: 400}
];


To generate the stack layout, first define a stack generator, and then apply it to the data:

var stack = d3.stack()
.keys(["apples", "bananas", "cherries", "dates"])
.order(d3.stackOrderNone)
.offset(d3.stackOffsetNone);

var series = stack(data);


The resulting array has one element per series. Each series has one point per month, and each point has a lower and upper value defining the baseline and topline:

[
[[   0, 3840], [   0, 1600], [   0,  640], [   0,  320]], // apples
[[3840, 5760], [1600, 3040], [ 640, 1600], [ 320,  800]], // bananas
[[5760, 6720], [3040, 4000], [1600, 2240], [ 800, 1440]], // cherries
[[6720, 7120], [4000, 4400], [2240, 2640], [1440, 1840]], // dates
]


Each series in then typically passed to an area generator to render an area chart, or used to construct rectangles for a bar chart. Stack generators no longer modify the input data, so stack.out has been removed.

For an introduction to shapes, see Introducing d3-shape.

## Time Formats (d3-time-format)

Pursuant to the great namespace flattening, the format constructors have new names:

The format.parse method has also been removed in favor of separate d3.timeParse, d3.utcParse and d3.isoParse parser constructors. Thus, this code in 3.x:

var parseTime = d3.time.format("%c").parse;


Can be rewritten in 4.0 as:

var parseTime = d3.timeParse("%c");


The multi-scale time format d3.time.format.multi has been replaced by d3.scaleTime’s tick format. Time formats now coerce inputs to dates, and time parsers coerce inputs to strings. The %Z directive now allows more flexible parsing of time zone offsets, such as -0700, -07:00, -07, and Z. The %p directive is now parsed correctly when the locale’s period name is longer than two characters (e.g., “a.m.”).

The default U.S. English locale now uses 12-hour time and a more concise representation of the date. This aligns with local convention and is consistent with date.toLocaleString in Chrome, Firefox and Node:

var now = new Date;
d3.timeFormat("%c")(new Date); // "6/23/2016, 2:01:33 PM"
d3.timeFormat("%x")(new Date); // "6/23/2016"
d3.timeFormat("%X")(new Date); // "2:01:38 PM"


You can now set the default locale using d3.timeFormatDefaultLocale! The locales are published as JSON to npm.

The performance of time formatting and parsing has been improved, and the UTC formatter and parser have a cleaner implementation (that avoids temporarily overriding the Date global).

## Time Intervals (d3-time)

Pursuant to the great namespace flattening, the local time intervals have been renamed:

The UTC time intervals have likewise been renamed:

The local time range aliases have been renamed:

The UTC time range aliases have been renamed:

The behavior of interval.range (and the convenience aliases such as d3.timeDays) has been changed when step is greater than one. Rather than filtering the returned dates using the field number, interval.range now behaves like d3.range: it simply skips, returning every stepth date. For example, the following code in 3.x returns only odd days of the month:

d3.time.days(new Date(2016, 4, 28), new Date(2016, 5, 5), 2);
// [Sun May 29 2016 00:00:00 GMT-0700 (PDT),
//  Tue May 31 2016 00:00:00 GMT-0700 (PDT),
//  Wed Jun 01 2016 00:00:00 GMT-0700 (PDT),
//  Fri Jun 03 2016 00:00:00 GMT-0700 (PDT)]


Note the returned array of dates does not start on the start date because May 28 is even. Also note that May 31 and June 1 are one day apart, not two! The behavior of d3.timeDays in 4.0 is probably closer to what you expect:

d3.timeDays(new Date(2016, 4, 28), new Date(2016, 5, 5), 2);
// [Sat May 28 2016 00:00:00 GMT-0700 (PDT),
//  Mon May 30 2016 00:00:00 GMT-0700 (PDT),
//  Wed Jun 01 2016 00:00:00 GMT-0700 (PDT),
//  Fri Jun 03 2016 00:00:00 GMT-0700 (PDT)]


If you want a filtered view of a time interval (say to guarantee that two overlapping ranges are consistent, such as when generating time scale ticks), you can use the new interval.every method or its more general cousin interval.filter:

d3.timeDay.every(2).range(new Date(2016, 4, 28), new Date(2016, 5, 5));
// [Sun May 29 2016 00:00:00 GMT-0700 (PDT),
//  Tue May 31 2016 00:00:00 GMT-0700 (PDT),
//  Wed Jun 01 2016 00:00:00 GMT-0700 (PDT),
//  Fri Jun 03 2016 00:00:00 GMT-0700 (PDT)]


Time intervals now expose an interval.count method for counting the number of interval boundaries after a start date and before or equal to an end date. This replaces d3.time.dayOfYear and related methods in 3.x. For example, this code in 3.x:

var now = new Date;
d3.time.dayOfYear(now); // 165


Can be rewritten in 4.0 as:

var now = new Date;
d3.timeDay.count(d3.timeYear(now), now); // 165


Likewise, in place of 3.x’s d3.time.weekOfYear, in 4.0 you would say:

d3.timeWeek.count(d3.timeYear(now), now); // 24


The new interval.count is of course more general. For example, you can use it to compute hour-of-week for a heatmap:

d3.timeHour.count(d3.timeWeek(now), now); // 64


Here are all the equivalences from 3.x to 4.0:

D3 4.0 now also lets you define custom time intervals using d3.timeInterval. The d3.timeYear, d3.utcYear, d3.timeMillisecond and d3.utcMillisecond intervals have optimized implementations of interval.every, which is necessary to generate time ticks for very large or very small domains efficiently. More generally, the performance of time intervals has been improved, and time intervals now do a better job with respect to daylight savings in various locales.

## Timers (d3-timer)

In D3 3.x, the only way to stop a timer was for its callback to return true. For example, this timer stops after one second:

d3.timer(function(elapsed) {
console.log(elapsed);
return elapsed >= 1000;
});


var t = d3.timer(function(elapsed) {
console.log(elapsed);
if (elapsed >= 1000) {
t.stop();
}
});


The primary benefit of timer.stop is that timers are not required to self-terminate: they can be stopped externally, allowing for the immediate and synchronous disposal of associated resources, and the separation of concerns. The above is equivalent to:

var t = d3.timer(function(elapsed) {
console.log(elapsed);
});

d3.timeout(function() {
t.stop();
}, 1000);


This improvement extends to d3-transition: now when a transition is interrupted, its resources are immediately freed rather than having to wait for transition to start.

4.0 also introduces a new timer.restart method for restarting timers, for replacing the callback of a running timer, or for changing its delay or reference time. Unlike timer.stop followed by d3.timer, timer.restart maintains the invocation priority of an existing timer: it guarantees that the order of invocation of active timers remains the same. The d3.timer.flush method has been renamed to d3.timerFlush.

Some usage patterns in D3 3.x could cause the browser to hang when a background page returned to the foreground. For example, the following code schedules a transition every second:

setInterval(function() {
}, 1000);


If such code runs in the background for hours, thousands of queued transitions will try to run simultaneously when the page is foregrounded. D3 4.0 avoids this hang by freezing time in the background: when a page is in the background, time does not advance, and so no queue of timers accumulates to run when the page returns to the foreground. Use d3.timer instead of transitions to schedule a long-running animation, or use d3.timeout and d3.interval in place of setTimeout and setInterval to prevent transitions from being queued in the background:

d3.interval(function() {
d3.selectAll("div").transition().call(someAnimation); // GOOD
}, 1000);


By freezing time in the background, timers are effectively “unaware” of being backgrounded. It’s like nothing happened! 4.0 also now uses high-precision time (performance.now) where available; the current time is available as d3.now.

## Transitions (d3-transition)

The selection.transition method now takes an optional transition instance which can be used to synchronize a new transition with an existing transition. (This change is discussed further in What Makes Software Good?) For example:

var t = d3.transition()
.duration(750)
.ease(d3.easeLinear);

d3.selectAll(".apple").transition(t)
.style("fill", "red");

d3.selectAll(".orange").transition(t)
.style("fill", "orange");


Transitions created this way inherit timing from the closest ancestor element, and thus are synchronized even when the referenced transition has variable timing such as a staggered delay. This method replaces the deeply magical behavior of transition.each in 3.x; in 4.0, transition.each is identical to selection.each. Use the new transition.on method to listen to transition events.

The meaning of transition.delay has changed for chained transitions created by transition.transition. The specified delay is now relative to the previous transition in the chain, rather than the first transition in the chain; this makes it easier to insert interstitial pauses. For example:

d3.selectAll(".apple")
.transition() // First fade to green.
.style("fill", "green")
.transition() // Then red.
.style("fill", "red")
.transition() // Wait one second. Then brown, and remove.
.delay(1000)
.style("fill", "brown")
.remove();


Time is now frozen in the background; see d3-timer for more information. While it was previously the case that transitions did not run in the background, now they pick up where they left off when the page returns to the foreground. This avoids page hangs by not scheduling an unbounded number of transitions in the background. If you want to schedule an infinitely-repeating transition, use transition events, or use d3.timeout and d3.interval in place of setTimeout and setInterval.

The selection.interrupt method now cancels all scheduled transitions on the selected elements, in addition to interrupting any active transition. When transitions are interrupted, any resources associated with the transition are now released immediately, rather than waiting until the transition starts, improving performance. (See also timer.stop.) The new d3.interrupt method is an alternative to selection.interrupt for quickly interrupting a single node.

The new d3.active method allows you to select the currently-active transition on a given node, if any. This is useful for modifying in-progress transitions and for scheduling infinitely-repeating transitions. For example, this transition continuously oscillates between red and blue:

d3.select("circle")
.transition()
.on("start", function repeat() {
d3.active(this)
.style("fill", "red")
.transition()
.style("fill", "blue")
.transition()
.on("start", repeat);
});


The life cycle of a transition is now more formally defined and enforced. For example, attempting to change the duration of a running transition now throws an error rather than silently failing. The transition.remove method has been fixed if multiple transition names are in use: the element is only removed if it has no scheduled transitions, regardless of name. The transition.ease method now always takes an easing function, not a string. When a transition ends, the tweens are invoked one last time with t equal to exactly 1, regardless of the associated easing function.

As with selections in 4.0, all transition callback functions now receive the standard arguments: the element’s datum (d), the element’s index (i), and the element’s group (nodes), with this as the element. This notably affects transition.attrTween and transition.styleTween, which no longer pass the tween function the current attribute or style value as the third argument. The transition.attrTween and transition.styleTween methods can now be called in getter modes for debugging or to share tween definitions between transitions.

Homogenous transitions are now optimized! If all elements in a transition share the same tween, interpolator, or event listeners, this state is now shared across the transition rather than separately allocated for each element. 4.0 also uses an optimized default interpolator in place of d3.interpolate for transition.attr and transition.style. And transitions can now interpolate both CSS and SVG transforms.

For reusable components that support transitions, such as axes, a new transition.selection method returns the selection that corresponds to a given transition. There is also a new transition.merge method that is equivalent to selection.merge.

For the sake of parsimony, the multi-value map methods have been extracted to d3-selection-multi and are no longer part of the default bundle. The multi-value map methods have also been renamed to plural form to reduce overload: transition.attrs and transition.styles.

## Voronoi Diagrams (d3-voronoi)

The d3.geom.voronoi method has been renamed to d3.voronoi, and the voronoi.clipExtent method has been renamed to voronoi.extent. The undocumented polygon.point property in 3.x, which is the element in the input data corresponding to the polygon, has been renamed to polygon.data.

Calling voronoi now returns the full Voronoi diagram, which includes topological information: each Voronoi edge exposes edge.left and edge.right specifying the sites on either side of the edge, and each Voronoi cell is defined as an array of these edges and a corresponding site. The Voronoi diagram can be used to efficiently compute both the Voronoi and Delaunay tessellations for a set of points: diagram.polygons, diagram.links, and diagram.triangles. The new topology is also useful in conjunction with TopoJSON; see the Voronoi topology example.

The voronoi.polygons and diagram.polygons now require an extent; there is no longer an implicit extent of ±1e6. The voronoi.links, voronoi.triangles, diagram.links and diagram.triangles are now affected by the clip extent: as the Delaunay is computed as the dual of the Voronoi, two sites are only linked if the clipped cells are touching. To compute the Delaunay triangulation without respect to clipping, set the extent to null.

The Voronoi generator finally has well-defined behavior for coincident vertices: the first of a set of coincident points has a defined cell, while the subsequent duplicate points have null cells. The returned array of polygons is sparse, so by using array.forEach or array.map, you can easily skip undefined cells. The Voronoi generator also now correctly handles the case where no cell edges intersect the extent.

## Zooming (d3-zoom)

The zoom behavior d3.behavior.zoom has been renamed to d3.zoom. Zoom behaviors no longer store the active zoom transform (i.e., the visible region; the scale and translate) internally. The zoom transform is now stored on any elements to which the zoom behavior has been applied. The zoom transform is available as event.transform within a zoom event or by calling d3.zoomTransform on a given element. To zoom programmatically, use zoom.transform with a given selection or transition; see the zoom transitions example. The zoom.event method has been removed.

To make programmatic zooming easier, there are several new convenience methods on top of zoom.transform: zoom.translateBy, zoom.scaleBy and zoom.scaleTo. There is also a new API for describing zoom transforms. Zoom behaviors are no longer dependent on scales, but you can use transform.rescaleX, transform.rescaleY, transform.invertX or transform.invertY to transform a scale’s domain. 3.x’s event.scale is replaced with event.transform.k, and event.translate is replaced with event.transform.x and event.transform.y. The zoom.center method has been removed in favor of programmatic zooming.

The zoom behavior finally supports simple constraints on panning! The new zoom.translateExtent lets you define the viewable extent of the world: the currently-visible extent (the extent of the viewport, as defined by zoom.extent) is always contained within the translate extent. The zoom.size method has been replaced by zoom.extent, and the default behavior is now smarter: it defaults to the extent of the zoom behavior’s owner element, rather than being hardcoded to 960×500. (This also improves the default path chosen during smooth zoom transitions!)

The zoom behavior’s interaction has also improved. It now correctly handles concurrent wheeling and dragging, as well as concurrent touching and mousing. The zoom behavior now ignores wheel events at the limits of its scale extent, allowing you to scroll past a zoomable area. The zoomstart and zoomend events have been renamed start and end. By default, zoom behaviors now ignore right-clicks intended for the context menu; use zoom.filter to control which events are ignored. The zoom behavior also ignores emulated mouse events on iOS. The zoom behavior now consumes handled events, making it easier to combine with other interactive behaviors such as dragging.