Skip to content

Latest commit

 

History

History
308 lines (213 loc) · 11.3 KB

NOTES.md

File metadata and controls

308 lines (213 loc) · 11.3 KB
Raw

Browser Features caniuse

Interactions

Way drawing strategy:

Continuing Ways

Continuing ways is a tricky problem: for instance, fixing unclosed areas by completing the loop.

Potlatch does this by clicking on the way (since it does not show handles by default), showing handles, and if you click on the start or end, it immediately starts drawing. If you escape out of that draw, you can then edit the node itself.

JOSM, since it has an add mode, just selects the node if you click a start or end. You need to be in draw mode, and then clicking on a start/end node extends the way.

However JOSM also does a little trick where if you have a node selected and then enter draw mode, it'll start either a new way or a continuation of the way, depending on whether it's the start or the end.

Also what happens if you click on a point that happens to be the start or end of more than one way?

Relations and Turn Restrictions

http://wiki.openstreetmap.org/wiki/Relation:restriction

Pathological conditions

  • Ways with one node
  • Relations which contain themselves (circular references)
  • Nodes with no tags and no way attached
  • Ways which contain only nodes that are subsets of the nodes of other ways
  • Paths with intersecting boundaries (invalid geometries)
  • Nodes with tags that repeat what the way says

The Graph

The data model of OSM is something like

root -> relations (-> relations) -> ways -> nodes
   \                             \> nodes
    \-  ways -> nodes
     \- nodes

In English:

  • Relations have (ways, nodes, relations)
  • Ways have (nodes)
  • Nodes have ()

iD implements a persistent data structure over the OSM data model. Instead of updating the graph in place when edits are made and storing enough information about the change so that it can be undone in place, changes produce a new version of the graph data structure, and the previous version is left untouced. "Undo" is accomplished simply by reverting to the previous version.

The persistent data structure approach also takes advantage of the fact that the typical change modifies only a small portion of the graph. The unchanged majority of the graph is shared between revisions, keeping memory use to a minimum. For example, the iD.actionRemoveWayNode action removes a single node from a way. It produces new versions of three objects:

  • The Array of nodes in the way.
  • The way itself. The new version references the new Array of nodes.
  • The Graph. The new version references the new way.

The previous versions of these three objects are retained. Since the previous version of the Graph continues to reference the previous version of the way and its nodes, the action can be undone by restoring this version. Meanwhile, both versions of the Graph share references to all the other objects. Since these objects are never themselves mutated, this is safe.

In concrete terms, this approach dictates the following rule: all methods that produce a change in the state of the data model objects (Entity, Graph) or their constituent parts (e.g. nodes Array, tags Object) must return a new instance of the appropriate type, leaving the current instance unchanged.

UI

Rendering and UI code generally follows a convention for reusable elements established by d3.

Constructor functions typically return a function decorated with additional properties. The function can be called with this set to a d3 selection in order to set up the HTML structure. This is usually done via selection.call.

Accessors are implemented as a unified getter/setter function. When called with no arguments, it acts as a getter; when called with an argument it acts as a setter and returns self, for chaining. Accessors are preferable to constructor arguments; constructors typically take zero arguments.

Performance

See blog post: http://mapbox.com/osmdev/2012/11/20/getting-serious-about-svg/

Main performance concerns of iD:

Panning & zooming performance of the map

SVG redraws are costly, especially when they require all features to be reprojected.

Approaches:

  • Using CSS transforms for intermediate map states, and then redrawing when map movement stops
  • "In-between" projecting features to make reprojection cheaper

Memory overhead of objects

Many things will be stored by iD. With the graph structure in place, we'll be storing much more.

We also need to worry about memory leaks, which have been a big problem in Potlatch 2. Storing OSM data and versions leads to a lot of object-referencing in JavaScript.

Connection, Graph, Map

The Map is a display and manipulation element. It should have minimal particulars of how exactly to store or retrieve data. It gets data from Connection and asks for it from Graph.

Graph stores all of the objects and all of the versions of those objects. Connection requests objects over HTTP, parses them, and provides them to Graph.

loaded

The .loaded member of nodes and ways is because of relations, which refer to elements, so we want to have real references of those elements, but we don't have the data yet. Thus when the Connection encounters a new object but has a non-loaded representation of it, the non-loaded version is replaced.

Prior Art

Editors:

JOSM and Potlatch 2 appear to implement versioning in the same way, but having an undo stack:

// src/org/openstreetmap/josm/actions/MoveNodeAction.java
Main.main.undoRedo.add(new MoveCommand(n, coordinates));

// src/org/openstreetmap/josm/command/MoveCommand.java

/**
 * List of all old states of the objects.
 */
private List<OldState> oldState = new LinkedList<OldState>();

@Override public boolean executeCommand() {
// ...
}
@Override public void undoCommand() {
// ...
}

Transforms Performance

There are two kinds of transforms: SVG and CSS. CSS transforms of SVG elements are less efficient that SVG transforms of SVG elements. translate notation has equivalent performance to matrix notation.

SVG transforms are a roughly 2x speedup relative to CSS - 16fps vs 32fps in Google Chrome Beta.

However, using CSS transforms with HTML elements has vastly different and better performance than using them with SVG elements. For this reason, iD transforms a map-container element rather than a g element on panning movements.

Transforms in browsers

Matrix transforms are significantly slower than translate in webkit but nearly equivalent in Firefox. Chrome is about 4x faster than Firefox with transforms.

However, matrix transforms can also represent scale, and so they should be compared with transform + scale. If you add an identity scale (scale(1, 1)), then matrix and translate scale performance is similar in Chrome, though matrix still lags significantly in Safari and Firefox.

SVG point rounding performance

Rounding points in SVG gives a ~20% speedup.

And this is not just the effect of less d data:

SVG Corner Cases

One-way streets need markers to indicate that they're one-way. Unfortunately SVG line markers are based strictly off of vertices, so won't handle this case properly.

One way to resolve this is by using textPath with a glyph, like a gt sign or triangle character if available. This has a few concerns:

  • performance of textPath is known to suck in some cases. For simple cases, it is fine
  • can we be absolutely sure about direction of text?
  • glyphs need to be available. are webfonts svg-okay?

Or more importantly, we need to calculate the pixel length of a linestring, calculate the width of a glyph, and do the necessary math so that it fills enough of the line without overflowing.

See the textPath element and its quirks.

See:

Authenticating

The OAuth endpoint of OpenStreetMap does support CORS, which is great and allows iD to do browser-side authentication. This requires some hacks, mainly that a cookie is used to persist the token_secret between pageloads.

Making Edits

PUT /api/0.6/changeset/create
POST /api/0.6/changeset/135324/upload
PUT /api/0.6/changeset/135324/close

Browser problems that affect iD

See also: Kothic browser bugs.

one-way streets use glyphs and textPaths. letter-spacing is not supported in Firefox but is in webkit so we need to use spaces.

And trailing spaces are not included in getComputedTextLength:

webkit doesn't let querySelectorAll select camelcase elements:

Firefox does not fire a blur event when an element is removed from the DOM, unlike WebKit browsers.

Firefox does not support the focusout event.

Opera does not support pointer-events on HTML elements, only SVG elements.

IE10 does not support navigator.language, only navigator.userLanguage

Transients

The graph supports transient, which is storage for non-versioned mutable properties about entities that are stored outside of entities. For instance, extent is about an entity, but can be invalidated and stored without getting a new graph.

Data sources

deprecated.json from http://wiki.openstreetmap.org/wiki/Deprecated_features TODO: deal with deprecated 'class' tag does not deal with landuse=wood because of indecision we will not care about http://taginfo.openstreetmap.org/tags/bicycle_parking=sheffield

discarded.json entirely discarded tags https://github.com/systemed/potlatch2/blob/master/net/systemeD/halcyon/connection/XMLConnection.as#L24