Allow expressions in addition to keyPaths when creating indexes

[inexorabletash]

Imported from https://www.w3.org/Bugs/Public/show_bug.cgi?id=10000

Discussion here: http://lists.w3.org/Archives/Public/public-webapps/2010AprJun/1094.html

[inexorabletash]

Lots of stuff to work through here, e.g. do we make expressions side-effect free by running in isolated execution environments? (VM implementers say: that's not cheap!) Also, loading libraries (e.g. text segmentation, ...) means these are probably non-trivial expressions... so do we abandon the notion that these are pure functions over the values and give sites the ability to do dumb (dangerous?) things?

[inexorabletash]

Worklets might be the way to go for this. They do require developers to be consistent, but that seems like an acceptable constraint.

Very rough sketch: I haven't thought this through in detail, but something like:

• IDBDatabase gets a worklet property; code can be loaded in via db.worklet.import(...)
• IDBWorklet exposes a registerIndexExpression(name, func) global function
• When creating an index, you can specify name() as the keyPath

Developers need to remember to call db.worklet.import(...) on every connection to the DB that might alter an index (create or add/update values). This import is also async, so maybe a new step in the open() phase. So the boilerplate would look like:

var open = indexedDB.open('stuff');
// new "registerworklet" phase always precedes "upgradeneeded" (if needed) and "success"
open.onregisterworklet = function() {
  var db = open.result;
  db.worklet.import('my_worklet_code.js');
};
// "upgradeneeded" implicitly delayed until all worklet imports are complete
open.onupgradeneeded = function() {
  var db = open.result;
  var store = db.createObjectStore('records');
  store.createIndex('freeText', 'words()', {multiEntry: true});
  store.createIndex('area', 'area()');
};
open.onsuccess = function() {
  var db = open.result;
  ...
};

And my_worklet_code.js would look like:

registerIndexExpression('area', r => r.width * r.height);
registerIndexExpression('words', r => r.body.split(/\s+/));

[inexorabletash]

Whether the binding (name -> function) should occur in the owning context or the worklet context should be consistent with other uses of worklets.

[bfgeek]

This looks pretty fun :)

Looks like result of an index can only be primitives + array + binary?

Is it possible for a index to have multiple types? I.e.

registerIndexExpression('foo', r => { return r.bool ? r.str : r.arrayType });

[inexorabletash]

The "result" is a key, which (in Indexed DB) is a number, date, string, binary (proposed!), or array-of-key. See https://w3c.github.io/IndexedDB/#key-construct. Boolean is not supported (but see #76).

There's no requirement that keys within a store or index have the same type or "shape", so your example is valid.

[inexorabletash]

If you're watching this issue...

One of the primitives necessary for building full text search on top of indexes is text segmentation. @littledan has a proposal for adding this to ECMA-402 (the ECMAScript Intl object). He'd love feedback:

• Proposal: https://github.com/littledan/Segmenter/blob/master/README.md
• Slides: https://docs.google.com/presentation/d/1X2zBU3bZ4ergVMWfubCsdnHFzeaDgqiTRJVgvNGjQBs/edit#slide=id.g17f706c31b_0_63