Flat CLDR Locale Data
Among the semantic and API issues with ICU resource bundles is the fallback of nested (non-top-level) resource items.
ICU resource bundles mostly behave like Java and Windows resource bundles: They perform fallback on the top level, and, like Windows resource bundles, they support structured data. (In the case of ICU, the structure is recursive and arbitrarily nestable, unlike Windows which provides fixed types like dialog box resources.) The idea is that you don't want the x coordinate of a dialog box to fallback, but only let the lookup for the whole dialog box fall back.
Fallback of nested items is bolted on top of the regular logic for the purpose of implementing CLDR locale data fallbacks while more or less retaining the nested structure of LDML.
One way to deal with this is to generalize and make public API for optional nested fallback.
Another way to deal with this is to flatten the locale data in ICU locale resource bundles, just like the CLDR tools code internally flattens the data using XPaths.
We could flatten the paths to single keys for every LDML item that is an atomic unit and is subject to fallbacks. Aside from most simple LDML items, this includes whole collation tailorings: For example, we would have a flat key for collations/phonebook which would access a Table with Sequence, Version etc. because there is no nested fallback for those. CLDR tools already have data and logic for distinguishing paths and atomic units.
With flattened locale data, there is no need to supported nested-item fallbacks, and we can remove the code, internal APIs, and semantic confusions about them.
Looking at this another way, it would keep the resource bundle mechanism simple, and shift the complexity of dealing with CLDR locale data out to CLDR-specific code.
Paths Flattened to Keys
Logically, we would use top-level key strings that are LDML distinguishing XPaths.
TODO: Give examples of actual LDML XPaths.
The path keys would be fairly long. They would be shared via the pool bundle, but the pool bundle would probably be significantly larger than the current 15kB.
TODO(markus): Measure the size of a pool bundle that has every resource path in ICU turned into a single key.
The LDML2ICUConverter could generate compact path keys. The runtime code could either use such compact keys directly, or map from readable keys to compact ones.
We need to figure out whether it needs to be possible and easy to parse compact keys and map them back to full paths.
Keys could be compacted by turning most path elements into one or two characters each. Path elements from large sets could be left untouched (e.g., language and region codes) or mapped to short-ish strings if there is already an established mapping (e.g., 5-letter UN codes for time zones). The mapping for each path element needs to be unique only at the level at which it occurs because it will be prefixed by higher-level path elements.
For keeping keys parseable and easily truncatable, if we don't use path element separator punctuation, then we may need another way to indicate the start or end of a path element, for example by titlecasing the short form of each path element.
For example:
date/calendar/gregorian/months/format/narrow/1 → DCGrMFN1
languages/gsw → LGsw
In code, we might have an LDMLKey class that maps from a readable form to the compact form. We might have an LDMLElements enumeration and use it like
myKey = LDMLKey.make(LDMLElements.date, LDMLElements.calendar, LDMLElements.gregorian, ...);
value = de_AT_bundle.getString(myKey + "1"); // get the de_AT month name for January
Alternatively, keys could be compacted by calculating 18-bit hash codes for them and turning each hash code into a 3-character base64 string. If there are collisions, we may need to modify the hash function. This would probably be fastest and simplest, but make keys completely unreadable and unparseable. (In other words, while this might be the most compact key string storage, it may not lend itself to the kind of processing we need to do for CLDR data.)
One consideration is at what nodes in the LDML structure we need to be able to enumerate the available items. At such nodes we probably need keys that we can tokenize into path elements, and calculate the parent path key as a key prefix.