RB design discussion - 2009-08-21
(ICU4J) Loading entire data into heap as raw bytes, creating a Java object when accessed. Without Java object caching, it creates multiple identical objects (like former implementation). Even with caching, raw byte data is not released.
(ICU4J) Bundle size is too big. For example, time zone data (zoneinfo.res) is currently 155kb. ICU4J load entire res contents to heap. In many cases, client only need a single time zone, which require only several hundreds byte maximum.
(ICU4J) Getting a resource item from our own proprietary format always require a new Java object instantiation and probably always slower than getting a Java object from .class file.
Usability / Design
No public API for supporting inheritance (fallback to parents) for nested items.
(ICU4J) UResourceBundle does not support JDK 6 ResourceBundle.Control features, which allow user to customize look up path, controlling cache, etc.
(ICU4J) UResourceBundle extends JDK ResourceBundle, but its design is largely different from JDK.
Break ICU locale resources into smaller pieces - package items optimized for typical usage models.
Doug to come up with resource type categorization proposal.
ICU4J to use .class file to store locale resources (such as ListResourceBundle)
This could introduce ICU4C / J incompatibility
We could add "locale data API" compatible between C and J.
Duplicated (or substring) key optimization can be done by tooling.
With smaller bundles, we should be able to avoid the Java class file size limitation.
Large binary items might need to be externalized to avoid the Java class file limitation (e.g. collation rules)
We need a way to verify the contents is equivalent to .res. We should probably start with .txt <-> .class. In longer term, we want to generate locale resources directly from LDML.
We agreed that Yoshito to develop .class resource type based on the existing code for evaluation. #7116
Deprecated ICU4J UResourceBundle and start new one from scratch. Leave UResourceBundle class as is.
Low impact to ICU4J users. We've never heard of clients using ICU4J UResourceBundle API. For ICU4C, we know there are many clients using ures APIs, so we probably need incremental improvement.
getInt / getUInt
Yoshito/Doug think signed or unsigned should be clearly tagged at the resource level - which is not done in the current architecture.
We still need both unless we change add resource types to identify signed or unsigned.
We still need to support "table" without inheritance.
Resource array won't support array item inheritance.