When doing kdbGet() possible more keys are returned which might be confusing. When doing a second kdbGet() with a new keyset no keys might be returned, because it is up-to-date.

When doing kdbSet() a deep duplication is needed.

Idea: keep a duplicated keyset internally. Return (non-deep?) duplications exactly of the parts as requested.

no cache (current situation)
flat cache with COW
deep duplicated cache
Use global plugin to implement internal cache.
Use global mmap plugin and make kdbGet() to always retrieve everything.

Not yet decided.

Semantics can be provided without additional code or overhead in the core.

Global Validation

internal caches lead to duplication of memory consumption (most of it is avoided by reference counting, though)

in some cases caches cannot be avoided? -> filesys, databases?

Pros:

not more keys than needed
kdbGet avoids IO even if done somewhere else
KDB handles could be more locally

Cons:

not possible to access cache with current architecture, KDB high level API
implementation overhead
where should the caches be