Profiling performance of snapshot reads is showing that NormalizedNodeInputStreamReader.readQName() is dominating, taking 53% of total CPU cycles. Of this, 43% is spent in QNameFactory.create(String) and 19% creating Strings for that lookup.

Rework QNameFactory.create() to work on structured data holding namespace/revision/localName without having to concatenate, which will save the string manipulation as well as speeding up hash-based lookup (as component String hashcodes are likely to be pre-computed).