If the writer's and reader's schema are records, we allow the order of fields to be different in them. The ResolvingDecoder returns the fields in the writer's order. This is to avoid buffering. The current implementation uses FieldAdjust action on the parser stack. The number of such action symbols is equal to the number of reader fields. This causes performance problem because the number of calls to advance() is almost double compared to the number of calls in ValidatingDecoder.
This patch replaces the FieldAdjustAction symbols with FieldOrderAction symbols. We have FieldOrderActions one per record, which is expected to be much smaller than the number of fields. Though it changes the API for ResolvingDecoder slightly, there is no impact because we do not have any users yet.
I see a 10-15% improvement in my computer with Perf -S.