Attached is the patch against HEAD to support more frame types in
window functions, including these frame types:
- ROWS BETWEEN s PRECEDING AND e PRECEDING
- ROWS BETWEEN s PRECEDING AND CURRENT ROW
- ROWS BETWEEN s PRECEDING AND e FOLLOWING
- ROWS BETWEEN s PRECEDING AND UNBOUNDED FOLLOWING
- ROWS BETWEEN CURRENT ROW AND e FOLLOWING
- ROWS BETWEEN CURRENT ROW AND UNBOUNDED FOLLOWING
- ROWS BETWEEN s FOLLOWING AND e FOLLOWING
- ROWS BETWEEN s FOLLOWING AND UNBOUNDED FOLLOWING
- RANGE BETWEEN CURRENT ROW AND UNBOUNDED FOLLOWING
- ROWS s PRECEDING
which means that for ROWS types we now support almost all types but
for RANGE types we don't have "value PRECEDING" / "value FOLLOWING".
I'm planning to implement them until CommitFest:2010-01 so this is
"Request for Review" phase though I've arranged the patch to be
committable.
Aggregate cache mechanism is sometimes cleared as discussed the other
day. But it should be kept that the original cache mechanism for basic
(i.e. already implemented) frame types.
Some points to be reviewed are:
- Added WindowFrameDef in parsenode.h
- Is A_Const member for startOffset / endOffset is appropriate?
- Are those data types (including gram.y) well designed?
- For basic frame types, are aggregates still optimized as before?
- Added regression tests but enough?
- All error case covered? For example: ROWS s FOLLOWING AND e PRECEDING
- Added members to WindowAggState to track more information of frame
types, but isn't there more efficient way?
- Not modified docs yet
Regards,
--
Hitoshi Harada
