This also caused un-ordered sliced prefetches to crash as they rely on Window.
Regression in e16d0c176e that made OrderByList
piggy-back ExpressionList without porting the empty handling that the latter
provided.
Supporting explicit empty ordering on Window functions and slicing is arguably
a foot-gun design due to how backends will return undeterministic results but
this is a problem that requires a larger discussion.
Refs #35064.
Thanks Andrew Backer for the report and Mariusz for the review.
This avoids cast of Window(order_by) for DecimalFields on SQLite.
This was achieved by piggy-backing ExpressionList which already
implements a specialized as_sqlite() method to override the inherited
behaviour of Func through SQLiteNumericMixin.
Refs #31723.
Thanks Quoates for the report.
Regression in f387d024fc.
Just like `OrderByList` the `ExpressionList` expression used to wrap
`Window.partition_by` must implement `get_group_by_cols` to ensure the
necessary grouping when mixing window expressions with aggregate
annotations is performed against the partition members and not the
partition expression itself.
This is necessary because while `partition_by` is implemented as
a source expression of `Window` it's actually a fragment of the WINDOW
expression at the SQL level and thus it should result in a group by its
members and not the sum of them.
Thanks ElRoberto538 for the report.
Adds support for joint predicates against window annotations through
subquery wrapping while maintaining errors for disjointed filter
attempts.
The "qualify" wording was used to refer to predicates against window
annotations as it's the name of a specialized Snowflake extension to
SQL that is to window functions what HAVING is to aggregates.
While not complete the implementation should cover most of the common
use cases for filtering against window functions without requiring
the complex subquery pushdown and predicate re-aliasing machinery to
deal with disjointed predicates against columns, aggregates, and window
functions.
A complete disjointed filtering implementation should likely be
deferred until proper QUALIFY support lands or the ORM gains a proper
subquery pushdown interface.
Even though good databases tend to keep the result sorted by the/one
window expression and the planners are smart enough to not resort if not
required, it is not valid to rely on this.
MariaDB specifically did return whatever order it wanted, which is
completely okay. Now we sort towards the expected data for all databases.
Thanks Josh Smeaton, Mariusz Felisiak, Sergey Fedoseev, Simon Charettes,
Adam Chainz/Johnson and Tim Graham for comments and reviews and Jamie
Cockburn for initial patch.
