PhysicalStreamingWindow operator

[lnkuiper]

This PR implements a streaming variant of the PhysicalWindow operator, which is selected when we have a window function, and the following is true:

1. The OVER clause is empty
2. There is no IGNORE NULLS clause
3. The window function is one of {first_value, percent_rank, rank, rank_dense, row_number}

The operator does not work in parallel (but is now deterministic!), but does give a nice speedup for these specific cases because it does not materialize anything.

Quick benchmark on TPC-H SF1:

select sum(fv) from (select first_value(l_orderkey) over () as fv from lineitem) sq;

Old	New
0.116	0.008

[Mytherin]

Thanks for the PR! Looks great. @hawkfish maybe you want to have a look as well?

[hawkfish]

Looks great - just a few small nits.

[lnkuiper]

Thanks for the feedback! Hopefully good to go now.

[hawkfish]

Can we tweak the operator ordering?

[lnkuiper]

@hawkfish I've tweaked the operator ordering. Thanks for the feedback.

I was wondering what you thought about this:

# When we combine a streaming and non-streaming window function over the same window,
# We get a non-streaming window operator

Is this something we want? Or would we rather have a non-streaming and a streaming window for these queries?

[hawkfish]

@hawkfish I've tweaked the operator ordering. Thanks for the feedback.

I was wondering what you thought about this:

When we combine a streaming and non-streaming window function over the same window,
We get a non-streaming window operator

Is this something we want? Or would we rather have a non-streaming and a streaming window for these queries?

You mean if they have the same partitioning and ordering? Like maybe

SELECT last_value(i) OVER(), avg(i) OVER() FROM table;

?

I suspect it makes sense to divide them so that the streaming goes last and we don't have to materialize the streamable columns. But since they will be constants (or sequences?) maybe that is not such a big deal.

[lnkuiper]

@hawkfish Thanks, that makes sense. I've separated them.

[lnkuiper]

@Mytherin Should be good to go now. I believe the CI failures are unrelated.

[Mytherin]

Thanks! Looks good.