Issue #3207: ASOF JOIN Compilation

[hawkfish]

Part 1 of the ASOF JOIN implementation.

This PR adds ASOF to the grammar and converts it into an inequality join on a window:

SELECT p.key, p.ts, e.value 
FROM probe p ASOF JOIN event e
  ON p.key = e.key AND p.ts >= e.ts

-- Translates to

SELECT p.key, p.ts, e.value
FROM probe p LEFT JOIN (
  SELECT value, ts, LEAD(ts, 1, 'infinity') OVER (PARTITION BY key ORDER BY ts) AS ts_temp
  ) e
  ON p.key = e.key AND p.ts >= e.ts AND p.ts < e.ts_temp

where the left join is implemented using the IEJoin operator.

This is not very efficient (it does three sorts), and part 2 will add a new physical operator that uses only one partitioned sort and the outer partitioning from hash join. But it will hopefully make it easier to specify these joins in a future-proof manner.

[Mytherin]

Thanks for the PR! Looks good. Some comments:

[Mytherin]

Thanks for the changes! Looks great. Some additional comments:

[Mytherin]

Could we add some extra tests:

• Join on a timestamp range
• Join on a string range (it is fine if that throws a non-internal exception)
• Join on integer ranges (it is fine if that throws a non-internal exception)
• Multiple ranges in one join condition
• What if we invert the comparisons (e.g. e.begin <= p.ts AND e.end > p.ts)
• What if we have a different closed range (e.g. p.ts > e.begin AND p.ts <= e.end)
• What if we have a range that is not actually a range (e.g. p.ts >= e.begin and p.other_col < e.end)
• ASOF join on a constant (e.g. ON 1=1)
• Could we add some tests where we have NULL values in the tables?
• Could we add some tests where we have infinity values (e.g. NaN, or infinity::timestamp) in the tables?
• What happens if we use an ASOF join inside of a correlated subquery? (it is fine if that throws an exception)

[hawkfish]

Multiple ranges was covered by the last error test, and you can't have invalid ranges because the bounds are implied (i.e., end is synthesised). But I'll have a go at the rest.

[hawkfish]

Got all of these except the last one. How would you write an ASOF join inside a correlated subquery? Aren't the joins in correlated subqueries generated by the planner?

[Mytherin]

Ah, I meant an ASOF join inside a subquery that has correlated columns in the left and/or right side, see e.g. here

[Mytherin]

Thanks for the fixes! Looks great.