Provide alternate SQL to allow "tpch" Q15 validation with floats

[GregoryKimball]

As of March 2026, we are avoiding validation on "tpch" Q15.

velox-testing/.github/workflows/presto-test-composite.yml

Line 99 in f8d9835

# avoid Q15 on GPU as it will fail output comparison until the schema differences are resolved

The SQL is here:
https://github.com/rapidsai/velox-testing/blob/f8d98355cbccf9e19aa1e42c2eb3c81272f0b69d/common/testing/queries/tpch/queries.json

"Q15": "WITH revenue AS ( SELECT l_suppkey AS supplier_no, sum(l_extendedprice * (1 - l_discount)) AS total_revenue FROM lineitem WHERE l_shipdate >= CAST('1996-01-01' AS date) AND l_shipdate < CAST('1996-04-01' AS date) GROUP BY supplier_no ) SELECT s_suppkey, s_name, s_address, s_phone, total_revenue FROM supplier, revenue WHERE s_suppkey = supplier_no AND total_revenue = ( SELECT max(total_revenue) FROM revenue) ORDER BY s_suppkey"

I propose that we change the SQL run with floats to:

"Q15": "WITH revenue AS ( SELECT l_suppkey AS supplier_no, ROUND(SUM(l_extendedprice * (1 - l_discount)), 2) AS total_revenue FROM lineitem WHERE l_shipdate >= CAST('1996-01-01' AS date) AND l_shipdate < CAST('1996-04-01' AS date) GROUP BY supplier_no ) SELECT s_suppkey, s_name, s_address, s_phone, total_revenue FROM supplier, revenue WHERE s_suppkey = supplier_no AND total_revenue = ( SELECT ROUND(MAX(total_revenue), 2) FROM revenue) ORDER BY s_suppkey"

so that we can validate on Q15.

[GregoryKimball]

@devavret @quasiben @misiugodfrey What do you think?

[karthikeyann]

I expected query with round to have more nodes. But it actually less nodes.
Extra Project, LocalExchange in original TPC-H Q15 plan

[devavret]

@GregoryKimball There's just this one small issue: We can't round a float using libcudf. This used to be allowed but was deprecated for floating point types. https://github.com/rapidsai/cudf/blob/8d69225d91683a84bc464b88c7d1a5f3b1b95808/cpp/include/cudf/round.hpp#L66

Now only decimals and integers can be rounded. https://github.com/rapidsai/cudf/blob/8d69225d91683a84bc464b88c7d1a5f3b1b95808/cpp/include/cudf/round.hpp#L105

[GregoryKimball]

We can't round a float using libcudf

True, we would need to cast to decimal, round and cast back to float.