test(tpcds): add queries 28-63 (#9736)

Author: cpcloud

.codespellrc

@@ -1,6 +1,6 @@
 [codespell]
 # local codespell matches `./docs`, pre-commit codespell matches `docs`
-skip = *.lock,.direnv,.git,./docs/_freeze,./docs/_output/**,./docs/_inv/**,docs/_freeze/**,*.svg,*.css,*.html,*.js,ibis/backends/tests/tpc/queries/duckdb/ds/44.sql
+skip = *.lock,.direnv,.git,./docs/_freeze,./docs/_output/**,./docs/_inv/**,docs/_freeze/**,*.svg,*.css,*.html,*.js,ibis/backends/tests/tpc/queries/duckdb/ds/*.sql
 ignore-regex = \b(i[if]f|I[IF]F|AFE|alls)\b
 builtin = clear,rare,names
 ignore-words-list = tim,notin,ang

ibis/backends/tests/tpc/README.md

@@ -0,0 +1,25 @@
+# TPC queries with Ibis
+
+These tests perform correctness tests against backends that are able to run
+some of the TPC-H and TPC-DS queries.
+
+The text queries are assumed to be correct, and also that if transpiled
+correctly will produce the same results as the written Ibis expression.
+
+**This is the assertion being made in these tests.**
+
+The ground truth SQL text is taken from
+[DuckDB](https://github.com/duckdb/duckdb/tree/main/extension/tpcds/dsdgen/queries)
+and transpiled using SQLGlot to the dialect of whatever backend is under test.
+
+Some queries are altered from the upstream DucKDB repo to have static column
+names and to cast strings that are dates explicitly to dates so that pedantic
+engines like Trino will accept these queries. These alterations do not change
+the computed results of the queries.
+
+ClickHouse is a bit odd in that queries that contain a cross join with an `OR`
+condition common to all operands of the `OR` will effectively never finish.
+This is probably a bug in ClickHouse.
+
+For that case, the queries for clickhouse have been minimally rewritten to pass
+by extracting the common join condition out into a single `AND` operand.

ibis/backends/tests/tpc/conftest.py

@@ -52,9 +52,18 @@ def tpc_test(suite_name: Literal["h", "ds"], *, result_is_empty=False):
     def inner(test: Callable[..., ir.Table]):
         name = f"tpc{suite_name}"
 
-        @getattr(pytest.mark, name)
+        # so that clickhouse doesn't run forever when we hit one of its weird cross
+        # join performance black holes
+        #
+        # trino can sometimes take a while as well, especially in CI
+        #
+        # func_only=True doesn't include the fixture setup time in the duration
+        # of the test run, which is important since backends can take a hugely
+        # variable amount of time to load all the TPC-$WHATEVER tables.
+        @pytest.mark.timeout(60, func_only=True)
         @pytest.mark.usefixtures("backend")
         @pytest.mark.xdist_group(name)
+        @getattr(pytest.mark, name)
         @functools.wraps(test)
         def wrapper(*args, backend, **kwargs):
             backend_name = backend.name()
@@ -94,17 +103,25 @@ def wrapper(*args, backend, **kwargs):
 
             assert result_expr._find_backend(use_default=False) is backend.connection
             result = backend.connection.to_pandas(result_expr)
-            assert (result_is_empty and result.empty) or not result.empty
+
+            assert (result_is_empty and result.empty) or (
+                not result_is_empty and not result.empty
+            )
 
             expected = expected_expr.to_pandas()
 
             assert len(expected.columns) == len(result.columns)
-            assert all(r in e.lower() for r, e in zip(result.columns, expected.columns))
+            assert all(
+                r.lower() in e.lower() for r, e in zip(result.columns, expected.columns)
+            )
 
             expected.columns = result.columns
 
             expected = PandasData.convert_table(expected, result_expr.schema())
-            assert (result_is_empty and expected.empty) or not expected.empty
+
+            assert (result_is_empty and expected.empty) or (
+                not result_is_empty and not expected.empty
+            )
 
             assert len(expected) == len(result)
             assert result.columns.tolist() == expected.columns.tolist()