feat: Improve array(), map(), and struct

fixes ibis-project#8289

This does a lot of changes. It was hard for me to separate them out as I implemented them. But now that it's all hashed out, I can try to split this up into separate commits if you want. But that might be sorta hard in
some cases.

One this is adding support for passing in None to all these constructors.
These use the new `ibis.null(<type>)` API to return `op.Literal(None, <type>)`s

Make these constructors idempotent: you can
pass in existing Expressions into array(), etc.
The type argument for all of these now always has an effect, not just when passing in python literals. So basically it acts like a cast.

A big structural change is that now ops.Array has an optional
attribute "dtype", so if you pass in a 0-length sequence
of values the op still knows what dtype it is.

Several of the backends were always broken here, they just weren't getting caught. I marked them as broken, we can fix them in a followup.

You can test this locally with eg
`pytest -m <backend> -k factory ibis/backends/tests/test_array.py  ibis/backends/tests/test_map.py ibis/backends/tests/test_struct.py`

Also, fix a typing bug: map() can accept ArrayValues, not just ArrayColumns.

Also, fix executing Literal(None) on pandas and polars, 0-length arrays on polars

Also, fixing converting dtypes on clickhouse, Structs should be converted to nonnullable dtypes.

Also, implement ops.StructColumn on pandas and dask

Author: NickCrews

ibis/backends/dask/executor.py

@@ -155,11 +155,17 @@ def mapper(df, cases):
         return cls.partitionwise(mapper, kwargs, name=op.name, dtype=dtype)
 
     @classmethod
-    def visit(cls, op: ops.Array, exprs):
+    def visit(cls, op: ops.Array, exprs, dtype):
         return cls.rowwise(
             lambda row: np.array(row, dtype=object), exprs, name=op.name, dtype=object
         )
 
+    @classmethod
+    def visit(cls, op: ops.StructColumn, names, values):
+        return cls.rowwise(
+            lambda row: dict(zip(names, row)), values, name=op.name, dtype=object
+        )
+
     @classmethod
     def visit(cls, op: ops.ArrayConcat, arg):
         dtype = PandasType.from_ibis(op.dtype)

ibis/backends/exasol/compiler.py

@@ -75,6 +75,7 @@ class ExasolCompiler(SQLGlotCompiler):
         ops.StringSplit,
         ops.StringToDate,
         ops.StringToTimestamp,
+        ops.StructColumn,
         ops.TimeDelta,
         ops.TimestampAdd,
         ops.TimestampBucket,

ibis/backends/pandas/executor.py

@@ -49,12 +49,14 @@ def visit(cls, op: ops.Node, **kwargs):
 
     @classmethod
     def visit(cls, op: ops.Literal, value, dtype):
+        if value is None:
+            return None
         if dtype.is_interval():
-            value = pd.Timedelta(value, dtype.unit.short)
-        elif dtype.is_array():
-            value = np.array(value)
-        elif dtype.is_date():
-            value = pd.Timestamp(value, tz="UTC").tz_localize(None)
+            return pd.Timedelta(value, dtype.unit.short)
+        if dtype.is_array():
+            return np.array(value)
+        if dtype.is_date():
+            return pd.Timestamp(value, tz="UTC").tz_localize(None)
         return value
 
     @classmethod
@@ -220,9 +222,13 @@ def visit(cls, op: ops.FindInSet, needle, values):
         return pd.Series(result, name=op.name)
 
     @classmethod
-    def visit(cls, op: ops.Array, exprs):
+    def visit(cls, op: ops.Array, exprs, dtype):
         return cls.rowwise(lambda row: np.array(row, dtype=object), exprs)
 
+    @classmethod
+    def visit(cls, op: ops.StructColumn, names, values):
+        return cls.rowwise(lambda row: dict(zip(names, row)), values)
+
     @classmethod
     def visit(cls, op: ops.ArrayConcat, arg):
         return cls.rowwise(lambda row: np.concatenate(row.values), arg)

ibis/backends/polars/compiler.py

@@ -86,10 +86,14 @@ def _make_duration(value, dtype):
 def literal(op, **_):
     value = op.value
     dtype = op.dtype
+    if dtype.is_interval():
+        return _make_duration(value, dtype)
 
-    if dtype.is_array():
+    typ = PolarsType.from_ibis(dtype)
+    if value is None:
+        return pl.lit(None, dtype=typ)
+    elif dtype.is_array():
         value = pl.Series("", value)
-        typ = PolarsType.from_ibis(dtype)
         val = pl.lit(value, dtype=typ)
         return val.implode()
     elif dtype.is_struct():
@@ -98,14 +102,11 @@ def literal(op, **_):
             for k, v in value.items()
         ]
         return pl.struct(values)
-    elif dtype.is_interval():
-        return _make_duration(value, dtype)
     elif dtype.is_null():
         return pl.lit(value)
     elif dtype.is_binary():
         return pl.lit(value)
     else:
-        typ = PolarsType.from_ibis(dtype)
         return pl.lit(op.value, dtype=typ)
 
 
@@ -980,9 +981,11 @@ def array_concat(op, **kw):
 
 
 @translate.register(ops.Array)
-def array_column(op, **kw):
-    cols = [translate(col, **kw) for col in op.exprs]
-    return pl.concat_list(cols)
+def array_literal(op, **kw):
+    if len(op.exprs) > 0:
+        return pl.concat_list([translate(col, **kw) for col in op.exprs])
+    else:
+        return pl.lit([], dtype=PolarsType.from_ibis(op.dtype))
 
 
 @translate.register(ops.ArrayCollect)

ibis/backends/risingwave/compiler.py

@@ -8,7 +8,7 @@
 import ibis.expr.datatypes as dt
 import ibis.expr.operations as ops
 from ibis.backends.postgres.compiler import PostgresCompiler
-from ibis.backends.sql.compiler import ALL_OPERATIONS
+from ibis.backends.sql.compiler import ALL_OPERATIONS, SQLGlotCompiler
 from ibis.backends.sql.datatypes import RisingWaveType
 from ibis.backends.sql.dialects import RisingWave
 
@@ -51,6 +51,11 @@ def visit_Correlation(self, op, *, left, right, how, where):
             op, left=left, right=right, how=how, where=where
         )
 
+    # def visit_StructColumn(self, op, *, names, values):
+    # The parent Postgres compiler uses the ROW() function,
+    # but the grandparent SQLGlot compiler uses the correct syntax
+    # return SQLGlotCompiler.visit_StructColumn(self, op, names=names, values=values)
+
     def visit_TimestampTruncate(self, op, *, arg, unit):
         unit_mapping = {
             "Y": "year",

ibis/backends/sql/compiler.py

@@ -970,8 +970,11 @@ def visit_InSubquery(self, op, *, rel, needle):
             query = sg.select(STAR).from_(query)
         return needle.isin(query=query)
 
-    def visit_Array(self, op, *, exprs):
-        return self.f.array(*exprs)
+    def visit_Array(self, op, *, exprs, dtype):
+        result = self.f.array(*exprs)
+        if len(exprs) == 0:
+            return self.cast(result, dtype)
+        return result
 
     def visit_StructColumn(self, op, *, names, values):
         return sge.Struct.from_arg_list(

ibis/backends/sql/datatypes.py

@@ -1007,8 +1007,10 @@ class ClickHouseType(SqlglotType):
     def from_ibis(cls, dtype: dt.DataType) -> sge.DataType:
         """Convert a sqlglot type to an ibis type."""
         typ = super().from_ibis(dtype)
-        if dtype.nullable and not (dtype.is_map() or dtype.is_array()):
-            # map cannot be nullable in clickhouse
+        # nested types cannot be nullable in clickhouse
+        if dtype.nullable and not (
+            dtype.is_map() or dtype.is_array() or dtype.is_struct()
+        ):
             return sge.DataType(this=typecode.NULLABLE, expressions=[typ])
         else:
             return typ

ibis/backends/sqlite/compiler.py

@@ -60,6 +60,7 @@ class SQLiteCompiler(SQLGlotCompiler):
         ops.TimestampDiff,
         ops.StringToDate,
         ops.StringToTimestamp,
+        ops.StructColumn,
         ops.TimeDelta,
         ops.DateDelta,
         ops.TimestampDelta,

ibis/backends/tests/test_array.py

@@ -30,6 +30,7 @@
     PySparkAnalysisException,
     TrinoUserError,
 )
+from ibis.common.annotations import ValidationError
 from ibis.common.collections import frozendict
 
 pytestmark = [
@@ -66,11 +67,64 @@
     pytest.mark.notimpl(["druid", "oracle"], raises=Exception),
 ]
 
+mark_notyet_datafusion = pytest.mark.notyet(
+    "datafusion",
+    raises=Exception,
+    reason="datafusion can't handle array casts yet. https://github.com/apache/datafusion/issues/10464",
+)
+
 # NB: We don't check whether results are numpy arrays or lists because this
 # varies across backends. At some point we should unify the result type to be
 # list.
 
 
+def test_array_factory(con):
+    a = ibis.array([1, 2, 3])
+    assert con.execute(a) == [1, 2, 3]
+
+    a2 = ibis.array(a)
+    assert con.execute(a2) == [1, 2, 3]
+
+
+@mark_notyet_datafusion
+def test_array_factory_typed(con):
+    typed = ibis.array([1, 2, 3], type="array<string>")
+    assert con.execute(typed) == ["1", "2", "3"]
+
+    typed2 = ibis.array(ibis.array([1, 2, 3]), type="array<string>")
+    assert con.execute(typed2) == ["1", "2", "3"]
+
+
+@mark_notyet_datafusion
+@pytest.mark.notimpl("flink", raises=Py4JJavaError)
+@pytest.mark.notimpl(["pandas", "dask"], raises=ValueError)
+def test_array_factory_empty(con):
+    with pytest.raises(ValidationError):
+        ibis.array([])
+
+    empty_typed = ibis.array([], type="array<string>")
+    assert empty_typed.type() == dt.Array(value_type=dt.string)
+    assert con.execute(empty_typed) == []
+
+
+@mark_notyet_datafusion
+@pytest.mark.notyet(
+    "clickhouse", raises=ClickHouseDatabaseError, reason="nested types can't be NULL"
+)
+def test_array_factory_null(con):
+    with pytest.raises(ValidationError):
+        ibis.array(None)
+    with pytest.raises(ValidationError):
+        ibis.array(None, type="int64")
+    none_typed = ibis.array(None, type="array<string>")
+    assert none_typed.type() == dt.Array(value_type=dt.string)
+    assert con.execute(none_typed) is None
+    # Execute a real value here, so the backends that don't support arrays
+    # actually xfail as we expect them to.
+    # Otherwise would have to @mark.xfail every test in this file besides this one.
+    assert con.execute(ibis.array([1, 2])) == [1, 2]
+
+
 def test_array_column(backend, alltypes, df):
     expr = ibis.array(
         [alltypes["double_col"], alltypes["double_col"], 5.0, ibis.literal(6.0)]
@@ -913,11 +967,6 @@ def test_zip_null(con, fn):
 
 
 @builtin_array
-@pytest.mark.notyet(
-    ["clickhouse"],
-    raises=ClickHouseDatabaseError,
-    reason="https://github.com/ClickHouse/ClickHouse/issues/41112",
-)
 @pytest.mark.notimpl(["postgres"], raises=PsycoPg2SyntaxError)
 @pytest.mark.notimpl(["risingwave"], raises=PsycoPg2ProgrammingError)
 @pytest.mark.notimpl(["datafusion"], raises=com.OperationNotDefinedError)

ibis/backends/tests/test_generic.py

@@ -1230,9 +1230,7 @@ def query(t, group_cols):
     snapshot.assert_match(str(ibis.to_sql(t3, dialect=con.name)), "out.sql")
 
 
-@pytest.mark.notimpl(
-    ["dask", "pandas", "oracle", "exasol"], raises=com.OperationNotDefinedError
-)
+@pytest.mark.notimpl(["oracle", "exasol"], raises=com.OperationNotDefinedError)
 @pytest.mark.notimpl(["druid"], raises=AssertionError)
 @pytest.mark.notyet(
     ["datafusion", "impala", "mssql", "mysql", "sqlite"],