def execute(
self,
expr: ir.Expr,
timecontext: Mapping | None = None,
params: Mapping[ir.Scalar, Any] | None = None,
limit: str = 'default',
**kwargs: Any,
) -> Any:
"""Execute an expression."""
if isinstance(expr, types.Table):
return self.compile(expr, timecontext, params, **kwargs).toPandas()
elif isinstance(expr, types.Column):
# expression must be named for the projection
if not expr.has_name():
expr = expr.name("tmp")
return self.compile(
expr.to_projection(), timecontext, params, **kwargs
).toPandas()[expr.get_name()]
elif isinstance(expr, types.Scalar):
compiled = self.compile(expr, timecontext, params, **kwargs)
if isinstance(compiled, Column):
# attach result column to a fake DataFrame and
# select the result
compiled = self._session.range(0, 1).select(compiled)
return compiled.toPandas().iloc[0, 0]
else:
raise com.IbisError(
f"Cannot execute expression of type: {type(expr)}"
)
def execute(
self,
expr: ir.Expr,
params: Mapping[ir.Expr, object] = None,
limit: str = 'default',
**kwargs: Any,
):
if isinstance(expr, ir.TableExpr):
frame = self.compile(expr, params, **kwargs)
table = _to_pyarrow_table(frame)
return table.to_pandas()
elif isinstance(expr, ir.ColumnExpr):
# expression must be named for the projection
expr = expr.name('tmp').to_projection()
frame = self.compile(expr, params, **kwargs)
table = _to_pyarrow_table(frame)
return table['tmp'].to_pandas()
elif isinstance(expr, ir.ScalarExpr):
if expr.op().root_tables():
# there are associated datafusion tables so convert the expr
# to a selection which we can directly convert to a datafusion
# plan
expr = expr.name('tmp').to_projection()
frame = self.compile(expr, params, **kwargs)
else:
# doesn't have any tables associated so create a plan from a
# dummy datafusion table
compiled = self.compile(expr, params, **kwargs)
frame = self._context.empty_table().select(compiled)
table = _to_pyarrow_table(frame)
return table[0][0].as_py()
else:
raise com.IbisError(
f"Cannot execute expression of type: {type(expr)}"
)
def coerce_to_output(
result: Any,
expr: ir.Expr,
index: Optional[pd.Index] = None) -> Union[pd.Series, pd.DataFrame]:
""" Cast the result to either a Series or DataFrame.
This method casts result of an execution to a Series or DataFrame,
depending on the type of the expression and shape of the result.
Parameters
----------
result: Any
The result to cast
expr: ibis.expr.types.Expr
The expression associated with the result
index: pd.Index
Optional. If passed, scalar results will be broadcasted according
to the index.
Returns
-------
result: A Series or DataFrame
"""
result_name = getattr(expr, '_name', None)
if isinstance(expr, (ir.DestructColumn, ir.StructColumn)):
return ibis.util.coerce_to_dataframe(result, expr.type().names)
elif isinstance(expr, (ir.DestructScalar, ir.StructScalar)):
# Here there are two cases, if this is groupby aggregate,
# then the result e a Series of tuple/list, or
# if this is non grouped aggregate, then the result
return ibis.util.coerce_to_dataframe(result, expr.type().names)
elif isinstance(result, pd.Series):
return result.rename(result_name)
elif isinstance(result, np.ndarray):
return pd.Series(result, name=result_name)
elif isinstance(expr.op(), ops.Reduction):
# We either wrap a scalar into a single element Series
# or broadcast the scalar to a multi element Series
if index is None:
return pd.Series(result, name=result_name)
else:
return pd.Series(
np.repeat(result, len(index)),
index=index,
name=result_name,
)
else:
raise ValueError(f"Cannot coerce_to_output. Result: {result}")
def maybe_wrap_scalar(result: Any, expr: ir.Expr) -> Any:
"""
A partial implementation of `coerce_to_output` in the pandas backend.
Currently only wraps scalars, but will change when udfs are added to the
dask backend.
"""
result_name = expr.get_name()
if isinstance(result, dd.core.Scalar) and isinstance(
expr.op(), ops.Reduction):
# TODO - computation
return dd.from_pandas(pd.Series(result.compute(), name=result_name),
npartitions=1)
else:
return result.rename(result_name)
def resolve(self, expr: ir.Expr, scope: Scope) -> List[ir.ColumnExpr]:
schema = expr.schema()
columns = schema.names
name_locs = schema._name_locs
start = name_locs[self.start]
stop = name_locs.get(self.stop, len(expr.columns) - 1) + 1 # inclusive
return [expr[columns[i]] for i in range(start, stop)]
def predicate(counts: Counter,
expr: ir.Expr) -> tuple[Sequence[ir.Table] | bool, None]:
op = expr.op()
if isinstance(op, ops.Join):
return [op.left, op.right], None
elif isinstance(op, ops.PhysicalTable):
return lin.halt, None
elif isinstance(op, ops.SelfReference):
return lin.proceed, None
elif isinstance(op, (ops.Selection, ops.Aggregation)):
counts[op] += 1
return [op.table], None
elif isinstance(op, ops.TableNode):
counts[op] += 1
return lin.proceed, None
elif isinstance(op, ops.TableColumn):
return op.table.op() not in counts, None
else:
return lin.proceed, None
def to_op_dag(expr: ir.Expr) -> Graph:
"""Convert `expr` into a directed acyclic graph.
Parameters
----------
expr
An ibis expression
Returns
-------
Graph
A directed acyclic graph of ibis operations
"""
stack = [expr.op()]
dag = {}
while stack:
if (node := stack.pop()) not in dag:
dag[node] = children = node._flat_ops
stack.extend(children)
def coerce_to_output(
result: Any,
expr: ir.Expr,
index: Optional[pd.Index] = None) -> Union[dd.Series, dd.DataFrame]:
"""Cast the result to either a Series of DataFrame, renaming as needed.
Reimplementation of `coerce_to_output` in the pandas backend, but
creates dask objects and adds special handling for dd.Scalars.
Parameters
----------
result: Any
The result to cast
expr: ibis.expr.types.Expr
The expression associated with the result
index: pd.Index
Optional. If passed, scalar results will be broadcasted according
to the index.
Returns
-------
result: A `dd.Series` or `dd.DataFrame`
Raises
------
ValueError
If unable to coerce result
Examples
--------
For dataframe outputs, see ``_coerce_to_dataframe``. Examples below use
pandas objects for legibility, but functionality is the same on dask
objects.
>>> coerce_to_output(pd.Series(1), expr)
0 1
Name: result, dtype: int64
>>> coerce_to_output(1, expr)
0 1
Name: result, dtype: int64
>>> coerce_to_output(1, expr, [1,2,3])
1 1
2 1
3 1
Name: result, dtype: int64
>>> coerce_to_output([1,2,3], expr)
0 [1, 2, 3]
Name: result, dtype: object
"""
result_name = expr.get_name()
dataframe_exprs = (
ir.DestructColumn,
ir.StructColumn,
ir.DestructScalar,
ir.StructScalar,
)
if isinstance(expr, dataframe_exprs):
return _coerce_to_dataframe(result,
expr.type().names,
expr.type().types)
elif isinstance(result, (pd.Series, dd.Series)):
# Series from https://github.com/ibis-project/ibis/issues/2711
return result.rename(result_name)
elif isinstance(expr.op(), ops.Reduction):
if isinstance(result, dd.core.Scalar):
# wrap the scalar in a series
out_dtype = _pandas_dtype_from_dd_scalar(result)
out_len = 1 if index is None else len(index)
meta = make_meta_series(dtype=out_dtype, name=result_name)
# Specify `divisions` so that the created Dask object has
# known divisions (to be concatenatable with Dask objects
# created using `dd.from_pandas`)
series = dd.from_delayed(
_wrap_dd_scalar(result, result_name, out_len),
meta=meta,
divisions=(0, out_len - 1),
)
return series
else:
return dd.from_pandas(pd_util.coerce_to_output(
result, expr, index),
npartitions=1)
else:
raise ValueError(f"Cannot coerce_to_output. Result: {result}")
def _is_row_order_preserving(expr: ir.Expr):
if isinstance(expr.op(), (ops.Reduction, ops.WindowOp)):
return (lin.halt, False)
else:
return (lin.proceed, True)
def coerce_to_output(
result: Any,
expr: ir.Expr,
index: Optional[pd.Index] = None) -> Union[pd.Series, pd.DataFrame]:
"""Cast the result to either a Series or DataFrame.
This method casts result of an execution to a Series or DataFrame,
depending on the type of the expression and shape of the result.
Parameters
----------
result: Any
The result to cast
expr: ibis.expr.types.Expr
The expression associated with the result
index: pd.Index
Optional. If passed, scalar results will be broadcasted according
to the index.
Returns
-------
result: A Series or DataFrame
Examples
--------
For dataframe outputs, see ``ibis.util.coerce_to_dataframe``.
>>> coerce_to_output(pd.Series(1), expr)
0 1
Name: result, dtype: int64
>>> coerce_to_output(1, expr)
0 1
Name: result, dtype: int64
>>> coerce_to_output(1, expr, [1,2,3])
1 1
2 1
3 1
Name: result, dtype: int64
>>> coerce_to_output([1,2,3], expr)
0 [1, 2, 3]
Name: result, dtype: object
"""
result_name = expr._safe_name
if isinstance(expr, (ir.DestructColumn, ir.StructColumn)):
return _coerce_to_dataframe(result, expr.type())
elif isinstance(expr, (ir.DestructScalar, ir.StructScalar)):
# Here there are two cases, if this is groupby aggregate,
# then the result e a Series of tuple/list, or
# if this is non grouped aggregate, then the result
return _coerce_to_dataframe(result, expr.type())
elif isinstance(result, pd.Series):
return result.rename(result_name)
elif isinstance(expr, ir.ScalarExpr):
if index is None:
# Wrap `result` into a single-element Series.
return pd.Series([result], name=result_name)
else:
# Broadcast `result` to a multi-element Series according to the
# given `index`.
return pd.Series(
np.repeat(result, len(index)),
index=index,
name=result_name,
)
elif isinstance(result, np.ndarray):
return pd.Series(result, name=result_name)
else:
raise ValueError(f"Cannot coerce_to_output. Result: {result}")
def main_execute(expr: ir.Expr,
scope: Optional[Mapping] = None,
aggcontext: Optional[agg_ctx.AggregationContext] = None,
clients: Sequence[ibis.client.Client] = (),
params: Optional[Mapping] = None,
**kwargs: Any):
"""Execute an ibis expression against the pandas backend.
Parameters
----------
expr
scope
aggcontext
clients
params
"""
toposorted, dependencies = toposort(expr)
params = toolz.keymap(get_node, params if params is not None else {})
# Add to scope the objects that have no dependencies and are not ibis
# nodes. We have to filter out nodes for cases--such as zero argument
# UDFs--that do not have any dependencies yet still need to be evaluated.
full_scope = toolz.merge(
scope if scope is not None else {},
{
key: key
for key, parents in dependencies.items()
if not parents and not isinstance(key, ops.Node)
},
params,
)
if not clients:
clients = list(find_backends(expr))
if aggcontext is None:
aggcontext = agg_ctx.Summarize()
# give backends a chance to inject scope if needed
execute_first_scope = execute_first(expr.op(),
*clients,
scope=full_scope,
aggcontext=aggcontext,
**kwargs)
full_scope = toolz.merge(full_scope, execute_first_scope)
nodes = [node for node in toposorted if node not in full_scope]
# compute the nodes that are not currently in scope
for node in nodes:
# allow clients to pre compute nodes as they like
pre_executed_scope = pre_execute(node,
*clients,
scope=full_scope,
aggcontext=aggcontext,
**kwargs)
# merge the existing scope with whatever was returned from pre_execute
execute_scope = toolz.merge(full_scope, pre_executed_scope)
# if after pre_execute our node is in scope, then there's nothing to do
# in this iteration
if node in execute_scope:
full_scope = execute_scope
else:
# If we're evaluating a literal then we can be a bit quicker about
# evaluating the dispatch graph
if isinstance(node, ops.Literal):
executor = execute_literal
else:
executor = execute_node
# Gather the inputs we've already computed that the current node
# depends on
execute_args = [
full_scope[get_node(arg)] for arg in dependencies[node]
]
# execute the node with its inputs
execute_node_result = executor(
node,
*execute_args,
aggcontext=aggcontext,
scope=execute_scope,
clients=clients,
**kwargs,
)
# last change to perform any additional computation on the result
# before it gets added to scope for the next node
full_scope[node] = post_execute(
node,
execute_node_result,
clients=clients,
aggcontext=aggcontext,
scope=full_scope,
)
# the last node in the toposorted graph is the root and maps to the desired
# result in scope
last_node = toposorted[-1]
result = full_scope[last_node]
return result
def _fmt_value_expr(expr: ir.Expr, *, aliases: Aliases) -> str:
"""Format a value expression.
Forwards the call on to the specific operation dispatch rule.
"""
return fmt_value(expr.op(), aliases=aliases)
def needs_parens(expr: ir.Expr):
op = expr.op()
if isinstance(op, ops.Alias):
op = op.arg.op()
return isinstance(op, _NEEDS_PARENS_OPS)