Be more careful about reference counts in zero-copy handoff, add pyarrow.Array.to_pandas method

Change-Id: Ic66c86f6900ff95463228667305760f44d71185c

python/pyarrow/array.pyx

@@ -22,6 +22,7 @@
 import numpy as np
 
 from pyarrow.includes.libarrow cimport *
+from pyarrow.includes.common cimport PyObject_to_object
 cimport pyarrow.includes.pyarrow as pyarrow
 
 import pyarrow.config
@@ -35,6 +36,8 @@ from pyarrow.scalar import NA
 from pyarrow.schema cimport Schema
 import pyarrow.schema as schema
 
+cimport cpython
+
 
 def total_allocated_bytes():
     cdef MemoryPool* pool = pyarrow.get_memory_pool()
@@ -111,6 +114,24 @@ cdef class Array:
     def slice(self, start, end):
         pass
 
+    def to_pandas(self):
+        """
+        Convert to an array object suitable for use in pandas
+
+        See also
+        --------
+        Column.to_pandas
+        Table.to_pandas
+        RecordBatch.to_pandas
+        """
+        cdef:
+            PyObject* np_arr
+
+        check_status(pyarrow.ConvertArrayToPandas(
+            self.sp_array, <PyObject*> self, &np_arr))
+
+        return PyObject_to_object(np_arr)
+
 
 cdef class NullArray(Array):
     pass

python/pyarrow/includes/common.pxd

@@ -47,3 +47,10 @@ cdef extern from "arrow/api.h" namespace "arrow" nogil:
         c_bool IsKeyError()
         c_bool IsNotImplemented()
         c_bool IsInvalid()
+
+
+cdef inline object PyObject_to_object(PyObject* o):
+    # Cast to "object" increments reference count
+    cdef object result = <object> o
+    cpython.Py_DECREF(result)
+    return result

python/pyarrow/includes/pyarrow.pxd

@@ -34,10 +34,10 @@ cdef extern from "pyarrow/api.h" namespace "pyarrow" nogil:
                                 shared_ptr[CArray]* out)
 
     CStatus ConvertArrayToPandas(const shared_ptr[CArray]& arr,
-                                 object py_ref, PyObject** out)
+                                 PyObject* py_ref, PyObject** out)
 
     CStatus ConvertColumnToPandas(const shared_ptr[CColumn]& arr,
-                                  object py_ref, PyObject** out)
+                                  PyObject* py_ref, PyObject** out)
 
     MemoryPool* get_memory_pool()
 

python/pyarrow/table.pyx

@@ -22,6 +22,7 @@
 from cython.operator cimport dereference as deref
 
 from pyarrow.includes.libarrow cimport *
+from pyarrow.includes.common cimport PyObject_to_object
 cimport pyarrow.includes.pyarrow as pyarrow
 
 import pyarrow.config
@@ -32,6 +33,7 @@ from pyarrow.schema cimport box_data_type, box_schema
 
 from pyarrow.compat import frombytes, tobytes
 
+cimport cpython
 
 cdef class ChunkedArray:
     '''
@@ -100,8 +102,10 @@ cdef class Column:
 
         import pandas as pd
 
-        check_status(pyarrow.ConvertColumnToPandas(self.sp_column, self, &arr))
-        return pd.Series(<object>arr, name=self.name)
+        check_status(pyarrow.ConvertColumnToPandas(self.sp_column,
+                                                   <PyObject*> self, &arr))
+
+        return pd.Series(PyObject_to_object(arr), name=self.name)
 
     cdef _check_nullptr(self):
         if self.column == NULL:
@@ -248,9 +252,10 @@ cdef class RecordBatch:
         data = []
         for i in range(self.batch.num_columns()):
             arr = self.batch.column(i)
-            check_status(pyarrow.ConvertArrayToPandas(arr, self, &np_arr))
+            check_status(pyarrow.ConvertArrayToPandas(arr, <PyObject*> self,
+                                                      &np_arr))
             names.append(frombytes(self.batch.column_name(i)))
-            data.append(<object> np_arr)
+            data.append(PyObject_to_object(np_arr))
 
         return pd.DataFrame(dict(zip(names, data)), columns=names)
 
@@ -375,9 +380,10 @@ cdef class Table:
         for i in range(self.table.num_columns()):
             col = self.table.column(i)
             column = self.column(i)
-            check_status(pyarrow.ConvertColumnToPandas(col, column, &arr))
+            check_status(pyarrow.ConvertColumnToPandas(
+                col, <PyObject*> column, &arr))
             names.append(frombytes(col.get().name()))
-            data.append(<object> arr)
+            data.append(PyObject_to_object(arr))
 
         return pd.DataFrame(dict(zip(names, data)), columns=names)
 

python/pyarrow/tests/test_array.py

@@ -15,6 +15,8 @@
 # specific language governing permissions and limitations
 # under the License.
 
+import sys
+
 import pyarrow
 import pyarrow.formatting as fmt
 
@@ -71,3 +73,30 @@ def test_long_array_format():
   99
 ]"""
     assert result == expected
+
+
+def test_to_pandas_zero_copy():
+    import gc
+
+    arr = pyarrow.from_pylist(range(10))
+
+    for i in range(10):
+        np_arr = arr.to_pandas()
+        assert sys.getrefcount(np_arr) == 2
+        np_arr = None  # noqa
+
+    assert sys.getrefcount(arr) == 2
+
+    for i in range(10):
+        arr = pyarrow.from_pylist(range(10))
+        np_arr = arr.to_pandas()
+        arr = None
+        gc.collect()
+
+        # Ensure base is still valid
+
+        # Because of py.test's assert inspection magic, if you put getrefcount
+        # on the line being examined, it will be 1 higher than you expect
+        base_refcount = sys.getrefcount(np_arr.base)
+        assert base_refcount == 2
+        np_arr.sum()

python/src/pyarrow/adapters/pandas.cc

@@ -628,7 +628,6 @@ class ArrowDeserializer {
     PyAcquireGIL lock;
 
     // Zero-Copy. We can pass the data pointer directly to NumPy.
-    OwnedRef py_ref(py_ref_);
     npy_intp dims[1] = {col_->length()};
     out_ = reinterpret_cast<PyArrayObject*>(PyArray_SimpleNewFromData(1, dims,
                 type, data));
@@ -645,7 +644,7 @@ class ArrowDeserializer {
       return Status::OK();
     } else {
       // PyArray_SetBaseObject steals our reference to py_ref_
-      py_ref.release();
+      Py_INCREF(py_ref_);
     }
 
     // Arrow data is immutable.