Compare upsert rows with PyArrow instead of Python

get_rows_to_update compared the matched rows one cell at a time, one PyArrow call
per row and per column, so an upsert got slower with every column of the table.
It compares one column at a time now: 20k matched rows over 200 columns takes
about 0.15s instead of about 20s.

PyArrow cannot compare struct columns, so a struct is compared field by field,
recursing into the structs its fields hold. Lists and maps are still compared in
Python, a slice at a time so the objects of a whole column are never held at once.

When PyArrow refuses to compare two columns because their types differ, the source
is cast to the type of the target. The full table cast this removes used to do
that, and without it a naive timestamp and a zoned one holding the same instant
are reported as different on every run.

Closes #3860

Author: GaspardMerten

pyiceberg/table/upsert_util.py

@@ -53,17 +53,86 @@ def has_duplicate_rows(df: pyarrow_table, join_cols: list[str]) -> bool:
     return len(df.select(join_cols).group_by(join_cols).aggregate([([], "count_all")]).filter(pc.field("count_all") > 1)) > 0
 
 
+# How many values of a column are turned into Python objects at a time, when PyArrow cannot
+# compare them itself
+_PYTHON_COMPARISON_SLICE = 10_000
+
+
+def _get_changed_struct_mask(source_column: pa.ChunkedArray, target_column: pa.ChunkedArray) -> pa.ChunkedArray:
+    """Compare two struct columns field by field, which PyArrow can do even though it cannot compare the structs."""
+    # `struct_field` carries the null of the struct into its fields, so the fields of two null
+    # structs compare equal and only the struct itself decides for those rows
+    changed = pc.not_equal(pc.is_null(source_column), pc.is_null(target_column))
+
+    for index in range(source_column.type.num_fields):
+        changed = pc.or_(changed, _get_changed_mask(pc.struct_field(source_column, index), pc.struct_field(target_column, index)))
+
+    return changed
+
+
+def _get_changed_mask(source_column: pa.ChunkedArray, target_column: pa.ChunkedArray) -> pa.ChunkedArray:
+    """Return a boolean mask that flags the positions where the two columns differ, treating two nulls as equal."""
+    try:
+        differs = pc.not_equal(source_column, target_column)
+    except (pa.ArrowNotImplementedError, pa.ArrowInvalid):
+        # PyArrow cannot compare columns with complex types
+        # See: https://github.com/apache/arrow/issues/35785
+        if pa.types.is_struct(source_column.type) and source_column.type == target_column.type:
+            return _get_changed_struct_mask(source_column, target_column)
+
+        # Two columns PyArrow refuses to compare may still hold the same values in another type:
+        # a naive timestamp against a zoned one, or the string of a dataframe against the
+        # large_string a scan reads. Comparing those in Python would call every row changed, on
+        # every run, and would leave a struct out of the comparison by field above. The types have
+        # to differ for this to make progress, the cast leaves them equal and the next round
+        # settles it one way or the other
+        if source_column.type != target_column.type:
+            try:
+                return _get_changed_mask(source_column.cast(target_column.type), target_column)
+            except pa.ArrowException:
+                # Whatever PyArrow makes of the cast, the comparison in Python below still holds
+                pass
+
+        # A list or a map is left to be compared in Python, value by value. A slice at a time,
+        # so that the objects of a whole column are never held at once
+        return pa.chunked_array(
+            [
+                [
+                    source_val != target_val
+                    for source_val, target_val in zip(
+                        source_column.slice(offset, _PYTHON_COMPARISON_SLICE).to_pylist(),
+                        target_column.slice(offset, _PYTHON_COMPARISON_SLICE).to_pylist(),
+                        strict=True,
+                    )
+                ]
+                for offset in range(0, len(source_column), _PYTHON_COMPARISON_SLICE)
+            ]
+            or [[]],
+            type=pa.bool_(),
+        )
+
+    # `not_equal` is null as soon as either side is null, and a null differs from a value
+    # but not from another null
+    return pc.fill_null(differs, pc.not_equal(pc.is_null(source_column), pc.is_null(target_column)))
+
+
 def get_rows_to_update(source_table: pa.Table, target_table: pa.Table, join_cols: list[str]) -> pa.Table:
     """
     Return a table with rows that need to be updated in the target table based on the join columns.
 
     The table is joined on the identifier columns, and then checked if there are any updated rows.
     Those are selected and everything is renamed correctly.
     """
-    all_columns = set(source_table.column_names)
-    join_cols_set = set(join_cols)
+    if set(source_table.column_names) != set(target_table.column_names):
+        raise ValueError(
+            f"Source table's field names are not matching the target's field names: "
+            f"{source_table.column_names}, {target_table.column_names}"
+        )
 
-    non_key_cols = list(all_columns - join_cols_set)
+    # Kept in the order of the source rather than taken from a set difference, whose order
+    # varies from one process to the next
+    join_cols_set = set(join_cols)
+    non_key_cols = [col for col in source_table.column_names if col not in join_cols_set]
 
     if has_duplicate_rows(target_table, join_cols):
         raise ValueError("Target table has duplicate rows, aborting upsert")
@@ -72,10 +141,6 @@ def get_rows_to_update(source_table: pa.Table, target_table: pa.Table, join_cols
         # When the target table is empty, there is nothing to update :)
         return source_table.schema.empty_table()
 
-    # We need to compare non_key_cols in Python as PyArrow
-    # 1. Cannot do a join when non-join columns have complex types
-    # 2. Cannot compare columns with complex types
-    # See: https://github.com/apache/arrow/issues/35785
     SOURCE_INDEX_COLUMN_NAME = "__source_index"
     TARGET_INDEX_COLUMN_NAME = "__target_index"
 
@@ -86,39 +151,38 @@ def get_rows_to_update(source_table: pa.Table, target_table: pa.Table, join_cols
         ) from None
 
     # Step 1: Prepare source index with join keys and a marker index
-    # Cast to target table schema, so we can do the join
+    # Only the join columns are cast, so the width of the table does not weigh on the join
     # See: https://github.com/apache/arrow/issues/37542
     source_index = (
-        source_table.cast(target_table.schema)
-        .select(join_cols_set)
+        source_table.select(join_cols)
+        .cast(target_table.select(join_cols).schema)
         .append_column(SOURCE_INDEX_COLUMN_NAME, pa.array(range(len(source_table))))
     )
 
     # Step 2: Prepare target index with join keys and a marker
-    target_index = target_table.select(join_cols_set).append_column(TARGET_INDEX_COLUMN_NAME, pa.array(range(len(target_table))))
+    target_index = target_table.select(join_cols).append_column(TARGET_INDEX_COLUMN_NAME, pa.array(range(len(target_table))))
 
     # Step 3: Perform an inner join to find which rows from source exist in target
-    matching_indices = source_index.join(target_index, keys=list(join_cols_set), join_type="inner")
-
-    # Step 4: Compare all rows using Python
-    to_update_indices = []
-    for source_idx, target_idx in zip(
-        matching_indices[SOURCE_INDEX_COLUMN_NAME].to_pylist(),
-        matching_indices[TARGET_INDEX_COLUMN_NAME].to_pylist(),
-        strict=True,
-    ):
-        source_row = source_table.slice(source_idx, 1)
-        target_row = target_table.slice(target_idx, 1)
-
-        for key in non_key_cols:
-            source_val = source_row.column(key)[0].as_py()
-            target_val = target_row.column(key)[0].as_py()
-            if source_val != target_val:
-                to_update_indices.append(source_idx)
-                break
-
-    # Step 5: Take rows from source table using the indices and cast to target schema
-    if to_update_indices:
-        return source_table.take(to_update_indices)
-    else:
+    matching_indices = source_index.join(target_index, keys=join_cols, join_type="inner")
+
+    if len(matching_indices) == 0:
         return source_table.schema.empty_table()
+
+    source_indices = matching_indices[SOURCE_INDEX_COLUMN_NAME]
+    target_indices = matching_indices[TARGET_INDEX_COLUMN_NAME]
+
+    # Step 4: Compare the matched rows one column at a time. Comparing them cell by cell instead
+    # would allocate a PyArrow scalar per cell, which does not fit in memory on a wide table.
+    changed = pa.chunked_array([pa.repeat(False, len(matching_indices))])
+    for col in non_key_cols:
+        changed = pc.or_(
+            changed,
+            _get_changed_mask(source_table.column(col).take(source_indices), target_table.column(col).take(target_indices)),
+        )
+        # Once every matched row has changed, the columns that are left cannot add anything, and
+        # asking is far cheaper than taking and comparing them
+        if pc.all(changed).as_py():
+            break
+
+    # Step 5: Take rows from source table using the indices
+    return source_table.take(source_indices.filter(changed))