PyArrow: Avoid buffer-overflow by avoid doing a sort (#1555)

Second attempt of #1539

This was already being discussed back here:
#208 (comment)

This PR changes from doing a sort, and then a single pass over the table
to the approach where we determine the unique partition tuples filter on
them individually.

Fixes #1491

Because the sort caused buffers to be joined where it would overflow in
Arrow. I think this is an issue on the Arrow side, and it should
automatically break up into smaller buffers. The `combine_chunks` method
does this correctly.

Now:

```
0.42877754200890195
Run 1 took: 0.2507691659993725
Run 2 took: 0.24833179199777078
Run 3 took: 0.24401691700040828
Run 4 took: 0.2419595829996979
Average runtime of 0.28 seconds
```

Before:

```
Run 0 took: 1.0768639159941813
Run 1 took: 0.8784021250030492
Run 2 took: 0.8486490420036716
Run 3 took: 0.8614017910003895
Run 4 took: 0.8497851670108503
Average runtime of 0.9 seconds
```

So it comes with a nice speedup as well :)

---------

Co-authored-by: Kevin Liu <[email protected]>

Author: Fokko

pyiceberg/io/pyarrow.py

@@ -27,8 +27,10 @@
 
 import concurrent.futures
 import fnmatch
+import functools
 import itertools
 import logging
+import operator
 import os
 import re
 import uuid
@@ -2174,7 +2176,10 @@ def _partition_value(self, partition_field: PartitionField, schema: Schema) -> A
             raise ValueError(
                 f"Cannot infer partition value from parquet metadata as there are more than one partition values for Partition Field: {partition_field.name}. {lower_value=}, {upper_value=}"
             )
-        return lower_value
+
+        source_field = schema.find_field(partition_field.source_id)
+        transform = partition_field.transform.transform(source_field.field_type)
+        return transform(lower_value)
 
     def partition(self, partition_spec: PartitionSpec, schema: Schema) -> Record:
         return Record(**{field.name: self._partition_value(field, schema) for field in partition_spec.fields})
@@ -2558,38 +2563,8 @@ class _TablePartition:
     arrow_table_partition: pa.Table
 
 
-def _get_table_partitions(
-    arrow_table: pa.Table,
-    partition_spec: PartitionSpec,
-    schema: Schema,
-    slice_instructions: list[dict[str, Any]],
-) -> list[_TablePartition]:
-    sorted_slice_instructions = sorted(slice_instructions, key=lambda x: x["offset"])
-
-    partition_fields = partition_spec.fields
-
-    offsets = [inst["offset"] for inst in sorted_slice_instructions]
-    projected_and_filtered = {
-        partition_field.source_id: arrow_table[schema.find_field(name_or_id=partition_field.source_id).name]
-        .take(offsets)
-        .to_pylist()
-        for partition_field in partition_fields
-    }
-
-    table_partitions = []
-    for idx, inst in enumerate(sorted_slice_instructions):
-        partition_slice = arrow_table.slice(**inst)
-        fieldvalues = [
-            PartitionFieldValue(partition_field, projected_and_filtered[partition_field.source_id][idx])
-            for partition_field in partition_fields
-        ]
-        partition_key = PartitionKey(raw_partition_field_values=fieldvalues, partition_spec=partition_spec, schema=schema)
-        table_partitions.append(_TablePartition(partition_key=partition_key, arrow_table_partition=partition_slice))
-    return table_partitions
-
-
 def _determine_partitions(spec: PartitionSpec, schema: Schema, arrow_table: pa.Table) -> List[_TablePartition]:
-    """Based on the iceberg table partition spec, slice the arrow table into partitions with their keys.
+    """Based on the iceberg table partition spec, filter the arrow table into partitions with their keys.
 
     Example:
     Input:
@@ -2598,54 +2573,50 @@ def _determine_partitions(spec: PartitionSpec, schema: Schema, arrow_table: pa.T
      'n_legs': [2, 2, 2, 4, 4, 4, 4, 5, 100],
      'animal': ["Flamingo", "Parrot", "Parrot", "Dog", "Horse", "Horse", "Horse","Brittle stars", "Centipede"]}.
     The algorithm:
-    Firstly we group the rows into partitions by sorting with sort order [('n_legs', 'descending'), ('year', 'descending')]
-    and null_placement of "at_end".
-    This gives the same table as raw input.
-    Then we sort_indices using reverse order of [('n_legs', 'descending'), ('year', 'descending')]
-    and null_placement : "at_start".
-    This gives:
-    [8, 7, 4, 5, 6, 3, 1, 2, 0]
-    Based on this we get partition groups of indices:
-    [{'offset': 8, 'length': 1}, {'offset': 7, 'length': 1}, {'offset': 4, 'length': 3}, {'offset': 3, 'length': 1}, {'offset': 1, 'length': 2}, {'offset': 0, 'length': 1}]
-    We then retrieve the partition keys by offsets.
-    And slice the arrow table by offsets and lengths of each partition.
+    - We determine the set of unique partition keys
+    - Then we produce a set of partitions by filtering on each of the combinations
+    - We combine the chunks to create a copy to avoid GIL congestion on the original table
     """
-    partition_columns: List[Tuple[PartitionField, NestedField]] = [
-        (partition_field, schema.find_field(partition_field.source_id)) for partition_field in spec.fields
-    ]
-    partition_values_table = pa.table(
-        {
-            str(partition.field_id): partition.transform.pyarrow_transform(field.field_type)(arrow_table[field.name])
-            for partition, field in partition_columns
-        }
-    )
+    # Assign unique names to columns where the partition transform has been applied
+    # to avoid conflicts
+    partition_fields = [f"_partition_{field.name}" for field in spec.fields]
+
+    for partition, name in zip(spec.fields, partition_fields):
+        source_field = schema.find_field(partition.source_id)
+        arrow_table = arrow_table.append_column(
+            name, partition.transform.pyarrow_transform(source_field.field_type)(arrow_table[source_field.name])
+        )
+
+    unique_partition_fields = arrow_table.select(partition_fields).group_by(partition_fields).aggregate([])
+
+    table_partitions = []
+    # TODO: As a next step, we could also play around with yielding instead of materializing the full list
+    for unique_partition in unique_partition_fields.to_pylist():
+        partition_key = PartitionKey(
+            field_values=[
+                PartitionFieldValue(field=field, value=unique_partition[name])
+                for field, name in zip(spec.fields, partition_fields)
+            ],
+            partition_spec=spec,
+            schema=schema,
+        )
+        filtered_table = arrow_table.filter(
+            functools.reduce(
+                operator.and_,
+                [
+                    pc.field(partition_field_name) == unique_partition[partition_field_name]
+                    if unique_partition[partition_field_name] is not None
+                    else pc.field(partition_field_name).is_null()
+                    for field, partition_field_name in zip(spec.fields, partition_fields)
+                ],
+            )
+        )
+        filtered_table = filtered_table.drop_columns(partition_fields)
 
-    # Sort by partitions
-    sort_indices = pa.compute.sort_indices(
-        partition_values_table,
-        sort_keys=[(col, "ascending") for col in partition_values_table.column_names],
-        null_placement="at_end",
-    ).to_pylist()
-    arrow_table = arrow_table.take(sort_indices)
-
-    # Get slice_instructions to group by partitions
-    partition_values_table = partition_values_table.take(sort_indices)
-    reversed_indices = pa.compute.sort_indices(
-        partition_values_table,
-        sort_keys=[(col, "descending") for col in partition_values_table.column_names],
-        null_placement="at_start",
-    ).to_pylist()
-    slice_instructions: List[Dict[str, Any]] = []
-    last = len(reversed_indices)
-    reversed_indices_size = len(reversed_indices)
-    ptr = 0
-    while ptr < reversed_indices_size:
-        group_size = last - reversed_indices[ptr]
-        offset = reversed_indices[ptr]
-        slice_instructions.append({"offset": offset, "length": group_size})
-        last = reversed_indices[ptr]
-        ptr = ptr + group_size
-
-    table_partitions: List[_TablePartition] = _get_table_partitions(arrow_table, spec, schema, slice_instructions)
+        # The combine_chunks seems to be counter-intuitive to do, but it actually returns
+        # fresh buffers that don't interfere with each other when it is written out to file
+        table_partitions.append(
+            _TablePartition(partition_key=partition_key, arrow_table_partition=filtered_table.combine_chunks())
+        )
 
     return table_partitions

pyiceberg/partitioning.py

@@ -29,6 +29,7 @@
     Optional,
     Tuple,
     TypeVar,
+    Union,
 )
 from urllib.parse import quote_plus
 
@@ -393,14 +394,14 @@ class PartitionFieldValue:
 
 @dataclass(frozen=True)
 class PartitionKey:
-    raw_partition_field_values: List[PartitionFieldValue]
+    field_values: List[PartitionFieldValue]
     partition_spec: PartitionSpec
     schema: Schema
 
     @cached_property
     def partition(self) -> Record:  # partition key transformed with iceberg internal representation as input
         iceberg_typed_key_values = {}
-        for raw_partition_field_value in self.raw_partition_field_values:
+        for raw_partition_field_value in self.field_values:
             partition_fields = self.partition_spec.source_id_to_fields_map[raw_partition_field_value.field.source_id]
             if len(partition_fields) != 1:
                 raise ValueError(f"Cannot have redundant partitions: {partition_fields}")
@@ -427,25 +428,45 @@ def partition_record_value(partition_field: PartitionField, value: Any, schema:
     the final partition record value.
     """
     iceberg_type = schema.find_field(name_or_id=partition_field.source_id).field_type
-    iceberg_typed_value = _to_partition_representation(iceberg_type, value)
-    transformed_value = partition_field.transform.transform(iceberg_type)(iceberg_typed_value)
-    return transformed_value
+    return _to_partition_representation(iceberg_type, value)
 
 
 @singledispatch
 def _to_partition_representation(type: IcebergType, value: Any) -> Any:
+    """Strip the logical type into the physical type.
+
+    It can be that the value is already transformed into its physical type,
+    in this case it will return the original value. Keep in mind that the
+    bucket transform always will return an int, but an identity transform
+    can return date that still needs to be transformed into an int (days
+    since epoch).
+    """
     return TypeError(f"Unsupported partition field type: {type}")
 
 
 @_to_partition_representation.register(TimestampType)
 @_to_partition_representation.register(TimestamptzType)
-def _(type: IcebergType, value: Optional[datetime]) -> Optional[int]:
-    return datetime_to_micros(value) if value is not None else None
+def _(type: IcebergType, value: Optional[Union[int, datetime]]) -> Optional[int]:
+    if value is None:
+        return None
+    elif isinstance(value, int):
+        return value
+    elif isinstance(value, datetime):
+        return datetime_to_micros(value)
+    else:
+        raise ValueError(f"Unknown type: {value}")
 
 
 @_to_partition_representation.register(DateType)
-def _(type: IcebergType, value: Optional[date]) -> Optional[int]:
-    return date_to_days(value) if value is not None else None
+def _(type: IcebergType, value: Optional[Union[int, date]]) -> Optional[int]:
+    if value is None:
+        return None
+    elif isinstance(value, int):
+        return value
+    elif isinstance(value, date):
+        return date_to_days(value)
+    else:
+        raise ValueError(f"Unknown type: {value}")
 
 
 @_to_partition_representation.register(TimeType)

pyiceberg/table/__init__.py

@@ -453,8 +453,10 @@ def append(self, df: pa.Table, snapshot_properties: Dict[str, str] = EMPTY_DICT)
         with self._append_snapshot_producer(snapshot_properties) as append_files:
             # skip writing data files if the dataframe is empty
             if df.shape[0] > 0:
-                data_files = _dataframe_to_data_files(
-                    table_metadata=self.table_metadata, write_uuid=append_files.commit_uuid, df=df, io=self._table.io
+                data_files = list(
+                    _dataframe_to_data_files(
+                        table_metadata=self.table_metadata, write_uuid=append_files.commit_uuid, df=df, io=self._table.io
+                    )
                 )
                 for data_file in data_files:
                     append_files.append_data_file(data_file)

pyproject.toml

@@ -1220,6 +1220,7 @@ markers = [
   "adls: marks a test as requiring access to adls compliant storage (use with --adls.account-name, --adls.account-key, and --adls.endpoint args)",
   "integration: marks integration tests against Apache Spark",
   "gcs: marks a test as requiring access to gcs compliant storage (use with --gs.token, --gs.project, and --gs.endpoint)",
+  "benchmark: collection of tests to validate read/write performance before and after a change"
 ]
 
 # Turns a warning into an error

tests/benchmark/test_benchmark.py

@@ -0,0 +1,72 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+import statistics
+import timeit
+import urllib
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+import pytest
+
+from pyiceberg.transforms import DayTransform
+
+
+@pytest.fixture(scope="session")
+def taxi_dataset(tmp_path_factory: pytest.TempPathFactory) -> pa.Table:
+    """Reads the Taxi dataset to disk"""
+    taxi_dataset = "https://d37ci6vzurychx.cloudfront.net/trip-data/yellow_tripdata_2022-01.parquet"
+    taxi_dataset_dest = tmp_path_factory.mktemp("taxi_dataset") / "yellow_tripdata_2022-01.parquet"
+    urllib.request.urlretrieve(taxi_dataset, taxi_dataset_dest)
+
+    return pq.read_table(taxi_dataset_dest)
+
+
+@pytest.mark.benchmark
+def test_partitioned_write(tmp_path_factory: pytest.TempPathFactory, taxi_dataset: pa.Table) -> None:
+    """Tests writing to a partitioned table with something that would be close a production-like situation"""
+    from pyiceberg.catalog.sql import SqlCatalog
+
+    warehouse_path = str(tmp_path_factory.mktemp("warehouse"))
+    catalog = SqlCatalog(
+        "default",
+        uri=f"sqlite:///{warehouse_path}/pyiceberg_catalog.db",
+        warehouse=f"file://{warehouse_path}",
+    )
+
+    catalog.create_namespace("default")
+
+    tbl = catalog.create_table("default.taxi_partitioned", schema=taxi_dataset.schema)
+
+    with tbl.update_spec() as spec:
+        spec.add_field("tpep_pickup_datetime", DayTransform())
+
+    # Profiling can sometimes be handy as well
+    # with cProfile.Profile() as pr:
+    #     tbl.append(taxi_dataset)
+    #
+    # pr.print_stats(sort=True)
+
+    runs = []
+    for run in range(5):
+        start_time = timeit.default_timer()
+        tbl.append(taxi_dataset)
+        elapsed = timeit.default_timer() - start_time
+
+        print(f"Run {run} took: {elapsed}")
+        runs.append(elapsed)
+
+    print(f"Average runtime of {round(statistics.mean(runs), 2)} seconds")