Add PuffinWriter for writing deletion vectors

pyiceberg/table/puffin.py

@@ -14,7 +14,10 @@
 # KIND, either express or implied.  See the License for the
 # specific language governing permissions and limitations
 # under the License.
+import io
 import math
+import zlib
+from collections.abc import Iterable
 from typing import TYPE_CHECKING, Literal
 
 from pydantic import Field
@@ -27,6 +30,7 @@
 
 # Short for: Puffin Fratercula arctica, version 1
 MAGIC_BYTES = b"PFA1"
+DELETION_VECTOR_MAGIC = b"\xd1\xd3\x39\x64"
 EMPTY_BITMAP = FrozenBitMap()
 MAX_JAVA_SIGNED = int(math.pow(2, 31)) - 1
 PROPERTY_REFERENCED_DATA_FILE = "referenced-data-file"
@@ -62,6 +66,35 @@ def _deserialize_bitmap(pl: bytes) -> list[BitMap]:
     return bitmaps
 
 
+def _serialize_bitmaps(bitmaps: dict[int, BitMap]) -> bytes:
+    """
+    Serialize a dictionary of bitmaps into a byte array.
+
+    The format is:
+    - 8 bytes: number of bitmaps (little-endian)
+    - For each bitmap:
+        - 4 bytes: key (little-endian)
+        - n bytes: serialized bitmap
+    """
+    with io.BytesIO() as out:
+        sorted_keys = sorted(bitmaps.keys())
+
+        # number of bitmaps
+        out.write(len(sorted_keys).to_bytes(8, "little"))
+
+        for key in sorted_keys:
+            if key < 0:
+                raise ValueError(f"Invalid unsigned key: {key}")
+            if key > MAX_JAVA_SIGNED:
+                raise ValueError(f"Key {key} is too large, max {MAX_JAVA_SIGNED} to maintain compatibility with Java impl")
+
+            # key
+            out.write(key.to_bytes(4, "little"))
+            # bitmap
+            out.write(bitmaps[key].serialize())
+        return out.getvalue()
+
+
 class PuffinBlobMetadata(IcebergBaseModel):
     type: Literal["deletion-vector-v1"] = Field()
     fields: list[int] = Field()
@@ -114,3 +147,105 @@ def __init__(self, puffin: bytes) -> None:
 
     def to_vector(self) -> dict[str, "pa.ChunkedArray"]:
         return {path: _bitmaps_to_chunked_array(bitmaps) for path, bitmaps in self._deletion_vectors.items()}
+
+
+class PuffinWriter:
+    _blobs: list[PuffinBlobMetadata]
+    _blob_payloads: list[bytes]
+    _created_by: str | None
+
+    def __init__(self, created_by: str | None = None) -> None:
+        self._blobs = []
+        self._blob_payloads = []
+        self._created_by = created_by
+
+    def set_blob(
+        self,
+        positions: Iterable[int],
+        referenced_data_file: str,
+    ) -> None:
+        # We only support one blob at the moment
+        self._blobs = []
+        self._blob_payloads = []
+
+        # 1. Create bitmaps from positions
+        bitmaps: dict[int, BitMap] = {}
+        for pos in positions:
+            key = pos >> 32
+            low_bits = pos & 0xFFFFFFFF
+            if key not in bitmaps:
+                bitmaps[key] = BitMap()
+            bitmaps[key].add(low_bits)
+
+        # Calculate the cardinality from the bitmaps
+        cardinality = sum(len(bm) for bm in bitmaps.values())
+
+        # 2. Serialize bitmaps for the vector payload
+        vector_payload = _serialize_bitmaps(bitmaps)
+
+        # 3. Construct the full blob payload for deletion-vector-v1
+        with io.BytesIO() as blob_payload_buffer:
+            # Magic bytes for DV
+            blob_payload_buffer.write(DELETION_VECTOR_MAGIC)
+            # The vector itself
+            blob_payload_buffer.write(vector_payload)
+
+            # The content for CRC calculation
+            crc_content = blob_payload_buffer.getvalue()
+            crc32 = zlib.crc32(crc_content)
+
+            # The full blob to be stored in the Puffin file
+            with io.BytesIO() as full_blob_buffer:
+                # Combined length of the vector and magic bytes stored as 4 bytes, big-endian
+                full_blob_buffer.write(len(crc_content).to_bytes(4, "big"))
+                # The content (magic + vector)
+                full_blob_buffer.write(crc_content)
+                # A CRC-32 checksum of the magic bytes and serialized vector as 4 bytes, big-endian
+                full_blob_buffer.write(crc32.to_bytes(4, "big"))
+
+                self._blob_payloads.append(full_blob_buffer.getvalue())
+
+        # 4. Create blob metadata
+        properties = {PROPERTY_REFERENCED_DATA_FILE: referenced_data_file, "cardinality": str(cardinality)}
+
+        self._blobs.append(
+            PuffinBlobMetadata(
+                type="deletion-vector-v1",
+                fields=[2147483645],  # Java INT_MAX - 2, reserved field id for deletion vectors
+                snapshot_id=-1,
+                sequence_number=-1,
+                offset=0,  # TODO: Use DeleteFileIndex data
+                length=0,  # TODO: Use DeleteFileIndex data
+                properties=properties,
+                compression_codec=None,
+            )
+        )
+
+    def finish(self) -> bytes:
+        with io.BytesIO() as out:
+            payload_buffer = io.BytesIO()
+            for blob_payload in self._blob_payloads:
+                payload_buffer.write(blob_payload)
+
+            updated_blobs_metadata: list[PuffinBlobMetadata] = []
+            current_offset = 4  # Start after file magic (4 bytes)
+            for i, blob_payload in enumerate(self._blob_payloads):
+                original_metadata_dict = self._blobs[i].model_dump(by_alias=True, exclude_none=True)
+                original_metadata_dict["offset"] = current_offset
+                original_metadata_dict["length"] = len(blob_payload)
+                updated_blobs_metadata.append(PuffinBlobMetadata(**original_metadata_dict))
+                current_offset += len(blob_payload)
+
+            footer = Footer(blobs=updated_blobs_metadata, properties={"created-by": self._created_by} if self._created_by else {})
+            footer_payload_bytes = footer.model_dump_json(by_alias=True, exclude_none=True).encode("utf-8")
+
+            # Final assembly
+            out.write(MAGIC_BYTES)
+            out.write(payload_buffer.getvalue())
+            out.write(MAGIC_BYTES)
+            out.write(footer_payload_bytes)
+            out.write(len(footer_payload_bytes).to_bytes(4, "little"))
+            out.write((0).to_bytes(4, "little"))  # flags
+            out.write(MAGIC_BYTES)
+
+            return out.getvalue()

tests/integration/test_puffin_spark_interop.py

@@ -0,0 +1,93 @@
+# 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 pytest
+from pyspark.sql import SparkSession
+
+from pyiceberg.catalog.rest import RestCatalog
+from pyiceberg.manifest import ManifestContent
+from pyiceberg.table.puffin import PuffinFile
+
+
+def run_spark_commands(spark: SparkSession, sqls: list[str]) -> None:
+    for sql in sqls:
+        spark.sql(sql)
+
+
+@pytest.mark.integration
+def test_read_spark_written_puffin_dv(spark: SparkSession, session_catalog: RestCatalog) -> None:
+    """Verify pyiceberg can read Puffin DVs written by Spark."""
+    identifier = "default.spark_puffin_format_test"
+
+    run_spark_commands(spark, [f"DROP TABLE IF EXISTS {identifier}"])
+    run_spark_commands(
+        spark,
+        [
+            f"""
+            CREATE TABLE {identifier} (id BIGINT)
+            USING iceberg
+            TBLPROPERTIES (
+                'format-version' = '3',
+                'write.delete.mode' = 'merge-on-read'
+            )
+            """,
+        ],
+    )
+
+    df = spark.range(1, 51)
+    df.coalesce(1).writeTo(identifier).append()
+
+    files_before = spark.sql(f"SELECT * FROM {identifier}.files").collect()
+    assert len(files_before) == 1, f"Expected 1 file, got {len(files_before)}"
+
+    run_spark_commands(spark, [f"DELETE FROM {identifier} WHERE id IN (10, 20, 30, 40)"])
+
+    table = session_catalog.load_table(identifier)
+    current_snapshot = table.current_snapshot()
+    assert current_snapshot is not None
+
+    manifests = current_snapshot.manifests(table.io)
+    delete_manifests = [m for m in manifests if m.content == ManifestContent.DELETES]
+    assert len(delete_manifests) > 0, "Expected delete manifest with DVs"
+
+    delete_manifest = delete_manifests[0]
+    entries = list(delete_manifest.fetch_manifest_entry(table.io))
+    assert len(entries) > 0, "Expected at least one delete file entry"
+
+    delete_entry = entries[0]
+    puffin_path = delete_entry.data_file.file_path
+    assert puffin_path.endswith(".puffin"), f"Expected Puffin file, got: {puffin_path}"
+
+    input_file = table.io.new_input(puffin_path)
+    with input_file.open() as f:
+        puffin_bytes = f.read()
+
+    puffin = PuffinFile(puffin_bytes)
+
+    assert len(puffin.footer.blobs) == 1, "Expected exactly one blob"
+
+    blob = puffin.footer.blobs[0]
+    assert blob.type == "deletion-vector-v1"
+    assert "referenced-data-file" in blob.properties
+    assert blob.properties["cardinality"] == "4"
+
+    dv_dict = puffin.to_vector()
+    assert len(dv_dict) == 1, "Expected one data file's deletions"
+
+    for _data_file_path, chunked_array in dv_dict.items():
+        positions = chunked_array.to_pylist()
+        assert len(positions) == 4, f"Expected 4 deleted positions, got {len(positions)}"
+        assert sorted(positions) == [9, 19, 29, 39], f"Unexpected positions: {positions}"

tests/table/test_puffin.py

@@ -14,12 +14,20 @@
 # KIND, either express or implied.  See the License for the
 # specific language governing permissions and limitations
 # under the License.
+import zlib
 from os import path
 
 import pytest
 from pyroaring import BitMap
 
-from pyiceberg.table.puffin import _deserialize_bitmap
+from pyiceberg.table.puffin import (
+    DELETION_VECTOR_MAGIC,
+    MAGIC_BYTES,
+    PROPERTY_REFERENCED_DATA_FILE,
+    PuffinFile,
+    PuffinWriter,
+    _deserialize_bitmap,
+)
 
 
 def _open_file(file: str) -> bytes:
@@ -71,3 +79,97 @@ def test_map_high_vals() -> None:
 
     with pytest.raises(ValueError, match="Key 4022190063 is too large, max 2147483647 to maintain compatibility with Java impl"):
         _ = _deserialize_bitmap(puffin)
+
+
+def test_puffin_round_trip() -> None:
+    # Define some deletion positions for a file
+    deletions = [5, (1 << 32) + 1, 5]  # Test with a high-bit position and duplicate
+
+    file_path = "path/to/data.parquet"
+
+    # Write the Puffin file
+    writer = PuffinWriter(created_by="my-test-app")
+    writer.set_blob(positions=deletions, referenced_data_file=file_path)
+    puffin_bytes = writer.finish()
+
+    # Read the Puffin file back
+    reader = PuffinFile(puffin_bytes)
+
+    # Assert footer metadata
+    assert reader.footer.properties["created-by"] == "my-test-app"
+    assert len(reader.footer.blobs) == 1
+
+    blob_meta = reader.footer.blobs[0]
+    assert blob_meta.properties[PROPERTY_REFERENCED_DATA_FILE] == file_path
+    assert blob_meta.properties["cardinality"] == str(len(set(deletions)))
+
+    # Assert the content of deletion vectors
+    read_vectors = reader.to_vector()
+
+    assert file_path in read_vectors
+    assert read_vectors[file_path].to_pylist() == sorted(set(deletions))
+
+
+def test_write_and_read_puffin_file() -> None:
+    writer = PuffinWriter()
+    writer.set_blob(positions=[1, 2, 3], referenced_data_file="file1.parquet")
+    writer.set_blob(positions=[4, 5, 6], referenced_data_file="file2.parquet")
+    puffin_bytes = writer.finish()
+
+    reader = PuffinFile(puffin_bytes)
+
+    assert len(reader.footer.blobs) == 1
+    blob = reader.footer.blobs[0]
+
+    assert blob.properties["referenced-data-file"] == "file2.parquet"
+    assert blob.properties["cardinality"] == "3"
+    assert blob.type == "deletion-vector-v1"
+    # Reserved field id of the row position column (Java MetadataColumns.ROW_POSITION, INT_MAX - 2);
+    # required for Java/Spark interoperability.
+    assert blob.fields == [2147483645]
+    assert blob.snapshot_id == -1
+    assert blob.sequence_number == -1
+    assert blob.compression_codec is None
+
+    vectors = reader.to_vector()
+    assert len(vectors) == 1
+    assert "file1.parquet" not in vectors
+    assert vectors["file2.parquet"].to_pylist() == [4, 5, 6]
+
+
+def test_deletion_vector_blob_framing_is_spec_compliant() -> None:
+    # PuffinFile reads only the serialized vector, skipping the blob's length prefix,
+    # deletion-vector magic and CRC-32. Assert that framing directly at the byte level so
+    # the bytes an external reader (Java/Spark) relies on stay spec-compliant.
+    positions = [0, 1, 5, (1 << 32) + 7]
+    writer = PuffinWriter()
+    writer.set_blob(positions=positions, referenced_data_file="file.parquet")
+    puffin_bytes = writer.finish()
+
+    # The Puffin file begins with the magic.
+    assert puffin_bytes[:4] == MAGIC_BYTES
+
+    blob = PuffinFile(puffin_bytes).footer.blobs[0]
+    blob_bytes = puffin_bytes[blob.offset : blob.offset + blob.length]
+
+    # Layout: length (4B big-endian) | DV magic (4B) | vector | CRC-32 (4B big-endian),
+    # where the length and CRC-32 both cover the magic bytes plus the vector.
+    length_prefix = int.from_bytes(blob_bytes[0:4], "big")
+    dv_magic = blob_bytes[4:8]
+    vector = blob_bytes[8 : 4 + length_prefix]
+    crc = int.from_bytes(blob_bytes[4 + length_prefix : 8 + length_prefix], "big")
+
+    assert dv_magic == DELETION_VECTOR_MAGIC
+    assert length_prefix == len(dv_magic) + len(vector)
+    assert blob.length == 4 + length_prefix + 4
+    assert crc == zlib.crc32(dv_magic + vector)
+
+
+def test_puffin_file_with_no_blobs() -> None:
+    writer = PuffinWriter()
+    puffin_bytes = writer.finish()
+
+    reader = PuffinFile(puffin_bytes)
+    assert len(reader.footer.blobs) == 0
+    assert len(reader.to_vector()) == 0
+    assert "created-by" not in reader.footer.properties