deltacat 0.2.9__py3-none-any.whl → 1.0.0__py3-none-any.whl

deltacat/compute/compactor_v2/model/merge_result.py

@@ -6,6 +6,7 @@ import numpy as np
 
 class MergeResult(NamedTuple):
     materialize_results: List[MaterializeResult]
+    input_record_count: np.int64
     deduped_record_count: np.int64
     peak_memory_usage_bytes: np.double
     telemetry_time_in_seconds: np.double

deltacat/compute/compactor_v2/steps/hash_bucket.py

@@ -5,7 +5,6 @@ from contextlib import nullcontext
 from typing import List, Optional, Tuple
 from deltacat.compute.compactor_v2.model.hash_bucket_input import HashBucketInput
 import numpy as np
-import pyarrow as pa
 import ray
 from deltacat import logs
 from deltacat.compute.compactor import (
@@ -14,12 +13,12 @@ from deltacat.compute.compactor import (
 )
 from deltacat.compute.compactor.model.delta_file_envelope import DeltaFileEnvelopeGroups
 from deltacat.compute.compactor_v2.model.hash_bucket_result import HashBucketResult
+from deltacat.compute.compactor_v2.utils.delta import read_delta_file_envelopes
 from deltacat.compute.compactor_v2.utils.primary_key_index import (
     group_hash_bucket_indices,
     group_by_pk_hash_bucket,
 )
 from deltacat.storage import interface as unimplemented_deltacat_storage
-from deltacat.types.media import StorageType
 from deltacat.utils.ray_utils.runtime import (
     get_current_ray_task_id,
     get_current_ray_worker_id,
@@ -39,57 +38,6 @@ if importlib.util.find_spec("memray"):
 logger = logs.configure_deltacat_logger(logging.getLogger(__name__))
 
 
-def _read_delta_file_envelopes(
-    annotated_delta: DeltaAnnotated,
-    read_kwargs_provider: Optional[ReadKwargsProvider],
-    deltacat_storage=unimplemented_deltacat_storage,
-    deltacat_storage_kwargs: Optional[dict] = None,
-) -> Tuple[Optional[List[DeltaFileEnvelope]], int, int]:
-
-    tables = deltacat_storage.download_delta(
-        annotated_delta,
-        max_parallelism=1,
-        file_reader_kwargs_provider=read_kwargs_provider,
-        storage_type=StorageType.LOCAL,
-        **deltacat_storage_kwargs,
-    )
-    annotations = annotated_delta.annotations
-    assert (
-        len(tables) == len(annotations),
-        f"Unexpected Error: Length of downloaded delta manifest tables "
-        f"({len(tables)}) doesn't match the length of delta manifest "
-        f"annotations ({len(annotations)}).",
-    )
-    if not tables:
-        return None, 0, 0
-
-    delta_stream_position = annotations[0].annotation_stream_position
-    delta_type = annotations[0].annotation_delta_type
-
-    for annotation in annotations:
-        assert annotation.annotation_stream_position == delta_stream_position, (
-            f"Annotation stream position does not match - {annotation.annotation_stream_position} "
-            f"!= {delta_stream_position}"
-        )
-        assert annotation.annotation_delta_type == delta_type, (
-            f"Annotation delta type does not match - {annotation.annotation_delta_type} "
-            f"!= {delta_type}"
-        )
-
-    delta_file_envelopes = []
-    table = pa.concat_tables(tables)
-    total_record_count = len(table)
-    total_size_bytes = int(table.nbytes)
-
-    delta_file = DeltaFileEnvelope.of(
-        stream_position=delta_stream_position,
-        delta_type=delta_type,
-        table=table,
-    )
-    delta_file_envelopes.append(delta_file)
-    return delta_file_envelopes, total_record_count, total_size_bytes
-
-
 def _group_file_records_by_pk_hash_bucket(
     annotated_delta: DeltaAnnotated,
     num_hash_buckets: int,
@@ -103,7 +51,7 @@ def _group_file_records_by_pk_hash_bucket(
         delta_file_envelopes,
         total_record_count,
         total_size_bytes,
-    ) = _read_delta_file_envelopes(
+    ) = read_delta_file_envelopes(
         annotated_delta,
         read_kwargs_provider,
         deltacat_storage,
@@ -187,7 +135,7 @@ def _timed_hash_bucket(input: HashBucketInput):
 @ray.remote
 def hash_bucket(input: HashBucketInput) -> HashBucketResult:
     with ProcessUtilizationOverTimeRange() as process_util:
-        logger.info(f"Starting hash bucket task...")
+        logger.info(f"Starting hash bucket task {input.hb_task_index}...")
 
         # Log node peak memory utilization every 10 seconds
         def log_peak_memory():
@@ -212,7 +160,7 @@ def hash_bucket(input: HashBucketInput) -> HashBucketResult:
             )
             emit_metrics_time = latency
 
-        logger.info(f"Finished hash bucket task...")
+        logger.info(f"Finished hash bucket task {input.hb_task_index}...")
         return HashBucketResult(
             hash_bucket_result[0],
             hash_bucket_result[1],

deltacat/compute/compactor_v2/steps/merge.py

@@ -6,28 +6,21 @@ import pyarrow as pa
 import ray
 import time
 import pyarrow.compute as pc
+import deltacat.compute.compactor_v2.utils.merge as merge_utils
 from uuid import uuid4
-from collections import defaultdict
 from deltacat import logs
-from typing import List, Optional
-from deltacat.types.media import DELIMITED_TEXT_CONTENT_TYPES
+from typing import List, Optional, Tuple
 from deltacat.compute.compactor_v2.model.merge_result import MergeResult
 from deltacat.compute.compactor.model.materialize_result import MaterializeResult
 from deltacat.compute.compactor.model.pyarrow_write_result import PyArrowWriteResult
-from deltacat.compute.compactor import (
-    RoundCompletionInfo,
-    DeltaFileEnvelope,
-)
+from deltacat.compute.compactor import RoundCompletionInfo, DeltaFileEnvelope
 from deltacat.utils.common import ReadKwargsProvider
-
 from contextlib import nullcontext
-from deltacat.types.tables import TABLE_CLASS_TO_SIZE_FUNC
 from deltacat.utils.ray_utils.runtime import (
     get_current_ray_task_id,
     get_current_ray_worker_id,
 )
 from deltacat.compute.compactor.utils import system_columns as sc
-
 from deltacat.utils.performance import timed_invocation
 from deltacat.utils.metrics import emit_timer_metrics
 from deltacat.utils.resources import (
@@ -36,7 +29,6 @@ from deltacat.utils.resources import (
 )
 from deltacat.compute.compactor_v2.utils.primary_key_index import (
     generate_pk_hash_column,
-    hash_group_index_to_hash_bucket_indices,
 )
 from deltacat.storage import (
     Delta,
@@ -77,14 +69,9 @@ def _drop_delta_type_rows(table: pa.Table, delta_type: DeltaType) -> pa.Table:
 
 
 def _build_incremental_table(
-    hash_bucket_index: int,
     df_envelopes_list: List[List[DeltaFileEnvelope]],
 ) -> pa.Table:
 
-    logger.info(
-        f"[Hash bucket index {hash_bucket_index}] Reading dedupe input for "
-        f"{len(df_envelopes_list)} delta file envelope lists..."
-    )
     hb_tables = []
     # sort by delta file stream position now instead of sorting every row later
     df_envelopes = [d for dfe_list in df_envelopes_list for d in dfe_list]
@@ -270,174 +257,120 @@ def _copy_all_manifest_files_from_old_hash_buckets(
     return materialize_result_list
 
 
-def _timed_merge(input: MergeInput) -> MergeResult:
-    def _materialize(
-        hash_bucket_index,
-        compacted_tables: List[pa.Table],
-    ) -> MaterializeResult:
-        compacted_table = pa.concat_tables(compacted_tables)
-        if input.compacted_file_content_type in DELIMITED_TEXT_CONTENT_TYPES:
-            # TODO (rkenmi): Investigate if we still need to convert this table to pandas DataFrame
-            # TODO (pdames): compare performance to pandas-native materialize path
-            df = compacted_table.to_pandas(split_blocks=True, self_destruct=True)
-            compacted_table = df
-        delta, stage_delta_time = timed_invocation(
-            input.deltacat_storage.stage_delta,
-            compacted_table,
-            input.write_to_partition,
-            max_records_per_entry=input.max_records_per_output_file,
-            content_type=input.compacted_file_content_type,
-            s3_table_writer_kwargs=input.s3_table_writer_kwargs,
-            **input.deltacat_storage_kwargs,
-        )
-        compacted_table_size = TABLE_CLASS_TO_SIZE_FUNC[type(compacted_table)](
-            compacted_table
-        )
-        logger.debug(
-            f"Time taken for materialize task"
-            f" to upload {len(compacted_table)} records"
-            f" of size {compacted_table_size} is: {stage_delta_time}s"
+def _compact_tables(
+    input: MergeInput, dfe_list: List[List[DeltaFileEnvelope]], hb_idx: int
+) -> Tuple[pa.Table, int, int]:
+    logger.info(
+        f"[Hash bucket index {hb_idx}] Reading dedupe input for "
+        f"{len(dfe_list)} delta file envelope lists..."
+    )
+    table = _build_incremental_table(dfe_list)
+
+    incremental_len = len(table)
+    logger.info(
+        f"[Hash bucket index {hb_idx}] Got the incremental table of length {incremental_len}"
+    )
+
+    if input.sort_keys:
+        # Incremental is sorted and merged, as sorting
+        # on non event based sort key does not produce consistent
+        # compaction results. E.g., compaction(delta1, delta2, delta3)
+        # will not be equal to compaction(compaction(delta1, delta2), delta3).
+        table = table.sort_by(input.sort_keys)
+
+    compacted_table = None
+
+    if (
+        input.round_completion_info
+        and input.round_completion_info.hb_index_to_entry_range
+        and input.round_completion_info.hb_index_to_entry_range.get(str(hb_idx))
+        is not None
+    ):
+        compacted_table = _download_compacted_table(
+            hb_index=hb_idx,
+            rcf=input.round_completion_info,
+            read_kwargs_provider=input.read_kwargs_provider,
+            deltacat_storage=input.deltacat_storage,
+            deltacat_storage_kwargs=input.deltacat_storage_kwargs,
         )
-        manifest = delta.manifest
-        manifest_records = manifest.meta.record_count
-        assert manifest_records == len(compacted_table), (
-            f"Unexpected Error: Materialized delta manifest record count "
-            f"({manifest_records}) does not equal compacted table record count "
-            f"({len(compacted_table)})"
+
+    hb_table_record_count = len(table) + (
+        len(compacted_table) if compacted_table else 0
+    )
+
+    table, merge_time = timed_invocation(
+        func=_merge_tables,
+        table=table,
+        primary_keys=input.primary_keys,
+        can_drop_duplicates=input.drop_duplicates,
+        compacted_table=compacted_table,
+    )
+    total_deduped_records = hb_table_record_count - len(table)
+
+    logger.info(
+        f"[Merge task index {input.merge_task_index}] Merged "
+        f"record count: {len(table)}, size={table.nbytes} took: {merge_time}s"
+    )
+
+    return table, incremental_len, total_deduped_records
+
+
+def _copy_manifests_from_hash_bucketing(
+    input: MergeInput, hb_index_copy_by_reference_ids: List[int]
+) -> List[MaterializeResult]:
+    materialized_results: List[MaterializeResult] = []
+
+    if input.round_completion_info:
+        referenced_materialized_results = (
+            _copy_all_manifest_files_from_old_hash_buckets(
+                hb_index_copy_by_reference_ids,
+                input.round_completion_info,
+                input.write_to_partition,
+                input.deltacat_storage,
+                input.deltacat_storage_kwargs,
+            )
         )
-        materialize_result = MaterializeResult.of(
-            delta=delta,
-            task_index=hash_bucket_index,
-            # TODO (pdames): Generalize WriteResult to contain in-memory-table-type
-            #  and in-memory-table-bytes instead of tight coupling to paBytes
-            pyarrow_write_result=PyArrowWriteResult.of(
-                len(manifest.entries),
-                TABLE_CLASS_TO_SIZE_FUNC[type(compacted_table)](compacted_table),
-                manifest.meta.content_length,
-                len(compacted_table),
-            ),
+        logger.info(
+            f"Copying {len(referenced_materialized_results)} manifest files by reference..."
         )
-        logger.info(f"Materialize result: {materialize_result}")
-        return materialize_result
+        materialized_results.extend(referenced_materialized_results)
 
+    return materialized_results
+
+
+def _timed_merge(input: MergeInput) -> MergeResult:
     task_id = get_current_ray_task_id()
     worker_id = get_current_ray_worker_id()
     with memray.Tracker(
         f"merge_{worker_id}_{task_id}.bin"
     ) if input.enable_profiler else nullcontext():
-        # In V2, we need to mitigate risk of running out of memory here in cases of
-        #  severe skew of primary key updates in deltas. By severe skew, we mean
-        #  one hash bucket require more memory than a worker instance have.
-        logger.info(
-            f"[Merge task {input.merge_task_index}] Getting delta file envelope "
-            f"groups for {len(input.dfe_groups_refs)} object refs..."
-        )
-
-        delta_file_envelope_groups_list = input.object_store.get_many(
-            input.dfe_groups_refs
-        )
-        hb_index_to_delta_file_envelopes_list = defaultdict(list)
-        for delta_file_envelope_groups in delta_file_envelope_groups_list:
-            assert input.hash_bucket_count == len(delta_file_envelope_groups), (
-                f"The hash bucket count must match the dfe size as {input.hash_bucket_count}"
-                f" != {len(delta_file_envelope_groups)}"
-            )
-
-            for hb_idx, dfes in enumerate(delta_file_envelope_groups):
-                if dfes:
-                    hb_index_to_delta_file_envelopes_list[hb_idx].append(dfes)
-
-        valid_hb_indices_iterable = hash_group_index_to_hash_bucket_indices(
-            input.hash_group_index, input.hash_bucket_count, input.num_hash_groups
-        )
+        total_input_records, total_deduped_records = 0, 0
+        materialized_results: List[MaterializeResult] = []
+        merge_file_groups = input.merge_file_groups_provider.create()
+        hb_index_copy_by_ref_ids = []
 
-        total_deduped_records = 0
-        total_dfes_found = 0
+        for merge_file_group in merge_file_groups:
+            if not merge_file_group.dfe_groups:
+                hb_index_copy_by_ref_ids.append(merge_file_group.hb_index)
+                continue
 
-        materialized_results: List[MaterializeResult] = []
-        hb_index_copy_by_reference = []
-        for hb_idx in valid_hb_indices_iterable:
-            dfe_list = hb_index_to_delta_file_envelopes_list.get(hb_idx)
-
-            if dfe_list:
-                total_dfes_found += 1
-                table = _build_incremental_table(hb_idx, dfe_list)
-
-                incremental_len = len(table)
-                logger.info(
-                    f"Got the incremental table of length {incremental_len} for hash bucket {hb_idx}"
-                )
-
-                if input.sort_keys:
-                    # Incremental is sorted and merged, as sorting
-                    # on non event based sort key does not produce consistent
-                    # compaction results. E.g., compaction(delta1, delta2, delta3)
-                    # will not be equal to compaction(compaction(delta1, delta2), delta3).
-                    table = table.sort_by(input.sort_keys)
-
-                compacted_table = None
-                if (
-                    input.round_completion_info
-                    and input.round_completion_info.hb_index_to_entry_range
-                    and input.round_completion_info.hb_index_to_entry_range.get(
-                        str(hb_idx)
-                    )
-                    is not None
-                ):
-
-                    compacted_table = _download_compacted_table(
-                        hb_index=hb_idx,
-                        rcf=input.round_completion_info,
-                        read_kwargs_provider=input.read_kwargs_provider,
-                        deltacat_storage=input.deltacat_storage,
-                        deltacat_storage_kwargs=input.deltacat_storage_kwargs,
-                    )
-
-                hb_table_record_count = len(table) + (
-                    len(compacted_table) if compacted_table else 0
-                )
-
-                table, merge_time = timed_invocation(
-                    func=_merge_tables,
-                    table=table,
-                    primary_keys=input.primary_keys,
-                    can_drop_duplicates=input.drop_duplicates,
-                    compacted_table=compacted_table,
-                )
-                total_deduped_records += hb_table_record_count - len(table)
-
-                logger.info(
-                    f"[Merge task index {input.merge_task_index}] Merged "
-                    f"record count: {len(table)}, size={table.nbytes} took: {merge_time}s"
-                )
-
-                materialized_results.append(_materialize(hb_idx, [table]))
-            else:
-                hb_index_copy_by_reference.append(hb_idx)
-
-        if input.round_completion_info and hb_index_copy_by_reference:
-            referenced_materialized_results = (
-                _copy_all_manifest_files_from_old_hash_buckets(
-                    hb_index_copy_by_reference,
-                    input.round_completion_info,
-                    input.write_to_partition,
-                    input.deltacat_storage,
-                    input.deltacat_storage_kwargs,
-                )
+            table, input_records, deduped_records = _compact_tables(
+                input, merge_file_group.dfe_groups, merge_file_group.hb_index
             )
-            logger.info(
-                f"Copying {len(referenced_materialized_results)} manifest files by reference..."
+            total_input_records += input_records
+            total_deduped_records += deduped_records
+            materialized_results.append(
+                merge_utils.materialize(input, merge_file_group.hb_index, [table])
             )
-            materialized_results.extend(referenced_materialized_results)
 
-        logger.info(
-            "Total number of materialized results produced for "
-            f"hash group index: {input.hash_group_index} is {len(materialized_results)}"
-        )
+        if hb_index_copy_by_ref_ids:
+            materialized_results.extend(
+                _copy_manifests_from_hash_bucketing(input, hb_index_copy_by_ref_ids)
+            )
 
-        assert total_dfes_found == len(hb_index_to_delta_file_envelopes_list), (
-            "The total dfe list does not match the input dfes from hash bucket as "
-            f"{total_dfes_found} != {len(hb_index_to_delta_file_envelopes_list)}"
+        logger.info(
+            f"[Hash group index: {input.merge_file_groups_provider.hash_group_index}]"
+            f" Total number of materialized results produced: {len(materialized_results)} "
         )
 
         peak_memory_usage_bytes = get_current_process_peak_memory_usage_in_bytes()
@@ -447,6 +380,7 @@ def _timed_merge(input: MergeInput) -> MergeResult:
 
         return MergeResult(
             materialized_results,
+            np.int64(total_input_records),
             np.int64(total_deduped_records),
             np.double(peak_memory_usage_bytes),
             np.double(0.0),
@@ -457,7 +391,7 @@ def _timed_merge(input: MergeInput) -> MergeResult:
 @ray.remote
 def merge(input: MergeInput) -> MergeResult:
     with ProcessUtilizationOverTimeRange() as process_util:
-        logger.info(f"Starting merge task...")
+        logger.info(f"Starting merge task {input.merge_task_index}...")
 
         # Log node peak memory utilization every 10 seconds
         def log_peak_memory():
@@ -480,11 +414,12 @@ def merge(input: MergeInput) -> MergeResult:
             )
             emit_metrics_time = latency
 
-        logger.info(f"Finished merge task...")
+        logger.info(f"Finished merge task {input.merge_task_index}...")
         return MergeResult(
             merge_result[0],
             merge_result[1],
             merge_result[2],
+            merge_result[3],
             np.double(emit_metrics_time),
             merge_result[4],
         )

deltacat/compute/compactor_v2/utils/delta.py

@@ -0,0 +1,97 @@
+import time
+from typing import List, Optional, Tuple
+
+from deltacat.compute.compactor import (
+    DeltaAnnotated,
+    DeltaFileEnvelope,
+)
+
+from deltacat.storage import interface as unimplemented_deltacat_storage
+from deltacat.types.media import StorageType
+from deltacat.utils.common import ReadKwargsProvider
+from deltacat import logs
+
+import pyarrow as pa
+import logging
+
+
+logger = logs.configure_deltacat_logger(logging.getLogger(__name__))
+
+
+def read_delta_file_envelopes(
+    annotated_delta: DeltaAnnotated,
+    read_kwargs_provider: Optional[ReadKwargsProvider],
+    deltacat_storage=unimplemented_deltacat_storage,
+    deltacat_storage_kwargs: Optional[dict] = None,
+) -> Tuple[Optional[List[DeltaFileEnvelope]], int, int]:
+    tables = deltacat_storage.download_delta(
+        annotated_delta,
+        max_parallelism=1,
+        file_reader_kwargs_provider=read_kwargs_provider,
+        storage_type=StorageType.LOCAL,
+        **deltacat_storage_kwargs,
+    )
+    annotations = annotated_delta.annotations
+    assert (
+        len(tables) == len(annotations),
+        f"Unexpected Error: Length of downloaded delta manifest tables "
+        f"({len(tables)}) doesn't match the length of delta manifest "
+        f"annotations ({len(annotations)}).",
+    )
+    if not tables:
+        return None, 0, 0
+
+    delta_stream_position = annotations[0].annotation_stream_position
+    delta_type = annotations[0].annotation_delta_type
+
+    for annotation in annotations:
+        assert annotation.annotation_stream_position == delta_stream_position, (
+            f"Annotation stream position does not match - {annotation.annotation_stream_position} "
+            f"!= {delta_stream_position}"
+        )
+        assert annotation.annotation_delta_type == delta_type, (
+            f"Annotation delta type does not match - {annotation.annotation_delta_type} "
+            f"!= {delta_type}"
+        )
+
+    delta_file_envelopes = []
+    table = pa.concat_tables(tables)
+    total_record_count = len(table)
+    total_size_bytes = int(table.nbytes)
+
+    delta_file = DeltaFileEnvelope.of(
+        stream_position=delta_stream_position,
+        delta_type=delta_type,
+        table=table,
+    )
+    delta_file_envelopes.append(delta_file)
+    return delta_file_envelopes, total_record_count, total_size_bytes
+
+
+def get_local_delta_file_envelopes(
+    uniform_deltas: List[DeltaAnnotated],
+    read_kwargs_provider: Optional[ReadKwargsProvider],
+    deltacat_storage=unimplemented_deltacat_storage,
+    deltacat_storage_kwargs: Optional[dict] = None,
+) -> Tuple[List[DeltaFileEnvelope], int]:
+    local_dfe_list = []
+    input_records_count = 0
+    logger.info(f"Getting {len(uniform_deltas)} DFE Tasks.")
+    dfe_start = time.monotonic()
+    for annotated_delta in uniform_deltas:
+        (
+            delta_file_envelopes,
+            total_record_count,
+            total_size_bytes,
+        ) = read_delta_file_envelopes(
+            annotated_delta,
+            read_kwargs_provider,
+            deltacat_storage,
+            deltacat_storage_kwargs,
+        )
+        if delta_file_envelopes:
+            local_dfe_list.extend(delta_file_envelopes)
+            input_records_count += total_record_count
+    dfe_end = time.monotonic()
+    logger.info(f"Retrieved {len(local_dfe_list)} DFE Tasks in {dfe_end - dfe_start}s.")
+    return local_dfe_list, input_records_count