[DOP-28871] Use bulk DB fetch in consumer

Author: dolfinus

data_rentgen/consumer/saver.py

@@ -0,0 +1,179 @@
+# SPDX-FileCopyrightText: 2024-2025 MTS PJSC
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from faststream import Logger
+from sqlalchemy.exc import DatabaseError, IntegrityError
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from data_rentgen.consumer.extractors import BatchExtractionResult
+from data_rentgen.services.uow import UnitOfWork
+
+
+class DatabaseSaver:
+    def __init__(
+        self,
+        session: AsyncSession,
+        logger: Logger,
+    ) -> None:
+        self.unit_of_work = UnitOfWork(session)
+        self.logger = logger
+
+    async def save(self, data: BatchExtractionResult):
+        self.logger.info("Saving to database")
+
+        await self.create_locations(data)
+        await self.create_datasets(data)
+        await self.create_dataset_symlinks(data)
+        await self.create_job_types(data)
+        await self.create_jobs(data)
+        await self.create_users(data)
+        await self.create_sql_queries(data)
+        await self.create_schemas(data)
+
+        try:
+            await self.create_runs_bulk(data)
+        except DatabaseError:
+            await self.create_runs_one_by_one(data)
+
+        await self.create_operations(data)
+        await self.create_inputs(data)
+        await self.create_outputs(data)
+        await self.create_column_lineage(data)
+
+        self.logger.info("Saved successfully")
+
+    async def create_locations(self, data: BatchExtractionResult):
+        self.logger.debug("Creating locations")
+        # It's hard to fetch locations in bulk, and number of locations is usually small,
+        # so using a row-by-row approach
+        for location_dto in data.locations():
+            async with self.unit_of_work:
+                location = await self.unit_of_work.location.create_or_update(location_dto)
+                location_dto.id = location.id
+
+    # To avoid deadlocks when parallel consumer instances insert/update the same row,
+    # commit changes for each row instead of committing the whole batch. Yes, this cloud be slow.
+    # But most entities are unchanged after creation, so we could just fetch them, and do nothing.
+    async def create_datasets(self, data: BatchExtractionResult):
+        self.logger.debug("Creating datasets")
+        dataset_pairs = await self.unit_of_work.dataset.get_bulk(data.datasets())
+        for dataset_dto, dataset in dataset_pairs:
+            if not dataset:
+                async with self.unit_of_work:
+                    dataset = await self.unit_of_work.dataset.create(dataset_dto)  # noqa: PLW2901
+            dataset_dto.id = dataset.id
+
+    async def create_dataset_symlinks(self, data: BatchExtractionResult):
+        self.logger.debug("Creating dataset symlinks")
+        dataset_symlinks_pairs = await self.unit_of_work.dataset_symlink.fetch_bulk(data.dataset_symlinks())
+        for dataset_symlink_dto, dataset_symlink in dataset_symlinks_pairs:
+            if not dataset_symlink:
+                async with self.unit_of_work:
+                    dataset_symlink = await self.unit_of_work.dataset_symlink.create(dataset_symlink_dto)  # noqa: PLW2901
+            dataset_symlink_dto.id = dataset_symlink.id
+
+    async def create_job_types(self, data: BatchExtractionResult):
+        self.logger.debug("Creating job types")
+        job_type_pairs = await self.unit_of_work.job_type.get_bulk(data.job_types())
+        for job_type_dto, job_type in job_type_pairs:
+            if not job_type:
+                async with self.unit_of_work:
+                    job_type = await self.unit_of_work.job_type.create(job_type_dto)  # noqa: PLW2901
+            job_type_dto.id = job_type.id
+
+    async def create_jobs(self, data: BatchExtractionResult):
+        self.logger.debug("Creating jobs")
+        job_pairs = await self.unit_of_work.job.get_bulk(data.jobs())
+        for job_dto, job in job_pairs:
+            async with self.unit_of_work:
+                if not job:
+                    job = await self.unit_of_work.job.create_or_update(job_dto)  # noqa: PLW2901
+                else:
+                    job = await self.unit_of_work.job.update(job, job_dto)  # noqa: PLW2901
+                job_dto.id = job.id
+
+    async def create_users(self, data: BatchExtractionResult):
+        self.logger.debug("Creating users")
+        user_pairs = await self.unit_of_work.user.fetch_bulk(data.users())
+        for user_dto, user in user_pairs:
+            if not user:
+                async with self.unit_of_work:
+                    user = await self.unit_of_work.user.create(user_dto)  # noqa: PLW2901
+            user_dto.id = user.id
+
+    async def create_sql_queries(self, data: BatchExtractionResult):
+        self.logger.debug("Creating sql queries")
+        sql_query_ids = await self.unit_of_work.sql_query.fetch_known_ids(data.sql_queries())
+        for sql_query_dto, sql_query_id in sql_query_ids:
+            if not sql_query_id:
+                async with self.unit_of_work:
+                    sql_query = await self.unit_of_work.sql_query.create(sql_query_dto)
+                    sql_query_dto.id = sql_query.id
+            else:
+                sql_query_dto.id = sql_query_id
+
+    async def create_schemas(self, data: BatchExtractionResult):
+        self.logger.debug("Creating schemas")
+        schema_ids = await self.unit_of_work.schema.fetch_known_ids(data.schemas())
+        for schema_dto, schema_id in schema_ids:
+            if not schema_id:
+                async with self.unit_of_work:
+                    schema = await self.unit_of_work.schema.create(schema_dto)
+                    schema_dto.id = schema.id
+            else:
+                schema_dto.id = schema_id
+
+    # In most cases, all the run tree created by some parent is send into one
+    # Kafka partition, and thus handled by just one worker.
+    # Cross fingers and create all runs in one transaction.
+    async def create_runs_bulk(self, data: BatchExtractionResult):
+        self.logger.debug("Creating runs in bulk")
+        async with self.unit_of_work:
+            await self.unit_of_work.run.create_or_update_bulk(data.runs())
+
+    # In case then child and parent runs are in different partitions,
+    # multiple workers may try to create/update the same run, leading to a deadlock.
+    # Fallback to creating runs one by one
+    async def create_runs_one_by_one(self, data: BatchExtractionResult):
+        self.logger.debug("Creating runs in one-by-one")
+        run_pairs = await self.unit_of_work.run.fetch_bulk(data.runs())
+        for run_dto, run in run_pairs:
+            try:
+                async with self.unit_of_work:
+                    if not run:
+                        await self.unit_of_work.run.create(run_dto)
+                    else:
+                        await self.unit_of_work.run.update(run, run_dto)
+            except IntegrityError:  # noqa: PERF203
+                # deadlock occurred, states in DB and RAM are out of sync,
+                # so we have to fetch run from DB
+                async with self.unit_of_work:
+                    await self.unit_of_work.run.create_or_update(run_dto)
+
+    # All events related to same operation are always send to the same Kafka partition,
+    # so other workers never insert/update the same operation in parallel.
+    # These rows can be inserted/updated in bulk, in one transaction.
+    async def create_operations(self, data: BatchExtractionResult):
+        async with self.unit_of_work:
+            self.logger.debug("Creating operations")
+            await self.unit_of_work.operation.create_or_update_bulk(data.operations())
+
+    async def create_inputs(self, data: BatchExtractionResult):
+        async with self.unit_of_work:
+            self.logger.debug("Creating inputs")
+            await self.unit_of_work.input.create_or_update_bulk(data.inputs())
+
+    async def create_outputs(self, data: BatchExtractionResult):
+        async with self.unit_of_work:
+            self.logger.debug("Creating outputs")
+            await self.unit_of_work.output.create_or_update_bulk(data.outputs())
+
+    async def create_column_lineage(self, data: BatchExtractionResult):
+        async with self.unit_of_work:
+            self.logger.debug("Creating dataset column relations")
+            await self.unit_of_work.dataset_column_relation.create_bulk_for_column_lineage(data.column_lineage())
+
+            self.logger.debug("Creating column lineage")
+            await self.unit_of_work.column_lineage.create_bulk(data.column_lineage())

data_rentgen/consumer/subscribers.py

@@ -14,10 +14,10 @@
 from pydantic import TypeAdapter
 from sqlalchemy.ext.asyncio import AsyncSession
 
-from data_rentgen.consumer.extractors import BatchExtractionResult, BatchExtractor
+from data_rentgen.consumer.extractors import BatchExtractor
+from data_rentgen.consumer.saver import DatabaseSaver
 from data_rentgen.dependencies.stub import Stub
 from data_rentgen.openlineage.run_event import OpenLineageRunEvent
-from data_rentgen.services.uow import UnitOfWork
 
 __all__ = [
     "runs_events_subscriber",
@@ -54,9 +54,8 @@ async def runs_events_subscriber(
     extracted = extractor.result
     logger.info("Got %r", extracted)
 
-    logger.info("Saving to database")
-    await save_to_db(extracted, session, logger)
-    logger.info("Saved successfully")
+    saver = DatabaseSaver(session, logger)
+    await saver.save(extracted)
 
     if malformed:
         logger.warning("Malformed messages: %d", len(malformed))
@@ -88,96 +87,6 @@ async def parse_messages(
         await asyncio.sleep(0)
 
 
-async def save_to_db(
-    data: BatchExtractionResult,
-    session: AsyncSession,
-    logger: Logger,
-) -> None:
-    # To avoid deadlocks when parallel consumer instances insert/update the same row,
-    # commit changes for each row instead of committing the whole batch. Yes, this cloud be slow.
-
-    unit_of_work = UnitOfWork(session)
-
-    logger.debug("Creating locations")
-    for location_dto in data.locations():
-        async with unit_of_work:
-            location = await unit_of_work.location.create_or_update(location_dto)
-            location_dto.id = location.id
-
-    logger.debug("Creating datasets")
-    for dataset_dto in data.datasets():
-        async with unit_of_work:
-            dataset = await unit_of_work.dataset.get_or_create(dataset_dto)
-            dataset_dto.id = dataset.id
-
-    logger.debug("Creating symlinks")
-    for dataset_symlink_dto in data.dataset_symlinks():
-        async with unit_of_work:
-            dataset_symlink = await unit_of_work.dataset_symlink.get_or_create(dataset_symlink_dto)
-            dataset_symlink_dto.id = dataset_symlink.id
-
-    logger.debug("Creating job types")
-    for job_type_dto in data.job_types():
-        async with unit_of_work:
-            job_type = await unit_of_work.job_type.get_or_create(job_type_dto)
-            job_type_dto.id = job_type.id
-
-    logger.debug("Creating jobs")
-    for job_dto in data.jobs():
-        async with unit_of_work:
-            job = await unit_of_work.job.create_or_update(job_dto)
-            job_dto.id = job.id
-
-    logger.debug("Creating sql queries")
-    for sql_query_dto in data.sql_queries():
-        async with unit_of_work:
-            sql_query = await unit_of_work.sql_query.get_or_create(sql_query_dto)
-            sql_query_dto.id = sql_query.id
-
-    logger.debug("Creating users")
-    for user_dto in data.users():
-        async with unit_of_work:
-            user = await unit_of_work.user.get_or_create(user_dto)
-            user_dto.id = user.id
-
-    logger.debug("Creating schemas")
-    for schema_dto in data.schemas():
-        async with unit_of_work:
-            schema = await unit_of_work.schema.get_or_create(schema_dto)
-            schema_dto.id = schema.id
-
-    # Some events related to specific run are send to the same Kafka partition,
-    # but at the same time we have parent_run which may be already inserted/updated by other worker
-    # (Kafka key maybe different for run and it's parent).
-    # In this case we cannot insert all the rows in one transaction, as it may lead to deadlocks.
-    logger.debug("Creating runs")
-    for run_dto in data.runs():
-        async with unit_of_work:
-            await unit_of_work.run.create_or_update(run_dto)
-
-    # All events related to same operation are always send to the same Kafka partition,
-    # so other workers never insert/update the same operation in parallel.
-    # These rows can be inserted/updated in bulk, in one transaction.
-    async with unit_of_work:
-        logger.debug("Creating operations")
-        await unit_of_work.operation.create_or_update_bulk(data.operations())
-
-        logger.debug("Creating inputs")
-        await unit_of_work.input.create_or_update_bulk(data.inputs())
-
-        logger.debug("Creating outputs")
-        await unit_of_work.output.create_or_update_bulk(data.outputs())
-
-    # If something went wrong here, at least we will have inputs/outputs
-    async with unit_of_work:
-        column_lineage = data.column_lineage()
-        logger.debug("Creating dataset column relations")
-        await unit_of_work.dataset_column_relation.create_bulk_for_column_lineage(column_lineage)
-
-        logger.debug("Creating column lineage")
-        await unit_of_work.column_lineage.create_bulk(column_lineage)
-
-
 async def report_malformed(
     messages: list[ConsumerRecord],
     message_id: str,

data_rentgen/db/repositories/column_lineage.py

@@ -6,7 +6,7 @@
 from typing import NamedTuple
 from uuid import UUID
 
-from sqlalchemy import ColumnElement, any_, func, select, tuple_
+from sqlalchemy import ARRAY, ColumnElement, Integer, any_, cast, func, select, tuple_
 from sqlalchemy.dialects.postgresql import insert
 
 from data_rentgen.db.models import ColumnLineage, DatasetColumnRelation
@@ -123,9 +123,20 @@ async def list_by_dataset_pairs(
         if not dataset_ids_pairs:
             return []
 
+        source_dataset_ids = [pair[0] for pair in dataset_ids_pairs]
+        target_dataset_ids = [pair[1] for pair in dataset_ids_pairs]
+        pairs = (
+            func.unnest(
+                cast(source_dataset_ids, ARRAY(Integer())),
+                cast(target_dataset_ids, ARRAY(Integer())),
+            )
+            .table_valued("source_dataset_id", "target_dataset_id")
+            .render_derived()
+        )
+
         where = [
             ColumnLineage.created_at >= since,
-            tuple_(ColumnLineage.source_dataset_id, ColumnLineage.target_dataset_id).in_(dataset_ids_pairs),
+            tuple_(ColumnLineage.source_dataset_id, ColumnLineage.target_dataset_id).in_(select(pairs)),
         ]
         if until:
             where.append(