airbyte-cdk 6.23.0__tar.gz → 6.23.0.dev0__tar.gz

{airbyte_cdk-6.23.0 → airbyte_cdk-6.23.0.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: airbyte-cdk
-Version: 6.23.0
+Version: 6.23.0.dev0
 Summary: A framework for writing Airbyte Connectors.
 License: MIT
 Keywords: airbyte,connector-development-kit,cdk

{airbyte_cdk-6.23.0 → airbyte_cdk-6.23.0.dev0}/airbyte_cdk/sources/declarative/concurrent_declarative_source.py

@@ -20,6 +20,9 @@ from airbyte_cdk.sources.declarative.extractors.record_filter import (
     ClientSideIncrementalRecordFilterDecorator,
 )
 from airbyte_cdk.sources.declarative.incremental.datetime_based_cursor import DatetimeBasedCursor
+from airbyte_cdk.sources.declarative.incremental.per_partition_with_global import (
+    PerPartitionWithGlobalCursor,
+)
 from airbyte_cdk.sources.declarative.interpolation import InterpolatedString
 from airbyte_cdk.sources.declarative.manifest_declarative_source import ManifestDeclarativeSource
 from airbyte_cdk.sources.declarative.models.declarative_component_schema import (
@@ -32,7 +35,7 @@ from airbyte_cdk.sources.declarative.parsers.model_to_component_factory import (
     ModelToComponentFactory,
 )
 from airbyte_cdk.sources.declarative.requesters import HttpRequester
-from airbyte_cdk.sources.declarative.retrievers import SimpleRetriever
+from airbyte_cdk.sources.declarative.retrievers import Retriever, SimpleRetriever
 from airbyte_cdk.sources.declarative.stream_slicers.declarative_partition_generator import (
     DeclarativePartitionFactory,
     StreamSlicerPartitionGenerator,
@@ -231,21 +234,7 @@ class ConcurrentDeclarativeSource(ManifestDeclarativeSource, Generic[TState]):
                         stream_state=stream_state,
                     )
 
-                    retriever = declarative_stream.retriever
-
-                    # This is an optimization so that we don't invoke any cursor or state management flows within the
-                    # low-code framework because state management is handled through the ConcurrentCursor.
-                    if declarative_stream and isinstance(retriever, SimpleRetriever):
-                        # Also a temporary hack. In the legacy Stream implementation, as part of the read,
-                        # set_initial_state() is called to instantiate incoming state on the cursor. Although we no
-                        # longer rely on the legacy low-code cursor for concurrent checkpointing, low-code components
-                        # like StopConditionPaginationStrategyDecorator and ClientSideIncrementalRecordFilterDecorator
-                        # still rely on a DatetimeBasedCursor that is properly initialized with state.
-                        if retriever.cursor:
-                            retriever.cursor.set_initial_state(stream_state=stream_state)
-                        # We zero it out here, but since this is a cursor reference, the state is still properly
-                        # instantiated for the other components that reference it
-                        retriever.cursor = None
+                    retriever = self._get_retriever(declarative_stream, stream_state)
 
                     partition_generator = StreamSlicerPartitionGenerator(
                         DeclarativePartitionFactory(
@@ -305,6 +294,60 @@ class ConcurrentDeclarativeSource(ManifestDeclarativeSource, Generic[TState]):
                             cursor=final_state_cursor,
                         )
                     )
+                elif (
+                    incremental_sync_component_definition
+                    and incremental_sync_component_definition.get("type", "")
+                    == DatetimeBasedCursorModel.__name__
+                    and self._stream_supports_concurrent_partition_processing(
+                        declarative_stream=declarative_stream
+                    )
+                    and hasattr(declarative_stream.retriever, "stream_slicer")
+                    and isinstance(
+                        declarative_stream.retriever.stream_slicer, PerPartitionWithGlobalCursor
+                    )
+                ):
+                    stream_state = state_manager.get_stream_state(
+                        stream_name=declarative_stream.name, namespace=declarative_stream.namespace
+                    )
+                    partition_router = declarative_stream.retriever.stream_slicer._partition_router
+
+                    perpartition_cursor = (
+                        self._constructor.create_concurrent_cursor_from_perpartition_cursor(
+                            state_manager=state_manager,
+                            model_type=DatetimeBasedCursorModel,
+                            component_definition=incremental_sync_component_definition,
+                            stream_name=declarative_stream.name,
+                            stream_namespace=declarative_stream.namespace,
+                            config=config or {},
+                            stream_state=stream_state,
+                            partition_router=partition_router,
+                        )
+                    )
+
+                    retriever = self._get_retriever(declarative_stream, stream_state)
+
+                    partition_generator = StreamSlicerPartitionGenerator(
+                        DeclarativePartitionFactory(
+                            declarative_stream.name,
+                            declarative_stream.get_json_schema(),
+                            retriever,
+                            self.message_repository,
+                        ),
+                        perpartition_cursor,
+                    )
+
+                    concurrent_streams.append(
+                        DefaultStream(
+                            partition_generator=partition_generator,
+                            name=declarative_stream.name,
+                            json_schema=declarative_stream.get_json_schema(),
+                            availability_strategy=AlwaysAvailableAvailabilityStrategy(),
+                            primary_key=get_primary_key_from_stream(declarative_stream.primary_key),
+                            cursor_field=perpartition_cursor.cursor_field.cursor_field_key,
+                            logger=self.logger,
+                            cursor=perpartition_cursor,
+                        )
+                    )
                 else:
                     synchronous_streams.append(declarative_stream)
             else:
@@ -395,6 +438,27 @@ class ConcurrentDeclarativeSource(ManifestDeclarativeSource, Generic[TState]):
                             return False
         return True
 
+    def _get_retriever(
+        self, declarative_stream: DeclarativeStream, stream_state: Mapping[str, Any]
+    ) -> Retriever:
+        retriever = declarative_stream.retriever
+
+        # This is an optimization so that we don't invoke any cursor or state management flows within the
+        # low-code framework because state management is handled through the ConcurrentCursor.
+        if declarative_stream and isinstance(retriever, SimpleRetriever):
+            # Also a temporary hack. In the legacy Stream implementation, as part of the read,
+            # set_initial_state() is called to instantiate incoming state on the cursor. Although we no
+            # longer rely on the legacy low-code cursor for concurrent checkpointing, low-code components
+            # like StopConditionPaginationStrategyDecorator and ClientSideIncrementalRecordFilterDecorator
+            # still rely on a DatetimeBasedCursor that is properly initialized with state.
+            if retriever.cursor:
+                retriever.cursor.set_initial_state(stream_state=stream_state)
+            # We zero it out here, but since this is a cursor reference, the state is still properly
+            # instantiated for the other components that reference it
+            retriever.cursor = None
+
+        return retriever
+
     @staticmethod
     def _select_streams(
         streams: List[AbstractStream], configured_catalog: ConfiguredAirbyteCatalog

{airbyte_cdk-6.23.0 → airbyte_cdk-6.23.0.dev0}/airbyte_cdk/sources/declarative/extractors/record_filter.py

@@ -59,13 +59,11 @@ class ClientSideIncrementalRecordFilterDecorator(RecordFilter):
 
     def __init__(
         self,
-        date_time_based_cursor: DatetimeBasedCursor,
-        substream_cursor: Optional[Union[PerPartitionWithGlobalCursor, GlobalSubstreamCursor]],
+        cursor: Union[DatetimeBasedCursor, PerPartitionWithGlobalCursor, GlobalSubstreamCursor],
         **kwargs: Any,
     ):
         super().__init__(**kwargs)
-        self._date_time_based_cursor = date_time_based_cursor
-        self._substream_cursor = substream_cursor
+        self._cursor = cursor
 
     def filter_records(
         self,
@@ -77,7 +75,7 @@ class ClientSideIncrementalRecordFilterDecorator(RecordFilter):
         records = (
             record
             for record in records
-            if (self._substream_cursor or self._date_time_based_cursor).should_be_synced(
+            if self._cursor.should_be_synced(
                 # Record is created on the fly to align with cursors interface; stream name is ignored as we don't need it here
                 # Record stream name is empty cause it is not used durig the filtering
                 Record(data=record, associated_slice=stream_slice, stream_name="")

{airbyte_cdk-6.23.0 → airbyte_cdk-6.23.0.dev0}/airbyte_cdk/sources/declarative/incremental/__init__.py

@@ -2,6 +2,10 @@
 # Copyright (c) 2022 Airbyte, Inc., all rights reserved.
 #
 
+from airbyte_cdk.sources.declarative.incremental.concurrent_partition_cursor import (
+    ConcurrentCursorFactory,
+    ConcurrentPerPartitionCursor,
+)
 from airbyte_cdk.sources.declarative.incremental.datetime_based_cursor import DatetimeBasedCursor
 from airbyte_cdk.sources.declarative.incremental.declarative_cursor import DeclarativeCursor
 from airbyte_cdk.sources.declarative.incremental.global_substream_cursor import (
@@ -21,6 +25,8 @@ from airbyte_cdk.sources.declarative.incremental.resumable_full_refresh_cursor i
 
 __all__ = [
     "CursorFactory",
+    "ConcurrentCursorFactory",
+    "ConcurrentPerPartitionCursor",
     "DatetimeBasedCursor",
     "DeclarativeCursor",
     "GlobalSubstreamCursor",

airbyte_cdk-6.23.0.dev0/airbyte_cdk/sources/declarative/incremental/concurrent_partition_cursor.py

@@ -0,0 +1,333 @@
+#
+# Copyright (c) 2023 Airbyte, Inc., all rights reserved.
+#
+
+import copy
+import logging
+import threading
+from collections import OrderedDict
+from copy import deepcopy
+from datetime import timedelta
+from typing import Any, Callable, Iterable, Mapping, MutableMapping, Optional
+
+from airbyte_cdk.sources.connector_state_manager import ConnectorStateManager
+from airbyte_cdk.sources.declarative.incremental.global_substream_cursor import (
+    Timer,
+    iterate_with_last_flag_and_state,
+)
+from airbyte_cdk.sources.declarative.partition_routers.partition_router import PartitionRouter
+from airbyte_cdk.sources.message import MessageRepository
+from airbyte_cdk.sources.streams.checkpoint.per_partition_key_serializer import (
+    PerPartitionKeySerializer,
+)
+from airbyte_cdk.sources.streams.concurrent.cursor import ConcurrentCursor, Cursor, CursorField
+from airbyte_cdk.sources.streams.concurrent.partitions.partition import Partition
+from airbyte_cdk.sources.types import Record, StreamSlice, StreamState
+
+logger = logging.getLogger("airbyte")
+
+
+class ConcurrentCursorFactory:
+    def __init__(self, create_function: Callable[..., ConcurrentCursor]):
+        self._create_function = create_function
+
+    def create(
+        self, stream_state: Mapping[str, Any], runtime_lookback_window: Optional[timedelta]
+    ) -> ConcurrentCursor:
+        return self._create_function(
+            stream_state=stream_state, runtime_lookback_window=runtime_lookback_window
+        )
+
+
+class ConcurrentPerPartitionCursor(Cursor):
+    """
+    Manages state per partition when a stream has many partitions, preventing data loss or duplication.
+
+    Attributes:
+        DEFAULT_MAX_PARTITIONS_NUMBER (int): Maximum number of partitions to retain in memory (default is 10,000).
+
+    - **Partition Limitation Logic**
+      Ensures the number of tracked partitions does not exceed the specified limit to prevent memory overuse. Oldest partitions are removed when the limit is reached.
+
+    - **Global Cursor Fallback**
+      New partitions use global state as the initial state to progress the state for deleted or new partitions. The history data added after the initial sync will be missing.
+
+    CurrentPerPartitionCursor expects the state of the ConcurrentCursor to follow the format {cursor_field: cursor_value}.
+    """
+
+    DEFAULT_MAX_PARTITIONS_NUMBER = 10000
+    _NO_STATE: Mapping[str, Any] = {}
+    _NO_CURSOR_STATE: Mapping[str, Any] = {}
+    _GLOBAL_STATE_KEY = "state"
+    _PERPARTITION_STATE_KEY = "states"
+    _KEY = 0
+    _VALUE = 1
+
+    def __init__(
+        self,
+        cursor_factory: ConcurrentCursorFactory,
+        partition_router: PartitionRouter,
+        stream_name: str,
+        stream_namespace: Optional[str],
+        stream_state: Any,
+        message_repository: MessageRepository,
+        connector_state_manager: ConnectorStateManager,
+        cursor_field: CursorField,
+    ) -> None:
+        self._global_cursor: Optional[StreamState] = {}
+        self._stream_name = stream_name
+        self._stream_namespace = stream_namespace
+        self._message_repository = message_repository
+        self._connector_state_manager = connector_state_manager
+        self._cursor_field = cursor_field
+
+        self._cursor_factory = cursor_factory
+        self._partition_router = partition_router
+
+        # The dict is ordered to ensure that once the maximum number of partitions is reached,
+        # the oldest partitions can be efficiently removed, maintaining the most recent partitions.
+        self._cursor_per_partition: OrderedDict[str, ConcurrentCursor] = OrderedDict()
+        self._semaphore_per_partition: OrderedDict[str, threading.Semaphore] = OrderedDict()
+        self._finished_partitions: set[str] = set()
+        self._lock = threading.Lock()
+        self._timer = Timer()
+        self._new_global_cursor: Optional[StreamState] = None
+        self._lookback_window: int = 0
+        self._parent_state: Optional[StreamState] = None
+        self._over_limit: int = 0
+        self._partition_serializer = PerPartitionKeySerializer()
+
+        self._set_initial_state(stream_state)
+
+    @property
+    def cursor_field(self) -> CursorField:
+        return self._cursor_field
+
+    @property
+    def state(self) -> MutableMapping[str, Any]:
+        states = []
+        for partition_tuple, cursor in self._cursor_per_partition.items():
+            if cursor.state:
+                states.append(
+                    {
+                        "partition": self._to_dict(partition_tuple),
+                        "cursor": copy.deepcopy(cursor.state),
+                    }
+                )
+        state: dict[str, Any] = {self._PERPARTITION_STATE_KEY: states}
+
+        if self._global_cursor:
+            state[self._GLOBAL_STATE_KEY] = self._global_cursor
+        if self._lookback_window is not None:
+            state["lookback_window"] = self._lookback_window
+        if self._parent_state is not None:
+            state["parent_state"] = self._parent_state
+        return state
+
+    def close_partition(self, partition: Partition) -> None:
+        # Attempt to retrieve the stream slice
+        stream_slice: Optional[StreamSlice] = partition.to_slice()  # type: ignore[assignment]
+
+        # Ensure stream_slice is not None
+        if stream_slice is None:
+            raise ValueError("stream_slice cannot be None")
+
+        partition_key = self._to_partition_key(stream_slice.partition)
+        self._cursor_per_partition[partition_key].close_partition(partition=partition)
+        with self._lock:
+            self._semaphore_per_partition[partition_key].acquire()
+            cursor = self._cursor_per_partition[partition_key]
+            if (
+                partition_key in self._finished_partitions
+                and self._semaphore_per_partition[partition_key]._value == 0
+            ):
+                if (
+                    self._new_global_cursor is None
+                    or self._new_global_cursor[self.cursor_field.cursor_field_key]
+                    < cursor.state[self.cursor_field.cursor_field_key]
+                ):
+                    self._new_global_cursor = copy.deepcopy(cursor.state)
+
+    def ensure_at_least_one_state_emitted(self) -> None:
+        """
+        The platform expect to have at least one state message on successful syncs. Hence, whatever happens, we expect this method to be
+        called.
+        """
+        if not any(
+            semaphore_item[1]._value for semaphore_item in self._semaphore_per_partition.items()
+        ):
+            self._global_cursor = self._new_global_cursor
+            self._lookback_window = self._timer.finish()
+            self._parent_state = self._partition_router.get_stream_state()
+        self._emit_state_message()
+
+    def _emit_state_message(self) -> None:
+        self._connector_state_manager.update_state_for_stream(
+            self._stream_name,
+            self._stream_namespace,
+            self.state,
+        )
+        state_message = self._connector_state_manager.create_state_message(
+            self._stream_name, self._stream_namespace
+        )
+        self._message_repository.emit_message(state_message)
+
+    def stream_slices(self) -> Iterable[StreamSlice]:
+        if self._timer.is_running():
+            raise RuntimeError("stream_slices has been executed more than once.")
+
+        slices = self._partition_router.stream_slices()
+        self._timer.start()
+        for partition in slices:
+            yield from self._generate_slices_from_partition(partition)
+
+    def _generate_slices_from_partition(self, partition: StreamSlice) -> Iterable[StreamSlice]:
+        # Ensure the maximum number of partitions is not exceeded
+        self._ensure_partition_limit()
+
+        cursor = self._cursor_per_partition.get(self._to_partition_key(partition.partition))
+        if not cursor:
+            cursor = self._create_cursor(
+                self._global_cursor,
+                self._lookback_window if self._global_cursor else 0,
+            )
+            self._cursor_per_partition[self._to_partition_key(partition.partition)] = cursor
+            self._semaphore_per_partition[self._to_partition_key(partition.partition)] = (
+                threading.Semaphore(0)
+            )
+
+        for cursor_slice, is_last_slice, _ in iterate_with_last_flag_and_state(
+            cursor.stream_slices(),
+            lambda: None,
+        ):
+            self._semaphore_per_partition[self._to_partition_key(partition.partition)].release()
+            if is_last_slice:
+                self._finished_partitions.add(self._to_partition_key(partition.partition))
+            yield StreamSlice(
+                partition=partition, cursor_slice=cursor_slice, extra_fields=partition.extra_fields
+            )
+
+    def _ensure_partition_limit(self) -> None:
+        """
+        Ensure the maximum number of partitions is not exceeded. If so, the oldest added partition will be dropped.
+        """
+        while len(self._cursor_per_partition) > self.DEFAULT_MAX_PARTITIONS_NUMBER - 1:
+            self._over_limit += 1
+            oldest_partition = self._cursor_per_partition.popitem(last=False)[
+                0
+            ]  # Remove the oldest partition
+            logger.warning(
+                f"The maximum number of partitions has been reached. Dropping the oldest partition: {oldest_partition}. Over limit: {self._over_limit}."
+            )
+
+    def _set_initial_state(self, stream_state: StreamState) -> None:
+        """
+        Initialize the cursor's state using the provided `stream_state`.
+
+        This method supports global and per-partition state initialization.
+
+        - **Global State**: If `states` is missing, the `state` is treated as global and applied to all partitions.
+          The `global state` holds a single cursor position representing the latest processed record across all partitions.
+
+        - **Lookback Window**: Configured via `lookback_window`, it defines the period (in seconds) for reprocessing records.
+          This ensures robustness in case of upstream data delays or reordering. If not specified, it defaults to 0.
+
+        - **Per-Partition State**: If `states` is present, each partition's cursor state is initialized separately.
+
+        - **Parent State**: (if available) Used to initialize partition routers based on parent streams.
+
+        Args:
+            stream_state (StreamState): The state of the streams to be set. The format of the stream state should be:
+                {
+                    "states": [
+                        {
+                            "partition": {
+                                "partition_key": "value"
+                            },
+                            "cursor": {
+                                "last_updated": "2023-05-27T00:00:00Z"
+                            }
+                        }
+                    ],
+                    "state": {
+                        "last_updated": "2023-05-27T00:00:00Z"
+                    },
+                    lookback_window: 10,
+                    "parent_state": {
+                        "parent_stream_name": {
+                            "last_updated": "2023-05-27T00:00:00Z"
+                        }
+                    }
+                }
+        """
+        if not stream_state:
+            return
+
+        if self._PERPARTITION_STATE_KEY not in stream_state:
+            # We assume that `stream_state` is in a global format that can be applied to all partitions.
+            # Example: {"global_state_format_key": "global_state_format_value"}
+            self._global_cursor = deepcopy(stream_state)
+            self._new_global_cursor = deepcopy(stream_state)
+
+        else:
+            self._lookback_window = int(stream_state.get("lookback_window", 0))
+
+            for state in stream_state[self._PERPARTITION_STATE_KEY]:
+                self._cursor_per_partition[self._to_partition_key(state["partition"])] = (
+                    self._create_cursor(state["cursor"])
+                )
+                self._semaphore_per_partition[self._to_partition_key(state["partition"])] = (
+                    threading.Semaphore(0)
+                )
+
+            # set default state for missing partitions if it is per partition with fallback to global
+            if self._GLOBAL_STATE_KEY in stream_state:
+                self._global_cursor = deepcopy(stream_state[self._GLOBAL_STATE_KEY])
+                self._new_global_cursor = deepcopy(stream_state[self._GLOBAL_STATE_KEY])
+
+        # Set initial parent state
+        if stream_state.get("parent_state"):
+            self._parent_state = stream_state["parent_state"]
+
+        # Set parent state for partition routers based on parent streams
+        self._partition_router.set_initial_state(stream_state)
+
+    def observe(self, record: Record) -> None:
+        if not record.associated_slice:
+            raise ValueError(
+                "Invalid state as stream slices that are emitted should refer to an existing cursor"
+            )
+        self._cursor_per_partition[
+            self._to_partition_key(record.associated_slice.partition)
+        ].observe(record)
+
+    def _to_partition_key(self, partition: Mapping[str, Any]) -> str:
+        return self._partition_serializer.to_partition_key(partition)
+
+    def _to_dict(self, partition_key: str) -> Mapping[str, Any]:
+        return self._partition_serializer.to_partition(partition_key)
+
+    def _create_cursor(
+        self, cursor_state: Any, runtime_lookback_window: int = 0
+    ) -> ConcurrentCursor:
+        cursor = self._cursor_factory.create(
+            stream_state=deepcopy(cursor_state),
+            runtime_lookback_window=timedelta(seconds=runtime_lookback_window),
+        )
+        return cursor
+
+    def should_be_synced(self, record: Record) -> bool:
+        return self._get_cursor(record).should_be_synced(record)
+
+    def _get_cursor(self, record: Record) -> ConcurrentCursor:
+        if not record.associated_slice:
+            raise ValueError(
+                "Invalid state as stream slices that are emitted should refer to an existing cursor"
+            )
+        partition_key = self._to_partition_key(record.associated_slice.partition)
+        if partition_key not in self._cursor_per_partition:
+            raise ValueError(
+                "Invalid state as stream slices that are emitted should refer to an existing cursor"
+            )
+        cursor = self._cursor_per_partition[partition_key]
+        return cursor

{airbyte_cdk-6.23.0 → airbyte_cdk-6.23.0.dev0}/airbyte_cdk/sources/declarative/incremental/global_substream_cursor.py

@@ -64,6 +64,9 @@ class Timer:
         else:
             raise RuntimeError("Global substream cursor timer not started")
 
+    def is_running(self) -> bool:
+        return self._start is not None
+
 
 class GlobalSubstreamCursor(DeclarativeCursor):
     """

{airbyte_cdk-6.23.0 → airbyte_cdk-6.23.0.dev0}/airbyte_cdk/sources/declarative/incremental/per_partition_cursor.py

@@ -303,6 +303,21 @@ class PerPartitionCursor(DeclarativeCursor):
             raise ValueError("A partition needs to be provided in order to get request body json")
 
     def should_be_synced(self, record: Record) -> bool:
+        if (
+            record.associated_slice
+            and self._to_partition_key(record.associated_slice.partition)
+            not in self._cursor_per_partition
+        ):
+            partition_state = (
+                self._state_to_migrate_from
+                if self._state_to_migrate_from
+                else self._NO_CURSOR_STATE
+            )
+            cursor = self._create_cursor(partition_state)
+
+            self._cursor_per_partition[
+                self._to_partition_key(record.associated_slice.partition)
+            ] = cursor
         return self._get_cursor(record).should_be_synced(
             self._convert_record_to_cursor_record(record)
         )