f32ddf78-b8ca-49ab-9751-2526b355995f

Committed 27 Mar 2025 01:39PM UTC coverage: 86.88% (+0.03%) from 86.854%

Build # f32ddf78-b8ca-49ab-9751-2526b355995f

Build Type

push

circleci

Committed by

web-flow

Commit Message

add localstack 4.3 blog to the README (#12445)

Coverage Stats

63268 of 72822 relevant lines covered (86.88%)

0.87 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

86.73

/localstack-core/localstack/services/lambda_/event_source_mapping/pollers/stream_poller.py

import json
import logging
import threading
from abc import abstractmethod
from collections import defaultdict
from datetime import datetime
from typing import Iterator

from botocore.client import BaseClient
from botocore.exceptions import ClientError

from localstack.aws.api.pipes import (
    OnPartialBatchItemFailureStreams,
)
from localstack.services.lambda_.event_source_mapping.event_processor import (
    BatchFailureError,
    CustomerInvocationError,
    EventProcessor,
    PartialBatchFailureError,
    PipeInternalError,
)
from localstack.services.lambda_.event_source_mapping.pipe_utils import (
    get_current_time,
    get_datetime_from_timestamp,
    get_internal_client,
)
from localstack.services.lambda_.event_source_mapping.pollers.poller import (
    EmptyPollResultsException,
    Poller,
    get_batch_item_failures,
)
from localstack.services.lambda_.event_source_mapping.pollers.sqs_poller import get_queue_url
from localstack.services.lambda_.event_source_mapping.senders.sender_utils import (
    batched,
)
from localstack.utils.aws.arns import parse_arn, s3_bucket_name
from localstack.utils.backoff import ExponentialBackoff
from localstack.utils.batch_policy import Batcher
from localstack.utils.strings import long_uid

LOG = logging.getLogger(__name__)


# TODO: fix this poller to support resharding
#   https://docs.aws.amazon.com/streams/latest/dev/kinesis-using-sdk-java-resharding.html
class StreamPoller(Poller):
    # Mapping of shard id => shard iterator
    shards: dict[str, str]
    # Iterator for round-robin polling from different shards because a batch cannot contain events from different shards
    # This is a workaround for not handling shards in parallel.
    iterator_over_shards: Iterator[tuple[str, str]] | None
    # ESM UUID is needed in failure processing to form s3 failure destination object key
    esm_uuid: str | None

    # The ARN of the processor (e.g., Pipe ARN)
    partner_resource_arn: str | None

    # Used for backing-off between retries and breaking the retry loop
    _is_shutdown: threading.Event

    # Collects and flushes a batch of records based on a batching policy
    shard_batcher: dict[str, Batcher[dict]]

    def __init__(
        self,
        source_arn: str,
        source_parameters: dict | None = None,
        source_client: BaseClient | None = None,
        processor: EventProcessor | None = None,
        partner_resource_arn: str | None = None,
        esm_uuid: str | None = None,
    ):
        super().__init__(source_arn, source_parameters, source_client, processor)
        self.partner_resource_arn = partner_resource_arn
        self.esm_uuid = esm_uuid
        self.shards = {}
        self.iterator_over_shards = None

        self._is_shutdown = threading.Event()

        self.shard_batcher = defaultdict(
            lambda: Batcher(
                max_count=self.stream_parameters.get("BatchSize", 100),
                max_window=self.stream_parameters.get("MaximumBatchingWindowInSeconds", 0),
            )
        )

    @abstractmethod
    def transform_into_events(self, records: list[dict], shard_id) -> list[dict]:
        pass

    @property
    @abstractmethod
    def stream_parameters(self) -> dict:
        pass

    @abstractmethod
    def initialize_shards(self) -> dict[str, str]:
        """Returns a shard dict mapping from shard id -> shard iterator
        The implementations for Kinesis and DynamoDB are similar but differ in various ways:
        * Kinesis uses "StreamARN" and DynamoDB uses "StreamArn" as source parameter
        * Kinesis uses "StreamStatus.ACTIVE" and DynamoDB uses "StreamStatus.ENABLED"
        * Only Kinesis supports the additional StartingPosition called "AT_TIMESTAMP" using "StartingPositionTimestamp"
        """
        pass

    @abstractmethod
    def stream_arn_param(self) -> dict:
        """Returns a dict of the correct key/value pair for the stream arn used in GetRecords.
        Either StreamARN for Kinesis or {} for DynamoDB (unsupported)"""
        pass

    @abstractmethod
    def failure_payload_details_field_name(self) -> str:
        pass

    @abstractmethod
    def get_approximate_arrival_time(self, record: dict) -> float:
        pass

    @abstractmethod
    def format_datetime(self, time: datetime) -> str:
        """Formats a datetime in the correct format for DynamoDB (with ms) or Kinesis (without ms)"""
        pass

    @abstractmethod
    def get_sequence_number(self, record: dict) -> str:
        pass

    def close(self):
        self._is_shutdown.set()

    def pre_filter(self, events: list[dict]) -> list[dict]:
        return events

    def post_filter(self, events: list[dict]) -> list[dict]:
        return events

    def poll_events(self):
        """Generalized poller for streams such as Kinesis or DynamoDB
        Examples of Kinesis consumers:
        * StackOverflow: https://stackoverflow.com/a/22403036/6875981
        * AWS Sample: https://github.com/aws-samples/kinesis-poster-worker/blob/master/worker.py
        Examples of DynamoDB consumers:
        * Blogpost: https://www.tecracer.com/blog/2022/05/getting-a-near-real-time-view-of-a-dynamodb-stream-with-python.html
        """
        # TODO: consider potential shard iterator timeout after 300 seconds (likely not relevant with short-polling):
        #   https://docs.aws.amazon.com/streams/latest/dev/troubleshooting-consumers.html#shard-iterator-expires-unexpectedly
        #  Does this happen if no records are received for 300 seconds?
        if not self.shards:
            self.shards = self.initialize_shards()

        if not self.shards:
            LOG.debug("No shards found for %s.", self.source_arn)
            raise EmptyPollResultsException(service=self.event_source(), source_arn=self.source_arn)
        else:
            LOG.debug("Event source %s has %d shards.", self.source_arn, len(self.shards))
            # Remove all shard batchers without corresponding shards
            for shard_id in self.shard_batcher.keys() - self.shards.keys():
                self.shard_batcher.pop(shard_id, None)

        # TODO: improve efficiency because this currently limits the throughput to at most batch size per poll interval
        # Handle shards round-robin. Re-initialize current shard iterator once all shards are handled.
        if self.iterator_over_shards is None:
            self.iterator_over_shards = iter(self.shards.items())

        current_shard_tuple = next(self.iterator_over_shards, None)
        if not current_shard_tuple:
            self.iterator_over_shards = iter(self.shards.items())
            current_shard_tuple = next(self.iterator_over_shards, None)

        # TODO Better handling when shards are initialised and the iterator returns nothing
        if not current_shard_tuple:
            raise PipeInternalError(
                "Failed to retrieve any shards for stream polling despite initialization."
            )

        try:
            self.poll_events_from_shard(*current_shard_tuple)
        except PipeInternalError:
            # TODO: standardize logging
            # Ignore and wait for the next polling interval, which will do retry
            pass

    def poll_events_from_shard(self, shard_id: str, shard_iterator: str):
        get_records_response = self.get_records(shard_iterator)
        records: list[dict] = get_records_response.get("Records", [])
        next_shard_iterator = get_records_response["NextShardIterator"]

        # We cannot reliably back-off when no records found since an iterator
        # may have to move multiple times until records are returned.
        # See https://docs.aws.amazon.com/streams/latest/dev/troubleshooting-consumers.html#getrecords-returns-empty
        # However, we still need to check if batcher should be triggered due to time-based batching.
        should_flush = self.shard_batcher[shard_id].add(records)
        if not should_flush:
            self.shards[shard_id] = next_shard_iterator
            return

        # Retrieve and drain all events in batcher
        collected_records = self.shard_batcher[shard_id].flush()
        # If there is overflow (i.e 1k BatchSize and 1.2K returned in flush), further split up the batch.
        for batch in batched(collected_records, self.stream_parameters.get("BatchSize")):
            # This could potentially lead to data loss if forward_events_to_target raises an exception after a flush
            # which would otherwise be solved with checkpointing.
            # TODO: Implement checkpointing, leasing, etc. from https://docs.aws.amazon.com/streams/latest/dev/kcl-concepts.html
            self.forward_events_to_target(shard_id, next_shard_iterator, batch)

    def forward_events_to_target(self, shard_id, next_shard_iterator, records):
        polled_events = self.transform_into_events(records, shard_id)

        abort_condition = None
        # Check MaximumRecordAgeInSeconds
        if maximum_record_age_in_seconds := self.stream_parameters.get("MaximumRecordAgeInSeconds"):
            arrival_timestamp_of_last_event = polled_events[-1]["approximateArrivalTimestamp"]
            now = get_current_time().timestamp()
            record_age_in_seconds = now - arrival_timestamp_of_last_event
            if record_age_in_seconds > maximum_record_age_in_seconds:
                abort_condition = "RecordAgeExpired"

        # TODO: implement format detection behavior (e.g., for JSON body):
        #  https://docs.aws.amazon.com/eventbridge/latest/userguide/eb-pipes-event-filtering.html
        #  Check whether we need poller-specific filter-preprocessing here without modifying the actual event!
        # convert to json for filtering (HACK for fixing parity with v1 and getting regression tests passing)
        # localstack.services.lambda_.event_source_listeners.kinesis_event_source_listener.KinesisEventSourceListener._filter_records
        # TODO: explore better abstraction for the entire filtering, including the set_data and get_data remapping
        #  We need better clarify which transformations happen before and after filtering -> fix missing test coverage
        parsed_events = self.pre_filter(polled_events)
        # TODO: advance iterator past matching events!
        #  We need to checkpoint the sequence number for each shard and then advance the shard iterator using
        #  GetShardIterator with a given sequence number
        #  https://docs.aws.amazon.com/kinesis/latest/APIReference/API_GetShardIterator.html
        #  Failing to do so kinda blocks the stream resulting in very high latency.
        matching_events = self.filter_events(parsed_events)
        matching_events_post_filter = self.post_filter(matching_events)

        # TODO: implement MaximumBatchingWindowInSeconds flush condition (before or after filter?)
        # Don't trigger upon empty events
        if len(matching_events_post_filter) == 0:
            # Update shard iterator if no records match the filter
            self.shards[shard_id] = next_shard_iterator
            return
        events = self.add_source_metadata(matching_events_post_filter)
        LOG.debug("Polled %d events from %s in shard %s", len(events), self.source_arn, shard_id)
        # TODO: A retry should probably re-trigger fetching the record from the stream again?!
        #  -> This could be tested by setting a high retry number, using a long pipe execution, and a relatively
        #  short record expiration age at the source. Check what happens if the record expires at the source.
        #  A potential implementation could use checkpointing based on the iterator position (within shard scope)
        # TODO: handle partial batch failure (see poller.py:parse_batch_item_failures)
        # TODO: think about how to avoid starvation of other shards if one shard runs into infinite retries
        attempts = 0
        error_payload = {}

        max_retries = self.stream_parameters.get("MaximumRetryAttempts", -1)
        # NOTE: max_retries == 0 means exponential backoff is disabled
        boff = ExponentialBackoff(max_retries=max_retries)
        while (
            not abort_condition
            and not self.max_retries_exceeded(attempts)
            and not self._is_shutdown.is_set()
        ):
            try:
                if attempts > 0:
                    # TODO: Should we always backoff (with jitter) before processing since we may not want multiple pollers
                    # all starting up and polling simultaneously
                    # For example: 500 persisted ESMs starting up and requesting concurrently could flood gateway
                    self._is_shutdown.wait(boff.next_backoff())

                self.processor.process_events_batch(events)
                boff.reset()

                # Update shard iterator if execution is successful
                self.shards[shard_id] = next_shard_iterator
                return
            except PartialBatchFailureError as ex:
                # TODO: add tests for partial batch failure scenarios
                if (
                    self.stream_parameters.get("OnPartialBatchItemFailure")
                    == OnPartialBatchItemFailureStreams.AUTOMATIC_BISECT
                ):
                    # TODO: implement and test splitting batches in half until batch size 1
                    #  https://docs.aws.amazon.com/eventbridge/latest/pipes-reference/API_PipeSourceKinesisStreamParameters.html
                    LOG.warning(
                        "AUTOMATIC_BISECT upon partial batch item failure is not yet implemented. Retrying the entire batch."
                    )
                error_payload = ex.error

                # Extract all sequence numbers from events in batch. This allows us to fail the whole batch if
                # an unknown itemidentifier is returned.
                batch_sequence_numbers = {
                    self.get_sequence_number(event) for event in matching_events
                }

                # If the batchItemFailures array contains multiple items, Lambda uses the record with the lowest sequence number as the checkpoint.
                # Lambda then retries all records starting from that checkpoint.
                failed_sequence_ids: list[int] | None = get_batch_item_failures(
                    ex.partial_failure_payload, batch_sequence_numbers
                )

                # If None is returned, consider the entire batch a failure.
                if failed_sequence_ids is None:
                    continue

                # This shouldn't be possible since a PartialBatchFailureError was raised
                if len(failed_sequence_ids) == 0:
                    assert failed_sequence_ids, (
                        "Invalid state encountered: PartialBatchFailureError raised but no batch item failures found."
                    )

                lowest_sequence_id: str = min(failed_sequence_ids, key=int)

                # Discard all successful events and re-process from sequence number of failed event
                _, events = self.bisect_events(lowest_sequence_id, events)
            except (BatchFailureError, Exception) as ex:
                if isinstance(ex, BatchFailureError):
                    error_payload = ex.error

                # FIXME partner_resource_arn is not defined in ESM
                LOG.debug(
                    "Attempt %d failed while processing %s with events: %s",
                    attempts,
                    self.partner_resource_arn or self.source_arn,
                    events,
                )
            finally:
                # Retry polling until the record expires at the source
                attempts += 1

        # Send failed events to potential DLQ
        abort_condition = abort_condition or "RetryAttemptsExhausted"
        failure_context = self.processor.generate_event_failure_context(
            abort_condition=abort_condition,
            error=error_payload,
            attempts_count=attempts,
            partner_resource_arn=self.partner_resource_arn,
        )
        self.send_events_to_dlq(shard_id, events, context=failure_context)
        # Update shard iterator if the execution failed but the events are sent to a DLQ
        self.shards[shard_id] = next_shard_iterator

    def get_records(self, shard_iterator: str) -> dict:
        """Returns a GetRecordsOutput from the GetRecords endpoint of streaming services such as Kinesis or DynamoDB"""
        try:
            get_records_response = self.source_client.get_records(
                # TODO: add test for cross-account scenario
                # Differs for Kinesis and DynamoDB but required for cross-account scenario
                **self.stream_arn_param(),
                ShardIterator=shard_iterator,
                Limit=self.stream_parameters["BatchSize"],
            )
            return get_records_response
        # TODO: test iterator expired with conditional error scenario (requires failure destinations)
        except self.source_client.exceptions.ExpiredIteratorException as e:
            LOG.debug(
                "Shard iterator %s expired for stream %s, re-initializing shards",
                shard_iterator,
                self.source_arn,
            )
            # TODO: test TRIM_HORIZON and AT_TIMESTAMP scenarios for this case. We don't want to start from scratch and
            #  might need to think about checkpointing here.
            self.shards = self.initialize_shards()
            raise PipeInternalError from e
        except ClientError as e:
            if "AccessDeniedException" in str(e):
                LOG.warning(
                    "Insufficient permissions to get records from stream %s: %s",
                    self.source_arn,
                    e,
                )
                raise CustomerInvocationError from e
            elif "ResourceNotFoundException" in str(e):
                # FIXME: The 'Invalid ShardId in ShardIterator' error is returned by DynamoDB-local. Unsure when/why this is returned.
                if "Invalid ShardId in ShardIterator" in str(e):
                    LOG.warning(
                        "Invalid ShardId in ShardIterator for %s. Re-initializing shards.",
                        self.source_arn,
                    )
                    self.shards = self.initialize_shards()
                else:
                    LOG.warning(
                        "Source stream %s does not exist: %s",
                        self.source_arn,
                        e,
                    )
                    raise CustomerInvocationError from e
            elif "TrimmedDataAccessException" in str(e):
                LOG.debug(
                    "Attempted to iterate over trimmed record or expired shard iterator %s for stream %s, re-initializing shards",
                    shard_iterator,
                    self.source_arn,
                )
                self.shards = self.initialize_shards()
            else:
                LOG.debug("ClientError during get_records for stream %s: %s", self.source_arn, e)
            raise PipeInternalError from e

    def send_events_to_dlq(self, shard_id, events, context) -> None:
        dlq_arn = self.stream_parameters.get("DeadLetterConfig", {}).get("Arn")
        if dlq_arn:
            failure_timstamp = get_current_time()
            dlq_event = self.create_dlq_event(shard_id, events, context, failure_timstamp)
            # Send DLQ event to DLQ target
            parsed_arn = parse_arn(dlq_arn)
            service = parsed_arn["service"]
            # TODO: use a sender instance here, likely inject via DI into poller (what if it updates?)
            if service == "sqs":
                # TODO: inject and cache SQS client using proper IAM role (supports cross-account operations)
                sqs_client = get_internal_client(dlq_arn)
                # TODO: check if the DLQ exists
                dlq_url = get_queue_url(dlq_arn)
                # TODO: validate no FIFO queue because they are unsupported
                sqs_client.send_message(QueueUrl=dlq_url, MessageBody=json.dumps(dlq_event))
            elif service == "sns":
                sns_client = get_internal_client(dlq_arn)
                sns_client.publish(TopicArn=dlq_arn, Message=json.dumps(dlq_event))
            elif service == "s3":
                s3_client = get_internal_client(dlq_arn)
                dlq_event_with_payload = {
                    **dlq_event,
                    "payload": {
                        "Records": events,
                    },
                }
                s3_client.put_object(
                    Bucket=s3_bucket_name(dlq_arn),
                    Key=get_failure_s3_object_key(self.esm_uuid, shard_id, failure_timstamp),
                    Body=json.dumps(dlq_event_with_payload),
                )
            else:
                LOG.warning("Unsupported DLQ service %s", service)

    def create_dlq_event(
        self, shard_id: str, events: list[dict], context: dict, failure_timestamp: datetime
    ) -> dict:
        first_record = events[0]
        first_record_arrival = get_datetime_from_timestamp(
            self.get_approximate_arrival_time(first_record)
        )

        last_record = events[-1]
        last_record_arrival = get_datetime_from_timestamp(
            self.get_approximate_arrival_time(last_record)
        )
        return {
            **context,
            self.failure_payload_details_field_name(): {
                "approximateArrivalOfFirstRecord": self.format_datetime(first_record_arrival),
                "approximateArrivalOfLastRecord": self.format_datetime(last_record_arrival),
                "batchSize": len(events),
                "endSequenceNumber": self.get_sequence_number(last_record),
                "shardId": shard_id,
                "startSequenceNumber": self.get_sequence_number(first_record),
                "streamArn": self.source_arn,
            },
            "timestamp": failure_timestamp.isoformat(timespec="milliseconds").replace(
                "+00:00", "Z"
            ),
            "version": "1.0",
        }

    def max_retries_exceeded(self, attempts: int) -> bool:
        maximum_retry_attempts = self.stream_parameters.get("MaximumRetryAttempts", -1)
        # Infinite retries until the source expires
        if maximum_retry_attempts == -1:
            return False
        return attempts > maximum_retry_attempts

    def bisect_events(
        self, sequence_number: str, events: list[dict]
    ) -> tuple[list[dict], list[dict]]:
        """Splits list of events in two, where a sequence number equals a passed parameter `sequence_number`.
        This is used for:
          - `ReportBatchItemFailures`: Discarding events in a batch following a failure when is set.
          - `BisectBatchOnFunctionError`: Used to split a failed batch in two when doing a retry (not implemented)."""
        for i, event in enumerate(events):
            if self.get_sequence_number(event) == sequence_number:
                return events[:i], events[i:]

        return events, []


def get_failure_s3_object_key(esm_uuid: str, shard_id: str, failure_datetime: datetime) -> str:
    """
    From https://docs.aws.amazon.com/lambda/latest/dg/kinesis-on-failure-destination.html:

    The S3 object containing the invocation record uses the following naming convention:
    aws/lambda/<ESM-UUID>/<shardID>/YYYY/MM/DD/YYYY-MM-DDTHH.MM.SS-<Random UUID>

    :return: Key for s3 object that invocation failure record will be put to
    """
    timestamp = failure_datetime.strftime("%Y-%m-%dT%H.%M.%S")
    year_month_day = failure_datetime.strftime("%Y/%m/%d")
    random_uuid = long_uid()
    return f"aws/lambda/{esm_uuid}/{shard_id}/{year_month_day}/{timestamp}-{random_uuid}"

localstack / localstack / f32ddf78-b8ca-49ab-9751-2526b355995f

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous