Batch Processing - Powertools for AWS Lambda (Python) (original) (raw)

The batch processing utility handles partial failures when processing batches from Amazon SQS, Amazon Kinesis Data Streams, and Amazon DynamoDB Streams.

stateDiagram-v2
    direction LR
    BatchSource: Amazon SQS <br/><br/> Amazon Kinesis Data Streams <br/><br/> Amazon DynamoDB Streams <br/><br/>
    LambdaInit: Lambda invocation
    BatchProcessor: Batch Processor
    RecordHandler: Record Handler function
    YourLogic: Your logic to process each batch item
    LambdaResponse: Lambda response

    BatchSource --> LambdaInit

    LambdaInit --> BatchProcessor
    BatchProcessor --> RecordHandler

    state BatchProcessor {
        [*] --> RecordHandler: Your function
        RecordHandler --> YourLogic
    }

    RecordHandler --> BatchProcessor: Collect results
    BatchProcessor --> LambdaResponse: Report items that failed processing

Key features¶

• Reports batch item failures to reduce number of retries for a record upon errors
• Simple interface to process each batch record
• Integrates with Event Source Data Classes and Parser (Pydantic) for self-documenting record schema
• Build your own batch processor by extending primitives

Background¶

When using SQS, Kinesis Data Streams, or DynamoDB Streams as a Lambda event source, your Lambda functions are triggered with a batch of messages.

If your function fails to process any message from the batch, the entire batch returns to your queue or stream. This same batch is then retried until either condition happens first: a) your Lambda function returns a successful response, b) record reaches maximum retry attempts, or c) records expire.

journey
  section Conditions
    Successful response: 5: Success
    Maximum retries: 3: Failure
    Records expired: 1: Failure

This behavior changes when you enable Report Batch Item Failures feature in your Lambda function event source configuration:

• SQS queues. Only messages reported as failure will return to the queue for a retry, while successful ones will be deleted.
• Kinesis data streams and DynamoDB streams. Single reported failure will use its sequence number as the stream checkpoint. Multiple reported failures will use the lowest sequence number as checkpoint. Warning: This utility lowers the chance of processing records more than once; it does not guarantee it

We recommend implementing processing logic in an idempotent manner wherever possible.

You can find more details on how Lambda works with either SQS, Kinesis, or DynamoDB in the AWS Documentation.

Getting started¶

For this feature to work, you need to (1) configure your Lambda function event source to use ReportBatchItemFailures, and (2) return a specific response to report which records failed to be processed.

You use your preferred deployment framework to set the correct configuration while this utility handles the correct response to be returned.

Required resources¶

The remaining sections of the documentation will rely on these samples. For completeness, this demonstrates IAM permissions and Dead Letter Queue where batch records will be sent after 2 retries were attempted.

You do not need any additional IAM permissions to use this utility, except for what each event source requires.

SQSKinesis Data StreamsDynamoDB Streams

Processing messages from SQS¶

Processing batches from SQS works in three stages:

1. Instantiate BatchProcessor and choose EventType.SQS for the event type
2. Define your function to handle each batch record, and use SQSRecord type annotation for autocompletion
3. Use process_partial_response to kick off processing

This code example uses Tracer and Logger for completion.

RecommendedAs a context managerSample responseSample event

1. Step 1. Creates a partial failure batch processor for SQS queues. See partial failure mechanics for details
2. Step 2. Defines a function to receive one record at a time from the batch
3. Step 3. Kicks off processing

The second record failed to be processed, therefore the processor added its message ID in the response.

FIFO queues¶

When working with SQS FIFO queues, a batch may include messages from different group IDs.

By default, we will stop processing at the first failure and mark unprocessed messages as failed to preserve ordering. However, this behavior may not be optimal for customers who wish to proceed with processing messages from a different group ID.

Enable the skip_group_on_error option for seamless processing of messages from various group IDs. This setup ensures that messages from a failed group ID are sent back to SQS, enabling uninterrupted processing of messages from the subsequent group ID.

RecommendedAs a context managerEnabling skip_group_on_error flag

1. Step 1. Creates a partial failure batch processor for SQS FIFO queues. See partial failure mechanics for details

Processing messages from Kinesis¶

Processing batches from Kinesis works in three stages:

1. Instantiate BatchProcessor and choose EventType.KinesisDataStreams for the event type
2. Define your function to handle each batch record, and use KinesisStreamRecord type annotation for autocompletion
3. Use process_partial_response to kick off processing

This code example uses Tracer and Logger for completion.

RecommendedAs a context managerSample responseSample event

1. Step 1. Creates a partial failure batch processor for Kinesis Data Streams. See partial failure mechanics for details

The second record failed to be processed, therefore the processor added its sequence number in the response.

Processing messages from DynamoDB¶

Processing batches from DynamoDB Streams works in three stages:

1. Instantiate BatchProcessor and choose EventType.DynamoDBStreams for the event type
2. Define your function to handle each batch record, and use DynamoDBRecord type annotation for autocompletion
3. Use process_partial_response to kick off processing

This code example uses Tracer and Logger for completion.

RecommendedAs a context managerSample responseSample event

1. Step 1. Creates a partial failure batch processor for DynamoDB Streams. See partial failure mechanics for details

The second record failed to be processed, therefore the processor added its sequence number in the response.

Error handling¶

By default, we catch any exception raised by your record handler function. This allows us to (1) continue processing the batch, (2) collect each batch item that failed processing, and (3) return the appropriate response correctly without failing your Lambda function execution.

Sample error handling with custom exceptionSample response

1. Any exception works here. See extending BatchProcessor section, if you want to override this behavior.
2. Exceptions raised in record_handler will propagate to process_partial_response.
   We catch them and include each failed batch item identifier in the response dictionary (see Sample response tab).

Partial failure mechanics¶

All batch items will be passed to the record handler for processing, even if exceptions are thrown - Here's the behavior after completing the batch:

• All records successfully processed. We will return an empty list of item failures {'batchItemFailures': []}
• Partial success with some exceptions. We will return a list of all item IDs/sequence numbers that failed processing
• All records failed to be processed. We will raise BatchProcessingError exception with a list of all exceptions raised when processing

The following sequence diagrams explain how each Batch processor behaves under different scenarios.

SQS Standard¶

Read more about Batch Failure Reporting feature in AWS Lambda.

Sequence diagram to explain how BatchProcessor works with SQS Standard queues.

sequenceDiagram
    autonumber
    participant SQS queue
    participant Lambda service
    participant Lambda function
    Lambda service->>SQS queue: Poll
    Lambda service->>Lambda function: Invoke (batch event)
    Lambda function->>Lambda service: Report some failed messages
    activate SQS queue
    Lambda service->>SQS queue: Delete successful messages
    SQS queue-->>SQS queue: Failed messages return
    Note over SQS queue,Lambda service: Process repeat
    deactivate SQS queue

SQS mechanism with Batch Item Failures

SQS FIFO¶

Read more about Batch Failure Reporting feature in AWS Lambda.

Sequence diagram to explain how SqsFifoPartialProcessor works with SQS FIFO queues without skip_group_on_error flag.

sequenceDiagram
    autonumber
    participant SQS queue
    participant Lambda service
    participant Lambda function
    Lambda service->>SQS queue: Poll
    Lambda service->>Lambda function: Invoke (batch event)
    activate Lambda function
    Lambda function-->Lambda function: Process 2 out of 10 batch items
    Lambda function--xLambda function: Fail on 3rd batch item
    Lambda function->>Lambda service: Report 3rd batch item and unprocessed messages as failure
    deactivate Lambda function
    activate SQS queue
    Lambda service->>SQS queue: Delete successful messages (1-2)
    SQS queue-->>SQS queue: Failed messages return (3-10)
    deactivate SQS queue

SQS FIFO mechanism with Batch Item Failures

Sequence diagram to explain how SqsFifoPartialProcessor works with SQS FIFO queues with skip_group_on_error flag.

sequenceDiagram
    autonumber
    participant SQS queue
    participant Lambda service
    participant Lambda function
    Lambda service->>SQS queue: Poll
    Lambda service->>Lambda function: Invoke (batch event)
    activate Lambda function
    Lambda function-->Lambda function: Process 2 out of 10 batch items
    Lambda function--xLambda function: Fail on 3rd batch item
    Lambda function-->Lambda function: Process messages from another MessageGroupID
    Lambda function->>Lambda service: Report 3rd batch item and all messages within the same MessageGroupID as failure
    deactivate Lambda function
    activate SQS queue
    Lambda service->>SQS queue: Delete successful messages processed
    SQS queue-->>SQS queue: Failed messages return
    deactivate SQS queue

SQS FIFO mechanism with Batch Item Failures

Kinesis and DynamoDB Streams¶

Read more about Batch Failure Reporting feature.

Sequence diagram to explain how BatchProcessor works with both Kinesis Data Streams and DynamoDB Streams.

For brevity, we will use Streams to refer to either services. For theory on stream checkpoints, see this blog post

sequenceDiagram
    autonumber
    participant Streams
    participant Lambda service
    participant Lambda function
    Lambda service->>Streams: Poll latest records
    Lambda service->>Lambda function: Invoke (batch event)
    activate Lambda function
    Lambda function-->Lambda function: Process 2 out of 10 batch items
    Lambda function--xLambda function: Fail on 3rd batch item
    Lambda function-->Lambda function: Continue processing batch items (4-10)
    Lambda function->>Lambda service: Report batch item as failure (3)
    deactivate Lambda function
    activate Streams
    Lambda service->>Streams: Checkpoints to sequence number from 3rd batch item
    Lambda service->>Streams: Poll records starting from updated checkpoint
    deactivate Streams

Kinesis and DynamoDB streams mechanism with single batch item failure

The behavior changes slightly when there are multiple item failures. Stream checkpoint is updated to the lowest sequence number reported.

Note that the batch item sequence number could be different from batch item number in the illustration.

sequenceDiagram
    autonumber
    participant Streams
    participant Lambda service
    participant Lambda function
    Lambda service->>Streams: Poll latest records
    Lambda service->>Lambda function: Invoke (batch event)
    activate Lambda function
    Lambda function-->Lambda function: Process 2 out of 10 batch items
    Lambda function--xLambda function: Fail on 3-5 batch items
    Lambda function-->Lambda function: Continue processing batch items (6-10)
    Lambda function->>Lambda service: Report batch items as failure (3-5)
    deactivate Lambda function
    activate Streams
    Lambda service->>Streams: Checkpoints to lowest sequence number
    Lambda service->>Streams: Poll records starting from updated checkpoint
    deactivate Streams

Kinesis and DynamoDB streams mechanism with multiple batch item failures

Processing messages asynchronously¶

New to AsyncIO? Read this comprehensive guide first.

You can use AsyncBatchProcessor class and async_process_partial_response function to process messages concurrently.

When is this useful?

Your use case might be able to process multiple records at the same time without conflicting with one another.

For example, imagine you need to process multiple loyalty points and incrementally save them in the database. While you await the database to confirm your records are saved, you could start processing another request concurrently.

The reason this is not the default behaviour is that not all use cases can handle concurrency safely (e.g., loyalty points must be updated in order).

Using tracer?

AsyncBatchProcessor uses asyncio.gather. This might cause side effects and reach trace limits at high concurrency.

Advanced¶

Pydantic integration¶

You can bring your own Pydantic models via model parameter when inheriting from SqsRecordModel, KinesisDataStreamRecord, or DynamoDBStreamRecordModel

Inheritance is importance because we need to access message IDs and sequence numbers from these records in the event of failure. Mypy is fully integrated with this utility, so it should identify whether you're passing the incorrect Model.

SQSSQS - Sample Event Kinesis Data StreamsKinesis - Sample Event DynamoDB StreamsDynamoDB - Sample Event

Working with full batch failures¶

By default, the BatchProcessor will raise BatchProcessingError if all records in the batch fail to process, we do this to reflect the failure in your operational metrics.

When working with functions that handle batches with a small number of records, or when you use errors as a flow control mechanism, this behavior might not be desirable as your function might generate an unnaturally high number of errors. When this happens, the Lambda service will scale down the concurrency of your function, potentially impacting performance.

For these scenarios, you can set the raise_on_entire_batch_failure option to False.

working_with_entire_batch_fail.py

Accessing processed messages¶

Use the context manager to access a list of all returned values from your record_handler function.

• When successful. We include a tuple with 1/ success, 2/ the result of record_handler, and 3/ the batch item
• When failed. We include a tuple with 1/ fail, 2/ exception as a string, and 3/ the batch item serialized as Event Source Data Class or Pydantic model.

If a Pydantic model fails validation early, we serialize its failure record as Event Source Data Class to be able to collect message ID/sequence numbers etc.

Accessing raw processed messagesSample processed messagesSample processed messages (Pydantic)

1. Context manager requires the records list. This is typically handled by process_partial_response.
2. Cause contains exception str if failed, or success otherwise.

1. Sample exception could have raised from within record_handler function.

1. Sample when a model fails validation early.
   Batch item (3rd item) is serialized to the respective Event Source Data Class event type.
2. Sample when model validated successfully but another exception was raised during processing.

Accessing Lambda Context¶

Within your record_handler function, you might need access to the Lambda context to determine how much time you have left before your function times out.

We can automatically inject the Lambda context into your record_handler if your function signature has a parameter named lambda_context. When using a context manager, you also need to pass the Lambda context object like in the example below.

RecommendedAs a context manager

Extending BatchProcessor¶

You might want to bring custom logic to the existing BatchProcessor to slightly override how we handle successes and failures.

For these scenarios, you can subclass BatchProcessor and quickly override success_handler and failure_handler methods:

• success_handler() is called for each successfully processed record
• failure_handler() is called for each failed record Note

These functions have a common record argument. For backward compatibility reasons, their type is not the same:

• success_handler: record type is dict[str, Any], the raw record data.
• failure_handler: record type can be an Event Source Data Class or your Pydantic model. During Pydantic validation errors, we fall back and serialize record to Event Source Data Class to not break the processing pipeline.

Let's suppose you'd like to add metrics to track successes and failures of your batch records.

Create your own partial processor¶

You can create your own partial batch processor from scratch by inheriting the BasePartialProcessor class, and implementing _prepare(), _clean(), _process_record() and _async_process_record().

classDiagram
    direction LR
    class BasePartialProcessor {
        <<interface>>
        +_prepare()
        +_clean()
        +_process_record_(record: Dict)
        +_async_process_record_()
    }

    class YourCustomProcessor {
        +_prepare()
        +_clean()
        +_process_record_(record: Dict)
        +_async_process_record_()
    }

    BasePartialProcessor <|-- YourCustomProcessor : implement

Visual representation to bring your own processor

• _process_record() – handles all processing logic for each individual message of a batch, including calling the record_handler (self.handler)
• _prepare() – called once as part of the processor initialization
• _clean() – teardown logic called once after _process_record completes
• _async_process_record() – If you need to implement asynchronous logic, use this method, otherwise define it in your class with empty logic
• response() - called upon completion of processing

You can utilize this class to instantiate a new processor and then pass it to the process_partial_response function.

Creating a custom batch processorDynamoDB table used for storing processed records.Sample event

Caveats¶

Tracer response auto-capture for large batch sizes¶

When using Tracer to capture responses for each batch record processing, you might exceed 64K of tracing data depending on what you return from your record_handler function, or how big is your batch size.

If that's the case, you can configure Tracer to disable response auto-capturing.

Testing your code¶

As there is no external calls, you can unit test your code with BatchProcessor quite easily.

Example:

Given a SQS batch where the first batch record succeeds and the second fails processing, we should have a single item reported in the function response.

getting_started_with_test.pygetting_started_with_test_app.pySample SQS event

FAQ¶

Choosing between method and context manager¶

Use context manager when you want access to the processed messages or handle BatchProcessingError exception when all records within the batch fail to be processed.

Integrating exception handling with Sentry.io¶

When using Sentry.io for error monitoring, you can override failure_handler to capture each processing exception with Sentry SDK:

Credits to Charles-Axel Dein