Implementing stateful Kinesis Data Streams processing in Lambda

Lambda functions can run continuous stream processing applications. A stream represents unbounded data that flows continuously through your application. To analyze information from this continuously updating input, you can bound the included records using a window defined in terms of time.

Tumbling windows are distinct time windows that open and close at regular intervals. By default, Lambda invocations are stateless—you cannot use them for processing data across multiple continuous invocations without an external database. However, with tumbling windows, you can maintain your state across invocations. This state contains the aggregate result of the messages previously processed for the current window. Your state can be a maximum of 1 MB per shard. If it exceeds that size, Lambda terminates the window early.

Each record in a stream belongs to a specific window. Lambda will process each record at least once, but doesn't guarantee that each record will be processed only once. In rare cases, such as error handling, some records might be processed more than once. Records are always processed in order the first time. If records are processed more than once, they might be processed out of order.

Aggregation and processing

Your user managed function is invoked both for aggregation and for processing the final results of that aggregation. Lambda aggregates all records received in the window. You can receive these records in multiple batches, each as a separate invocation. Each invocation receives a state. Thus, when using tumbling windows, your Lambda function response must contain a state property. If the response does not contain a state property, Lambda considers this a failed invocation. To satisfy this condition, your function can return a TimeWindowEventResponse object, which has the following JSON shape:

{
    "state": {
        "1": 282,
        "2": 715
    },
    "batchItemFailures": []
}

At the end of the window, the flag isFinalInvokeForWindow is set to true to indicate that this is the final state and that it’s ready for processing. After processing, the window completes and your final invocation completes, and then the state is dropped.

At the end of your window, Lambda uses final processing for actions on the aggregation results. Your final processing is synchronously invoked. After successful invocation, your function checkpoints the sequence number and stream processing continues. If invocation is unsuccessful, your Lambda function suspends further processing until a successful invocation.

{
    "Records": [
        {
            "kinesis": {
                "kinesisSchemaVersion": "1.0",
                "partitionKey": "1",
                "sequenceNumber": "49590338271490256608559692538361571095921575989136588898",
                "data": "SGVsbG8sIHRoaXMgaXMgYSB0ZXN0Lg==",
                "approximateArrivalTimestamp": 1607497475.000
            },
            "eventSource": "aws:kinesis",
            "eventVersion": "1.0",
            "eventID": "shardId-000000000006:49590338271490256608559692538361571095921575989136588898",
            "eventName": "aws:kinesis:record",
            "invokeIdentityArn": "arn:aws:iam::123456789012:role/lambda-kinesis-role",
            "awsRegion": "us-east-1",
            "eventSourceARN": "arn:aws:kinesis:us-east-1:123456789012:stream/lambda-stream"
        }
    ],
    "window": {
        "start": "2020-12-09T07:04:00Z",
        "end": "2020-12-09T07:06:00Z"
    },
    "state": {
        "1": 282,
        "2": 715
    },
    "shardId": "shardId-000000000006",
    "eventSourceARN": "arn:aws:kinesis:us-east-1:123456789012:stream/lambda-stream",
    "isFinalInvokeForWindow": false,
    "isWindowTerminatedEarly": false
}

Configuration

You can configure tumbling windows when you create or update an event source mapping. To configure a tumbling window, specify the window in seconds (TumblingWindowInSeconds). The following example AWS Command Line Interface (AWS CLI) command creates a streaming event source mapping that has a tumbling window of 120 seconds. The Lambda function defined for aggregation and processing is named tumbling-window-example-function.

aws lambda create-event-source-mapping \
--event-source-arn arn:aws:kinesis:us-east-1:123456789012:stream/lambda-stream \
--function-name tumbling-window-example-function \
--starting-position TRIM_HORIZON \
--tumbling-window-in-seconds 120

Lambda determines tumbling window boundaries based on the time when records were inserted into the stream. All records have an approximate timestamp available that Lambda uses in boundary determinations.

Tumbling window aggregations do not support resharding. When a shard ends, Lambda considers the current window to be closed, and any child shards will start their own window in a fresh state. When no new records are being added to the current window, Lambda waits for up to 2 minutes before assuming that the window is over. This helps ensure that the function reads all records in the current window, even if the records are added intermittently.

Tumbling windows fully support the existing retry policies maxRetryAttempts and maxRecordAge.

def lambda_handler(event, context):
    print('Incoming event: ', event)
    print('Incoming state: ', event['state'])

#Check if this is the end of the window to either aggregate or process.
    if event['isFinalInvokeForWindow']:
        # logic to handle final state of the window
        print('Destination invoke')
    else:
        print('Aggregate invoke')

#Check for early terminations
    if event['isWindowTerminatedEarly']:
        print('Window terminated early')

    #Aggregation logic
    state = event['state']
    for record in event['Records']:
        state[record['kinesis']['partitionKey']] = state.get(record['kinesis']['partitionKey'], 0) + 1

    print('Returning state: ', state)
    return {'state': state}