Lambda Functions¶

Serverless functions for TradAI platform orchestration, monitoring, and automation.

Overview¶

TradAI uses AWS Lambda for event-driven processing, scheduled tasks, and Step Functions workflow integration. All Lambdas use the lambda_handler decorator from tradai-common for consistent error handling and metrics.

from tradai.common.lambda_ import lambda_handler, LambdaResponse

@lambda_handler
def handler(event: dict, context) -> dict:
    # Business logic here
    return LambdaResponse.success({"result": "data"})

Lambda Categories¶

Workflow Orchestration¶

Lambda	Trigger	Purpose
backtest-consumer	SQS	Launches ECS backtesting tasks from queue
sqs-consumer	SQS	Launches ECS retraining tasks from queue
cleanup-resources	Step Functions	Cleans up orphaned ECS tasks on workflow failure

Model Lifecycle¶

Lambda	Trigger	Purpose
check-retraining-needed	Step Functions	Evaluates if model needs retraining
compare-models	Step Functions	Compares champion vs challenger models
promote-model	Step Functions	Promotes model to Production in MLflow
model-rollback	CloudWatch Alarm	Rolls back model to previous version
retraining-scheduler	EventBridge	Schedules and triggers model retraining

Monitoring & Health¶

Lambda	Trigger	Purpose
health-check	EventBridge	Checks ECS service health via Service Discovery
trading-heartbeat-check	EventBridge	Monitors live trading container heartbeats
drift-monitor	EventBridge	Detects model drift using PSI metrics
orphan-scanner	EventBridge	Finds and stops orphaned ECS tasks

Deployment & Validation¶

Lambda	Trigger	Purpose
validate-strategy	Step Functions	Validates strategy config before deployment
notify-completion	Step Functions	Sends SNS/Slack notifications

Data Services¶

Lambda	Trigger	Purpose
data-collection-proxy	API/Step Functions	Proxy to data-collection service

Common Patterns¶

Environment Variables¶

All Lambdas use these common environment variables:

Variable	Description
`ENVIRONMENT`	Environment name (dev/staging/prod)
`DYNAMODB_TABLE_NAME`	Default state repository table
`ALERT_SNS_TOPIC_ARN`	SNS topic for alerts
`ECS_CLUSTER`	ECS cluster name/ARN

Response Format¶

All Lambdas return LambdaResponse format:

# Success
{
    "statusCode": 200,
    "body": {"result": "data"}
}

# Error
{
    "statusCode": 400,
    "body": {"error": "Error message"}
}

Step Functions Integration¶

Lambdas in Step Functions workflows return data directly (not wrapped in statusCode):

# For Step Functions
return LambdaResponse.success({"decision": "PROMOTE"})
# Returns: {"decision": "PROMOTE"}

SQS Batch Processing¶

SQS-triggered Lambdas support partial batch failure:

return {
    "batchItemFailures": [
        {"itemIdentifier": "failed-message-id"}
    ]
}

Deployment¶

Lambdas are deployed via Pulumi in infra/modules/lambda_funcs.py:

# Example: Creating a Lambda function
lambda_func = aws.lambda_.Function(
    f"{name}-lambda",
    runtime="python3.11",
    handler="handler.handler",
    timeout=300,
    memory_size=256,
    environment=aws.lambda_.FunctionEnvironmentArgs(
        variables={
            "ENVIRONMENT": environment,
            "DYNAMODB_TABLE_NAME": state_table.name,
        }
    ),
)

IAM Permissions¶

Each Lambda has specific IAM permissions defined in infra/modules/iam.py. Common patterns:

ECS Lambdas: ecs:RunTask, ecs:DescribeTasks, ecs:ListTasks, ecs:StopTask
DynamoDB Lambdas: dynamodb:GetItem, dynamodb:PutItem, dynamodb:Query, dynamodb:Scan
SNS Lambdas: sns:Publish
CloudWatch Lambdas: cloudwatch:PutMetricData

Monitoring¶

CloudWatch Metrics¶

All Lambdas publish custom metrics to CloudWatch namespace TradAI/Lambda:

Metric	Description
`Invocations`	Total invocations
`Errors`	Error count
`Duration`	Execution time (ms)
`ConcurrentExecutions`	Concurrent executions

CloudWatch Alarms¶

Critical Lambdas have CloudWatch alarms configured:

# Example alarm for health-check failures
aws.cloudwatch.MetricAlarm(
    "health-check-alarm",
    comparison_operator="GreaterThanThreshold",
    evaluation_periods=3,
    metric_name="Errors",
    namespace="AWS/Lambda",
    period=300,
    statistic="Sum",
    threshold=1,
)