Serverless warming for latency reduction

Serverless Warming Implementation for Latency Reduction

Cold start—main problem in serverless functions for latency-sensitive apps. First invoke after idle period takes 200ms-2s depending on runtime, package size, and config. For API handling real user requests, unacceptable. Warming solves problem by keeping functions "warm".

Cold Start Nature

What happens at cold start:

1. Cloud finds available container/VM
2. Loads function image
3. Initializes runtime (Node.js, Python, JVM)
4. Executes initialization code (outside handler)
5. Executes handler

Steps 1-4 are overhead. Steps 1-3 provider controls, steps 4-5 developer controls.

Typical cold start times:

• Python 3.12 (AWS Lambda, 256MB)—200-400ms
• Node.js 20—100-300ms
• Java 17—800ms-2s (JVM startup)
• Go—50-150ms

Scheduled Warming

Simplest approach: invoke function every 5 minutes via CloudWatch Events / EventBridge to keep it warm.

# lambda_warmer.py—ping function
import json

def handler(event, context):
    if event.get('source') == 'warming':
        # This is ping from warmers, not real request
        return {'statusCode': 200, 'body': json.dumps({'warm': True})}

    # Real function logic
    return process_request(event)

# Terraform: CloudWatch rule for warming
resource "aws_cloudwatch_event_rule" "warmer" {
  name                = "lambda-warmer"
  schedule_expression = "rate(5 minutes)"
}

resource "aws_cloudwatch_event_target" "warmer" {
  rule  = aws_cloudwatch_event_rule.warmer.name
  arn   = aws_lambda_function.api.arn
  input = jsonencode({"source": "warming"})
}

Limitation: each EventBridge trigger spawns only one concurrent instance. For multiple desired warm instances need N parallel invokes.

Warming Multiple Parallel Instances

import boto3
import asyncio

lambda_client = boto3.client('lambda')

async def warm_instance(function_name: str, instance_num: int):
    lambda_client.invoke(
        FunctionName=function_name,
        InvocationType='RequestResponse',
        Payload=json.dumps({
            'source': 'warming',
            'instance': instance_num,
            'sleep': 10  # Keep instance busy 10 seconds
        })
    )

async def warm_function(function_name: str, concurrent_count: int = 5):
    """Spawn N parallel warmup invokes"""
    tasks = [warm_instance(function_name, i) for i in range(concurrent_count)]
    await asyncio.gather(*tasks)

While one invoke keeps instance busy (sleep 10s), Lambda creates new container for next parallel invoke. Result: 5 warm instances.

AWS Lambda Provisioned Concurrency

Official AWS solution: reserve initialized instances. Costs extra but guarantees P99 latency without cold start.

resource "aws_lambda_provisioned_concurrency_config" "api" {
  function_name                  = aws_lambda_function.api.function_name
  qualifier                      = aws_lambda_alias.live.name
  provisioned_concurrent_executions = 5
}

Auto Scaling Provisioned Concurrency—scale provisioning by schedule (more morning, less night):

resource "aws_appautoscaling_target" "lambda_pc" {
  max_capacity       = 20
  min_capacity       = 2
  resource_id        = "function:${aws_lambda_function.api.function_name}:live"
  scalable_dimension = "lambda:function:ProvisionedConcurrency"
  service_namespace  = "lambda"
}

resource "aws_appautoscaling_policy" "lambda_pc_tracking" {
  policy_type        = "TargetTrackingScaling"
  resource_id        = aws_appautoscaling_target.lambda_pc.resource_id
  scalable_dimension = aws_appautoscaling_target.lambda_pc.scalable_dimension
  service_namespace  = aws_appautoscaling_target.lambda_pc.service_namespace

  target_tracking_scaling_policy_configuration {
    target_value = 0.7  # 70% utilization of provisioning
    predefined_metric_specification {
      predefined_metric_type = "LambdaProvisionedConcurrencyUtilization"
    }
  }
}

Optimization of Initialization Code

Warming helps but reducing cold start itself—better strategy:

# BAD: create clients inside handler
def handler(event, context):
    dynamodb = boto3.resource('dynamodb')  # Every cold start
    db_client = psycopg2.connect(DSN)      # Creates connection
    ...

# GOOD: create clients at module level (once)
import boto3
import psycopg2

dynamodb = boto3.resource('dynamodb')  # Init at cold start
_connection = None  # Lazy connection pool

def get_connection():
    global _connection
    if _connection is None or _connection.closed:
        _connection = psycopg2.connect(DSN)
    return _connection

def handler(event, context):
    conn = get_connection()  # Reuses existing connection
    ...

Lambda SnapStart (Java)

AWS Lambda SnapStart for Java: creates snapshot of initialized function state. Java cold start reduced from 1-2s to 100-200ms.

resource "aws_lambda_function" "java_api" {
  ...
  snap_start {
    apply_on = "PublishedVersions"
  }
}

Timeline

• Scheduled warming (EventBridge)—0.5 day
• Parallel warming script—1 day
• Provisioned Concurrency + Auto Scaling—1-2 days
• Initialization code optimization—1-3 days (depends on codebase)