2024 - Best practices for serverless developers (SVS401)

Service-full serverless (Compose, configure, then code)

Orchestration	Choreography
A central coordination controlling services	Services collaboration
Workflows (AWS Step functions)	Events (EventBridge)

Step functions: standard vs. Express workflows

Standard	Express
Max 1 year	Max 5 min
Async	Sync & Async
Exactly once	At least once
	High throughput
	Cost effective
Pay per state transition	Pay per number of executions + GBs
More expensive	Cheaper

Also, a nested workflow (using both) can take both advantages.

Serverless compute

AWS Lambda invocation model

3 types of Lambda integrations:

• Synchronous: including request streaming
  e.g. client <-> API gateway <-> Lambda
• Asynchronous
  e.g. EventBridge <-> Lambda service queue <-> Lambda
• Poll-based: event source mapping
  e.g. kinesis <-polls-> Lambda <-> Lambda

Lambda function memory power

• Lambda memory can be between 128 MB to 10 GB.
• Memory allocates CPU power & network bandwidth.
• If code is bounded by memory, CPU, or network, adding memory may improve performance and reduce cost.

Lambda life cycle

INIT	INVOKE	INVOKE	INVOKE	SHUTDOWN
Cold Start	Warm Start

• INIT: (from 2025, it costs money)
  1. Extension init
     • create new execution environment
     • download code / image
  2. runtime init
  3. function init (start to be optimized by ourself)
• INVOKE
• SHUTDOWN:
  1. Runtime shutdown
  2. extension shutdown

Function init optimization

• Don’t load it until you need
  • Optimize dependencies, SDKs and other libraries
  • Minify / Uglify production code
  • Reduce deployment package size
  • Avoid monolithic functions
• Lazy initialize shared libraries
• Establishing connections
  • Handle reconnections in handler
  • Keep-alive in SDKs
• State during environment reuse
  • Keep data for subsequent invocations
• Use provisioned concurrency or SnapStart

Lambda Snapshot

• 在Cold start 與Warm start 中間，指定某版本（無法指定隨時為最新的）製作成snapshot
• 10x faster startup performance
• no additional cost

Concurrency

多一個concurrency 需要再啟動一次cold start，也可設定concurrency requests limit（throttled）

Provisioned concurrency

• 預訂數量個cold start，隨時可以進入warm start。
• Can be tracked through ProvisionedConcurrencyUtilzation
• 指定某版本（無法指定隨時為最新的）製作成snapshot

Scaling quotas

Multiple quotas can be set:

• Account concurrency quota (一個帳號所有function的quota）
• Scale concurrency quota（一個function的quota）

Synchronous invocations: Best practices

• Load test
  • Raise concurrency quota before burst
• Set appropriate timeouts
  • API gateway default is 29 seconds
• Retry with backoff
  • Client retry, use exponential backoff + jitter to not overload system
• Implement idempotency
  • Expect and account for retrying operations

Poll-based (stream) invocations: Best practices

• Processing
  • Use filtering to avoid processing message
  • Evenly distribute records using partition key
  • Poison message can stop processing, gracefully handle errors
• Performance
  • Increase memory/optimize code/increase batch size
  • Add partitions/shards
  • Kinesis: increase parallelization factor/enhanced fan-out
• Monitoring
  • Enhanced monitoring: IteratorAge, consumer_lag

AWS Fargate

A AWS managed ECS. Pay a constant rate per second for the task CPU and memory (not concurrency requests) → Lower requests per number is relatively cost.

Consideration between Functions & Containers

Condition	Functions (Lambda)	Containers (ECS / Fargates)
When you want	Fewer decision and fast. Faster prototype to production	Control of compute environment Tooling consistency
Architecture	Managed triggers, runtimes, scaling	Custom code, services, management
Resources	Memory of 128 MB to 10 GB + vCPU / Net. Max 15 min / single concurrency	Memory of 512MB - 120 GB, 0.25 -16 vCPU Unlimited run time / multi concurrency

Also, we can use both inside step functions

Production-ready serverless service

Retries & timeouts in the record handler

Set up configuration when connecting to services

from boto3 import client
from botocore.config import Config

config = Config(
    retries={
        'max_attempts': 5, 
        'mode': 'adaptive',
    },
    read_timeout=30,
    connect_timeout=10,
)

s3_client = client('s3', config=config)

CloudWatch loggin tips

• Add correlation id/custom id to all logs
• Reduce cost:
  • set INFO log level as default log level
  • Log only what matters most, otherwise, log as DEBUG level
  • Set retention policies on log groups (14 days)

Upgrade your runtime

Python 3.10 to 3.13 improves performance

---

Ref:

https://www.youtube.com/watch?v=5wokwEtddtc