Serverless Computing

Someone Else's Servers

Tusa Viktor @ Xively

@talien

github.com/talien

Computational Unit

• Different aspects
  • Business: What you pay for
    • Unit: $
  • Developer: What your program use
    • Unit: CPU/Memory/Time
  • Ops: What you manage
    • Number of servers/databases/log lines/metrics

History

• Mainframes
  • Time-shared usage
  • Not easy to extend capacity
• PCs/Data Centers
  • Roles are separated on different servers
  • Monolith applications
  • Easier to extend capacity
• Virtualization/Containerization
  • Fine-grained capacity on different apps
  • Still need to know the whole capacity
• Public Cloud/IaaS/PaaS
  • "Capacity is not my problem anymore"
  • Computational unit is still one container/server

Responsibilities

• Hardware management
  • Data Center (on-premise)
• OS management, patch management
  • IaaS
• App management
  • PaaS

Serverless

• New computational model
• Computational unit is 1 function call (Function-as-a-Service)
• If no function call happens, nothing to pay for
• Other category: Backend-as-a-Service, DB-as-a-Service

Function-as-a-Service

• Event driven architecture
• Triggers + Functions
  • Messages, Schedules, State changes
• Billing unit: GB-seconds
• Constraints:
  • memory
  • execution time (5 min in Lambda)
  • parallelization
• Stateless

Backend-as-a-Service

• Common aspects of application
  • Authentication, Authorization, User tracking
  • eg: Auth0, Firebase
• Heavily used in Mobile development
• API driven
• Easy to integrate

Serverless Systems

• AWS Lambda
  • First amongst public clouds
  • Nicely integrates with other part of AWS
• Azure Functions
  • Very similar to AWS Lambda
  • Two "Consumption Plans"
• Google Cloud Functions
  • Still in Beta
• OpenWhisk/IBM Bluemix
  • Open source
• Over kubernetes serverless:
  • Funktion, Kubeless, Fission

Serverless Architectures

• Event-driven, "nanoservices"
• Supports popular runtimes
  • Java, NodeJs, Python, C#
• IoT is main target
• Events, Triggers
  • HTTP triggers
    • eg.: AWS API Gateway
  • Pub/Sub triggers
  • Message queue triggers
  • State changes
    • Uploaded file to blob storage
  • Scheduled events ("cloud cron")

Serverless Cons

• Not easy to estimate costs upfront
• Can be expensive
  • 3mill req, each 1 sec, 512 MB memory
  • AWS: 15$, Azure: 20$
  • smallest instance: 5$/month both in Azure/AWS
• "Cold start"
  • First call after idle time is slow
• Function initialization is not deterministic

Serverless Cons #2

• Patching is not in our hand (fixed runtimes)
• No service discovery
• Cannot attach debugger
  • More logging needed
• Hard to test locally
  • AWS SAM Local
  • Azure Functions Runtime

Serverless Framework

• Formerly JAWS,
• First framework to utilize AWS Lambda/API GW
• CLI written in node
• Support major serverless platforms
  • AWS
  • Azure
  • Google Cloud
  • OpenWhisk

Bottom Line

• Lot of Cons, but what are the pros?
• Very easy to deploy
• Ideal for prototyping
• No need for classic ops (patch management, software install & config, ssh, etc).
• Free tier can handle small websites
• Use cases:
  • Event driven architecture 
  • IoT
  • Automation

The End!

Thank you!