Contact v.kazulkin@iplabs.de Vadym Kazulkin ip.labs GmbH - - PowerPoint PPT Presentation

Contact

Vadym Kazulkin

ip.labs GmbH

v.kazulkin@iplabs.de xing.com/profile/Vadym_Kazulkin @VKazulkin @ServerlessBonn (Meetup)

Christian Bannes

ip.labs GmbH

c.bannes@iplabs.de xing.com/profile/Christian_Bannes

Ip.labs GmbH

To FaaS... …or not to FaaS

Image: https://aws.amazon.com/de/lambda/features

Function as a Service = AWS Lambda

Image: https://aws.amazon.com/de/lambda/features

Serverless is more than Function as a Service

Image: https://www.infoq.com/articles/serverless-sea-change/

Typical Three-Tier

Image: https://www.infoq.com/articles/serverless-sea-change/

Function as a Service

Image: https://www.infoq.com/articles/serverless-sea-change/

Serverless Applications

Infrastructure cost

Image: https://stocksnap.io

Velocity of your team

Image: https://burst.shopify.com

What you really have to consider

Total Cost of Ownership

Lambda pricing model

Cost for lambda

REQUEST DURATION

Request Tier

$ 0.20

Per 1 Mio Requests

Duration Tier

$ 0.00001667

Per GB-Second

GB-Second

ONE SECOND ONE GB

Example

• 1 Mio requests
• Lambda with 512MiB
• Each lambda takes 200ms

0.5 GiB * 0.2 sec * 1 Mio = 100 000 GB-Seconds

Requests:

$0.20

GB-Seconds:

$1.67

Cost scales linearly with memory

How does this compare to VM pricing model?

Lambda performance compares to

• M4 (general purpose)
• M5 (general purpose)

EC2 Instance Families

M5 General Purpose Instance

Cost scales linearly with memory

Lamdba vs VM

https://www.youtube.com/watch?v=b1ebYEfQX1s

Lambda compute is 5x more expensive than M5

Threading Model

Idle Idle

Pay per wall clock time Pay per use

Image: https://aws.amazon.com

AWS monetizes the datacenter… … by selling time on their servers

Common pitfalls when working with lambda

More memory = more expensive?

Power tuning lambda

• Executes different

settings in parallel

• Outputs the optimal

setting

Image: https://github.com/alexcasalboni/aws-lambda-power-tuning

Paying for wait time

Image: https://blog.binaris.com/lambda-pricing-pitfalls/

The reality is…

Lambda is often just a small percentage

• f your total cost

$ 3.50

Per million API calls Api Gateway

Event Sources

• SQS and SNS are

charged for requests

• Kinesis charges for

shard hours & PUT requests

Image: https://blog.binaris.com/lambda-pricing-pitfalls/

Event Sources

• Cost for Kinesis grows

with slower rate

• Attractive at to
• perate at scale

Image: https://blog.binaris.com/lambda-pricing-pitfalls/

DynamoDB On-Demand

Image: https://aws.amazon.com/blogs/aws/amazon-dynamodb-on-demand-no-capacity-planning-and-pay-per-request-pricing/

Provisioned vs On-Demand

• Use On-Demand for

spiky workloads

• Use Provisioned for

constantly high workload

Cost saving with On-Demand

Image: https://twitter.com

Switch to "On-Demand"

Example

$2400 / Month $520 / Month 100 API requests per second 24/7

• API Gateway
• Lambda with 1GB and

300ms per request

• Load balancer
• 3 x M5.xlarge with

16GB each

Serverless vs VM

• "Pay as you go" can significantly reduce server cost

because you don’t pay for ide time

• For serverless you need right architecture
• Consider the Total Cost of Ownership

What you really have to consider:

Total Cost of Ownership

Ajay Nair (https://twitter.com/ajaynairthinks/status/1057696133076541441)

TCO Full Picture

Forrest Brazeal „The Business Case For Serverless” https://www.trek10.com/blog/business-case-for-serverless/

No Infrastructure Maintanence Auto Scaling Built in Fewer Engineers Required Lower Technical Debt Focus on Business Value & Innovation Faster Time to Market

Do more with (Server-)less

FaaS... …or not to FaaS ?

Image: https://stock.adobe.com

• 1. Application lifecycle
• 2. Workloads
• 3. Platform limitations
• 4. Cost at scale
• 5. Operational constraints
• 6. Organizational knowledge

Decision Checklist

Understand Application lifecycle

Christian Posta „Be as serverless as you can, but not more than that” https://blog.christianposta.com/serverless/its-not-about-microservices-vs-serverless/

Explore phase

• Quickly validate

hypotheses

• Rapidly experiment
• Run experiments as

cheaply as possible

Image: burst.shopify.com/photos/a-look-across-the-landscape-with-view-of-the-sea

Serverless is a perfect fit

Exploit phase

• Built something that does

provide customer value

• Build it on scale
• Build a profitable product

around it

Image: Robert Scoble via Flickr

partly serverless and partly not serverless architecture

Application lifecycle

• How much of my stack should I own

to be able to deliver business value?

• Outsource SLA, regulatory

compliance, price, and roadmap to my service provider?

Existing applications

• You can’t magically move

that all off to service providers

• You can try to modernize

parts of them

Strangler Pattern

• Add a proxy (API

Gateway or Application Loadbalancer), which sits between the legacy application and the user

• Add new services and

link it to the proxy

Marin Fowler „StrangerFigApplication” https://martinfowler.com/bliki/StranglerFigApplication.html

FinDev Concept

Activity-based costing on a digital operation-by-operation basis

• Figure out features which deliver

business value comparing to their cost

Aleksander Simovic & Mark Schwarz „FinDev and Serverless Microeconomics: Part 1” https://aws.amazon.com/de/blogs/enterprise-strategy/findev-and-serverless-microeconomics-part-1/

• 1. Application lifecycle
• 2. Workloads
• 3. Platform limitations
• 4. Cost at scale
• 5. Operational constraints imposed by other dependencies
• 6. Organizational knowledge

Understand your workloads

• Event-driven
• API-driven
• Batch Job
• ML/AI
• Big Data

Image: flickr.com/photos/everywhereatonce/294789504

Lambda Layers & Lambda Runtime API

Door opener for use case like:

• Big Data
• ML/AI

Understand your workloads

• Do we need to access specialized

hardware ?

• GPU access required?
• Another RAM/CPU ratio?
• Do we need constantly high

performance?

• Response time below 100 ms

(e.g. bidding or gaming platforms)

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

Understand your workloads

• Do we need high throughput ?
• Lambda‘s network bandwidth is limited

(an order of magnitude lower than a single modern SSD) shared between all functions packed on the same VM

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

Storage TradeOffs

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

Latency Scaling Pricing Consistency

S3 Storage TradeOffs

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

Latency Scaling Pricing Consistency

The State of Serverless Computing

“Chenggang Wu & Vikram Sreekanti - The State of Serverless Computing” Craft Conference 2019 - Budapest , Hungary https://www.youtube.com/watch?v=htLQiSPMUmk&list=LLYgjRSI2oCzI9eooyFrWR7A&index=6

Understand your workloads

• Do functions need to communicate with

each other?

• functions are not directly network

accessible, they must communicate via an intermediary service

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

• 1. Application lifecycle
• 2. Workloads
• 3. Platform limitations
• 4. Cost at scale
• 5. Operational constraints imposed by other dependencies
• 6. Organizational knowledge

Understand platform limitations

• Cold start
• Lambda with and without VPC for

each runtime

Ajay Nair „Become a Serverless Black Belt” https://www.youtube.com/watch?v=oQFORsso2go

Ajay Nair „Become a Serverless Black Belt” https://www.youtube.com/watch?v=oQFORsso2go

Lambda in VPC

Chris Munns "Announcing improved VPC networking for AWS Lambda functions" https://aws.amazon.com/de/blogs/compute/announcing-improved-vpc-networking-for-aws-lambda-functions/

Lambda in VPC

As function’s execution environment scales

• More network interfaces are created and

attached to the Lambda infrastructure.

• The exact number of network interfaces

created and attached is a factor of your function configuration and concurrency.

Chris Munns: "Announcing improved VPC networking for AWS Lambda functions” https://aws.amazon.com/de/blogs/compute/announcing-improved-vpc-networking-for-aws-lambda-functions/

Lambda in VPC

As your function scales, be mindful of several issues:

• Managing the IP address space in your

subnets

• Reaching the account level network

interface limit

• The potential to hit the API rate limit on

creating new network interfaces

Chris Munns: "Announcing improved VPC networking for AWS Lambda functions” https://aws.amazon.com/de/blogs/compute/announcing-improved-vpc-networking-for-aws-lambda-functions/

NEW: Lambda in VPC:

Leveraging AWS Hyperplane to provide NAT capabilities from the Lambda VPC to customer VPCs

Chris Munns: "Announcing improved VPC networking for AWS Lambda functions” https://aws.amazon.com/de/blogs/compute/announcing-improved-vpc-networking-for-aws-lambda-functions/

NEW: Lambda in VPC:

• The network interface creation happens

when Lambda function is created or its VPC settings are updated.

• Because the network interfaces are shared

across execution environments, only a handful of network interfaces are required per function

• Function scaling is no longer directly tied to

the number of network interfaces

• Hyperplane ENIs can scale to support

large numbers of concurrent function executions

Chris Munns: "Announcing improved VPC networking for AWS Lambda functions” https://aws.amazon.com/de/blogs/compute/announcing-improved-vpc-networking-for-aws-lambda-functions/

Amazon Aurora Serverless Data API as beta for MySql available

https://docs.aws.amazon.com/de_de/AmazonRDS/latest/AuroraUserGuide/data-api.html

Don‘t be scared

• f cold starts

To avoid cold starts them completely, you have to :

• Overpay
• Overprovision

Cold starts don’t really matter if the you make the call asynchronously

Understand platform limitations

• Cold start
• Lambda with and without VPC for

each runtime

• Invocation duration/timeouts
• Lambda 15min
• API Gateway integration 29sec
• Max Memory
• Lambda 3GB

Understand platform limitations

• Max concurrent invocations
• 1000-3000 parallel executions for

all Lambdas in each AWS account

• Scaling thresholds
• Dynamo DB provisioned and on-

demand capacities

• 1. Application lifecycle
• 2. Workloads
• 3. Platform limitations
• 4. Cost at scale
• 5. Operational constraints imposed by other dependencies
• 6. Organizational knowledge

Understand your cost at scale

• Lambda
• API Gateway
• Dynamo DB capacity choices
• Remote API calls
• Logging costs
• Monitoring costs

• 1. Application lifecycle
• 2. Workloads
• 3. Platform limitations
• 4. Cost at scale
• 5. Operational constraints
• 6. Organizational knowledge

Understand the operational constraints (imposed by other dependencies)

• Language runtime choice
• Can be forced by some dependencies to
• ther libraries

Yan Cui „Not so FaaS” https://blog.binaris.com/not-so-faas/

• 1. Application lifecycle
• 2. Workloads
• 3. Platform limitations
• 4. Cost at scale
• 5. Operational constraints imposed by other dependencies
• 6. Organizational knowledge

Understand organizational knowledge

• Do you already have

DevOps knowledge?

Tom McLaughlin Talk: What do we do when the server goes away?

• Monitoring & Alerting
• Chaos Engineering & Game Days
• Infrastructure as Code & Testing
• Help understand constraints
• f AWS services & choose the right
• ne

Tom McLaughlin „What do we do when the server goes away” https://speakerdeck.com/tmclaugh/serverless-devops-what-do-we-do-when-the-server-goes-away

Understand organizational knowledge

• Do you already have

DevOps knowledge?

• Are the developers willing to

learn new languages?

Future of Serverless

Berkeley View on Serverless Computing

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

Berkeley View on Serverless Computing

• Provide low latency and high IOPS

Serverless Ephemeral Storage

• Provide Serverless Durable Storage
• Improve Networking and Performance
• Improve Security
• Accommodate cost-performance

“A Berkeley View on Serverless Computing” https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-3.html

• Application lifecycle
• Workloads
• Platform limitations
• Cost at scale
• Operational constraints
• Organizational knowledge

FaaS or not to FaaS

Thank You!