Docker vs. Kubernetes vs. Serverless

Containers and Serverless Computing:
Selecting the Right Platform for Your Application
Phil Christensen
Sr. Solutions Architect, DevOps Engineer

©2017 Logicworks. All rights reserved.
About Logicworks and Me
2
Phil Christensen
Sr. Solutions Architect
Logicworks
§ 15+ years of experience in software
engineering and cloud
§ Formerly Sr. DevOps Engineer
§ Hold all 5 AWS Certifications
Logicworks designs, builds,
automates, and manages AWS clouds.
Cloud Migration
24/7 Management
Cloud Automation
Cloud Security

The standard deployment model for
applications is changing rapidly.

Containers Serverless
</> </> </>
</> </> </>
</> </> </>

Let’s Build Reusable Artifacts!
5
§ Decrease deployment time
§ Leverage the power of CI/CD
§ Every artifact created by a specific
recipe or manifest
§ A common interface reduces
operational complexity
§ The exact same artifact is run in local
development, production, or any
other part of the SDLC

Why Containers?
6
Docker containers
provide a packaging
standard and a
common interface for
running, deploying, and
managing workloads on
compute instances.
SERVER
APP1
Host OS
APP4
APP2
APP3
SERVER
Host OS
APP1
APP4
APP2
APP3
DockerAPP5
APP5

§ Container runs on top of your preferred
workstation OS
§ Run with reduced memory or CPU usage
in development
§ Finished image moved as-in to
deployment server
Local Development
7
The first benefit of a containerized
development approach is the ability
to deploy the exact same artifact to
your server that was created in
development.
DEV WORKSTATION
Host OS
APP
Docker
SERVER
Host OS
Docker
APP
APP
APP

§ Container image behaves the same on any host
§ Complex dependencies are only compiled or
assembled at image build time
§ Safely encapsulate applications requiring older
Linux distributions
Container Portability
8
Container portability simplifies
deployment process and reduces the
likelihood of deployment failure
QA
Host OS
APP
Docker
APP
APP
APP
STAGING
Host OS
Docker
APP
APP
APP
PRODUCTION
Host OS
Docker
APP
APP
APP
APP
APP

Benefits
§ Already deployed with Docker
§ Simplest configuration, integrates with Docker-
Compose
Considerations
§ Limited cloud integration
§ Service discovery by DNS only
§ One ELB per cluster
Option 1: Docker Swarm
Orchestration Options
10

Benefits
§ Most feature-filled orchestration layer
§ Large community support
§ Many deployment options
§ Powerful, built-in container discovery
Considerations
§ Most complex install/configuration
process
§ Difficult to stay up to date
Option 2: Kubernetes
11
Developer /
Operator
API Server Controller Manager
Scheduler
etcd
Kubernetes Master
Kubelet
Kubernetes Node
cAdvisor Kube-Proxy
Pod Pod Pod
Kubelet
Kubernetes Node
cAdvisor Kube-Proxy
Pod Pod Pod
Plugin Network (eg Flannel, Weavenet, etc)
Users

Benefits
§ Highest capacity for scale
§ Deeply integrated with AWS
ecosystem
Considerations
§ Amazon cloud-only deployment
§ Container discovery requires third-
party applications like Consul
Option 3: Amazon Elastic Container Service
12

Achieving High Availability
13
Docker Swarm Kubernetes AWS ECS
Management Tier A 3-5 node manager tier
responds to requests on a
single ELB, delegates to N
worker nodes
§ A 3-5 node master tier responds to API
calls and web requests, delegates to N
minion nodes.
§ Services can leverage individual ELBs
directly, or define NodePorts that are
routed through the master tier.
§ Control plane fully managed by
AWS
§ ALB and ELB route traffic to
appropriate containers
Management Tier
Failure
Manager nodes must
maintain a quorum, but a
failed manager will continue
to run services
Master nodes must maintain a quorum,
and a failed master tier will cause most
services to fail
No single point of failure in managed
control plane.
Worker Nodes
Replaced?
Lost worker nodes
automatically replaced
Lost minion nodes automatically replaced Worker nodes are easily added,
replaced, or removed
Updates Cluster can be upgraded in-
place
In-place cluster upgrades still maturing,
3rd-party distributions may differ
Agent upgrades can be performed in-
place

Cluster Capacity
14
Optimized For Optimized for multiple smaller
clusters per SDLC
Optimized for a single large cluster Optimized for one cluster per SDLC
Load Balancing Load balancer limitations can
arise with specific SSL or
DNS requirements
Namespaces help organize SDLCs, or
delegate access to sets of containers.
§ Best-in breed container
AutoScaling
§ Native ALB support
Node Support Supports 2000+ nodes Supports up to 5000 nodes Supports up to 1000 nodes
Container Limit Limited to 95,000 containers Limited to 300,000 containers Limited to 500,000 containers

How Deployments Work
15
Deployment
Details
§ Simple docker CLI
commands invoked on
management node.
§ Uses Docker-Compose for
configuration Limited
support for secrets
§ Support for rolling updates
§ Largest number of deployment options
§ Kubectl CLI can be run from anywhere
with access to master tier API
§ Native “Deployment” type provides
robust interface for updates, secrets,
and infrastructure management
§ Deepest integration with other AWS
services like CodeBuild, ECR
§ Best-in-breed support for ALB
Target Groups
§ Can be invoked via the AWS API
§ Support for Docker-Compose
manifests
AWS Resources
Need to be Pre-
Created?
Yes No, will manage and create necessary
AWS resources.
Yes

Example Container Pipeline
16
§ Modify the source
§ Commit changes
§ Webhook triggers
build
§ Pull source
§ Compile assets
§ Run tests
§ Build container
§ Push container
§ Private image
repository
§ AM-based
authentication
§ CloudWatch events
trigger Lambda on
push event
§ Lambda invokes
Kubernetes API to
publish Deployment
§ Native k8s deployment
process allows for
seamless updates

You are exchanging flexibility for scale.
What is Serverless?
18

§ Quick start-up times
§ Shared compute pool
§ No need for orchestration – cloud provider
handles it for you
§ You only pay for what you use
§ Deploy arbitrarily many development instances
at little to no cost
Why Serverless?
19
§ Less flexibility around development techniques
§ Limited to supported frameworks (Sorry Ruby
fans!)
§ Reduced access to native functionality (binding
to custom C libraries)
§ Limits on function execution time
Benefits Considerations

§ A typical server is up to 80% idle of most of
the time
§ Traditional development instances either
need to be parked, spun down, or deleted
when not in use
§ For bigger workloads, there’s a break-even
point where below a minimum throughput,
you’re still better off writing Lambda functions
§ Focus on the memory and execution time
that a typical transaction in your app will
need
Cost Effectiveness of Serverless
20
Exec Time @
Memory Used
m4.large Break-
Even Point
Requests Per
Second
100ms @ 128 MB 295,000 requests 81.9
200ms @ 512 MB 64,000 requests 17.8
200ms @ 1GB 34,000 requests 9.4
1 sec @ 1GB 7,100 requests 2.0
AWS Lambda Pricing in Context, by Andy Warzon @ trek10.com

§ Limited runtime to ensure one function
doesn’t dominate over others behind
the scenes
§ More complex workflow requires an
understanding of asynchronous
techniques
§ Step Functions used to aggregate
multiple functions into a larger
application
Asynchronous
21
Find Instances
Count Instances
Iterator
Iterate Instances
Done Select Backup Type
Create SnapshotCreate AmiEnd
Start

§ Lambda is dramatically less
expensive
§ It takes a great deal of compute
volume before Lambda
approaches EC2 costs
§ TCO is greatly reduced, little to
no need to manage infrastructure
Good Use Cases for Serverless
22
Periodic or Light Workloads API-Only Applications Long Running or Deeply Integrated
AWS Scripts or Applications
§ Tight integration between API
Gateway and Lambda allows for
easy API creation
§ API Gateway helpers can provide
authentication and other
functions
§ Frequently used management
scripts are best created as
Lambda functions
§ Execution time limits encourage
planning ahead for large
resource counts
</>API

Lowest Common Denominator
23

Autonomous IT
24
Serverless

Most Efficient Scaling
25
Serverless

Best Scale for Conventional Software
26
Amazon ECS

Exchange Simplicity for Flexibility
27

Standard Management Layer
for Undifferentiated Compute
28

§ Train your IT staff
§ Give expert advice about the right DevOps
tools for your application
§ Build a new AWS environment optimized for
containers
§ Provide 24x7 support for AWS infrastructure
Logicworks can help you:
Come Talk to Me!
29
www.logicworks.com
info@logicworks.com
(212) 625-5300
§ CONTACT US

155 Avenue of the Americas, Fifth Floor | New York, NY 10013
P:212.625.5300 | www.logicworks.com
Questions?

Docker vs. Kubernetes vs. Serverless

More Related Content

What's hot (20)

Similar to Docker vs. Kubernetes vs. Serverless (20)

Recently uploaded (20)

Docker vs. Kubernetes vs. Serverless