Cloud Native Akka & Kubernetes: the holy grail to elasticity? by Fabio Tiriticco Have you ever wondered how to actually achieve elasticity on the cloud, outside of your cushy local dev environment? Akka is the most mature choice to implement the traits of the Reactive Manifesto, thanks to the Actor model. Resilience is enabled by asynchronous message passing and Actor supervision. Location transparency, message routing and the “share nothing” approach enable seamless scalability, but once our app is deployed we need to rely on some external infrastructure to automatically scale up or down our services. At this point, Akka alone is not sufficient anymore. What else do we need? We found Docker & Kubernetes to be a perfect match for clustered Akka applications running in the cloud. Thanks to the Kubernetes API, one can write infrastructure logic that complements the business logic. This leads to a general approach of infrastructure-as-code tailored to each different business case, which in turn lets companies achieve maximum efficiency and resource usage. During this talk we will cover features from both Akka and Kubernetes, plus example code in Scala: familiarity with these technologies is recommended. Our example will rely on Google Cloud infrastructure.

1. CLOUD NATIVE AKKA & KUBERNETES:
THE HOLY GRAIL TO ELASTICITY?
👎 👍

2. Notes
• My co-speaker is on holiday
• Only relevant concepts are introduced
• Slides use pseudo-code. Real deal on GitHub
@adamsand0r
🎡🍸

3. About me
Fabio Tiriticco
Software Developer / Tech Lead
@ticofab

4. The Reactive Manifesto
Elasticity: “use just the right amount”
2013, J. Boner, R. Kuhn, D. Farley, M. Thompson

5. Akka’s building blocks: Actors
• contains state &
behaviour logic
Actor 1
Actor 3
Actor 2
• has an addressable
mailbox to receive
and send messages
myActor3.getName()
“Joe”
“What is your name?”
“My name is Joe”
“Bob”
“Tim”

6. My first journey in the Reactive world

7. My first journey in the Reactive world

8. My first journey in the Reactive world
Phone 1
actor
Phone 2
actor
Matcher
actor
Phone 3
actor
A possible architecture
for CloudMatch
Phone 4
actor
“You matched with Phone
4, his address is xxx”
“You matched
with Phone 1,
his address is
yyy”

9. Location transparency

10. Location transparency
Actor 1
Actor 2
“Bob”
“Tim”
// in Scala + Akka
timAddress ! WhatIsYourName
• Locally
• Remotely

11. Elasticity test system
Job
source Master
Worker
Worker
HTTP
Adjust rate

12. The Worker
class Worker extends Actor {
// ...
case job: Job =>
// .. simulate a CPU-intensive workload
master ! JobCompleted(job.number)
}

13. Elasticity test system
Job
source
Worker
Master
Worker
Worker
Scaling
controller
ADD NODEREMOVE NODE
TOO LITTLE POWER!TOO MUCH POWER!

14. Issue #1: how to distribute this?
Job
source
Master
Worker
Worker
Scaling
controller

15. Node 1
Akka Cluster to the rescue!
• Joining
• Up
• Leaving
• Unreacheable
• …
Seed
Node 1
Seed
Node 2
Node 2
Node 3
NOTE: nodes need to know where the seed is

16. Moving workers to different nodes
JVM 2 - worker
JVM 3 - worker
Job
source
Master
Worker
Worker
Scaling
controller
Single JVM
JVM 1 - seed
Issue #2: how to spin up and shut down nodes?

17. Docker + Kubernetes
• Node provisioning
• Resilience
• Mainstream adoption
• Cloud providers
• API

18. Kubernetes
controller
Scaling Controller → Kubernetes Controller
Job
source
Master
Worker
Worker
Scaling
controller

19. Kubernetes Controller
class KubernetesController extends Actor {
case AddNode =>
// use kubernetes api to scale up
case RemoveNode =>
// use kubernetes api to scale down
}

20. How does the master asks for changes?
class Master extends Actor {
// to add a node
kubernetesController ! AddNode
// to remove a node
kubernetesController ! RemoveNode
}

21. Kubernetes Architecture
Master
Actors
Kubernetes
Controller
Worker Worker Worker
akka-
master-3fw4
akka-
worker-1
akka-
worker-2
akka-
worker-3
Pods
akka-master
(Deployment)
akka-worker
(Deployment)
Controllers
Services
DNS
HTTP
Job Source
JVM seed JVMJVMJVM

22. A glance at the code, then a live recorded demo
📜 📽,

23. Burndown rate pattern upon changes
+
-
0

24. Measurements

25. HTTP vs Remote Actor Messaging
• Actor come with built-in elasticity
• Async messaging gives you more control
• Less complexity
• Actor messages do not necessarily need a
response
• HTTP is ubiquitous

26. HTTP vs Remote Actor Messaging
•Performance!
J. Janssen

27. Biggest learning
REACTIVE SYSTEM DESIGN
=
APPLICATION AND INFRA

28. And oh, about the holy grail
?
No holy grail in software development.

29. THANKS
github.com/ticofab/akka-cluster-kubernetes
@ticofab
@adamsand0r