This document discusses serverless computing and Nuclio, an open source serverless platform. It provides an overview of serverless and its benefits, describes Nuclio's features like high performance, support for any programming language or data source, and easy deployment. Examples are given for real-time serverless applications involving IoT, analytics, and AI.

1. iguazio © 2016
1
https://nuclio.io
Using Serverless for Big Data and AI

2. iguazio © 2016
2

3. iguazio © 2016
3

4. iguazio © 2016
4

5. iguazio © 2016
5
High-Level Abstractions and Serverless Solve the Complexity
Enabling faster business transformation

6. iguazio © 2016
6
The New Stack
Analytics and AI Frameworks
Packaged
Management Applications
Unified
Orchestration
Cloud Data
Services
UNSTRUCTURED
(FILE & OBJECT)
STRUCTURED
(NO / NEWSQL)
STREAM
S3
Glustre
iguazio
DynamoDB
Cassandra
Spanner
iguazio
Kafka
Kinesis
iguazio
Serverless Functions

7. iguazio © 2016
7
What is Serverless?
“No need to spend time and resources on server provisioning, maintenance,
updates, scaling, and capacity planning. Instead, all of these tasks and capabilities
are handled by a serverless platform and are completely abstracted away from the
developers and IT/operations teams” - CNCF Serverless Paper
Event
Sources
Function
Instance
invoker
Function
Instance
invoker
Function
Instance
invoker
Event
Sources
Event
Sources
Synchronous or
Asynchronous Invocation
FaaS Controller
Platform Services (Identity, Data, etc.)

8. iguazio © 2016
8
Serverless Invocation Modes
Function
Instance
invoker
Message
Function
Instance
invoker
Function
Instance
invoker
Exchange
Message Queue
(e.g. RabbitMQ)
HTTP
API GW
Function
Instance
invoker
Function
Instance
invoker
Function
Instance
invoker
Req
Function
Instance
invoker
Function
Instance
invoker
Function
Instance
invokerPartition 1
Messages
Synchronous Req/Rep
Message Stream
Async Message Queue
Kafka,
Kinesis, etc.
Function
Instance
invoker
Function
Instance
invoker
Function
Instance
invoker
Job
Job (Master/Worker)
Priority
Queue
Master
(dealer)
Dealer
Partition 2
Partition 3
Partition 4

9. iguazio © 2016
9
Using Serverless for Big Data and AI
• Serverless:
Easy to deploy functions, no scripts, Dockerfiles, build, etc.
Auto-scaling and event triggers are built-in
Use any language
• What’s missing in current implementations?
Slow performance, lack of concurrency, no stream partitioning
Limited number of event sources (mostly HTTP, some Kafka)
Hard to debug and diagnose
Hard to manage complex functions and dependencies

10. iguazio © 2016
10
nuclio: A New Real-Time Serverless Platform
Comprehensive, Open, Super Fast and runs ANYWHERE
Any source and
any workload
10
Event
Listeners
Real-time, Auto-scaling Function OS
Language Runtime
Function
Workers
Data
Bindings
Pluggable
Data Services
Pluggable
Event Sources
Platform Specific Controller
image
repo
Local or shared
Pluggable
log & monitor
Common Services
Dealer
Builder
CLI
Web UI
Packaged as a binary or container image
Platform API: config, log, monitor, security
HTTP, AMQP, MQTT, Kafka,
Kinesis, NATS, Files/DB
nuclio Platform Services
400K Events/sec per Process
Maximum performance
Dev, test, and run
ANYWHERE
Simple, fast and
secure data
integrationhttps://github.com/nuclio/nuclio , https://thenewstack.io/whats-next-serverless/
http://nuclio.io/

11. iguazio © 2016
11
Data Bindings
$ nuctl deploy <name> <source> [options]
Enabling Simple and Continuous Dev and Ops (CI/CD)
One click to test, deploy, upgrade or rollback code
Runs ANYWHERE, Self-healing and Auto-Scaling
LOCAL or CLOUD

12. iguazio © 2016
12
nuclio Application Examples
APIS, UI,
DASHBOARDS
DATA / EVENT
SOURCES
DATA
SERVICES
DATA
SERVICES
DATA
SOURCE
DESTINATION
IOT DEVICE /
GATEWAY
CENTRAL /
EDGE CLOUD
Real-time Analytics & AI Data Transformation & Movement
IoTAPIs, UI, Bots
APP
SERVICES
APP
SERVICES
SENSORS

13. iguazio © 2016
13
nuclio Perf Results: Single Process, Basic Echo Functions
https://github.com/v3io/http_blasterTested using:
Native
Prometheus
Integration
Lang Threads HTTP Req/sec Avg Latency
1 40,281 24 us
100 372,496 2.03 ms
1 30,156 38us
100 66,681 3.41 ms
Setup: a single AWS C5.9Xlarge instance
Other Serverless/FaaS platforms processed only
1-5K events/sec at magnitudes higher latency !
Up to 100X Faster
Py
Go

14. iguazio © 2016
14
nuclio Playground – Online Dev, Config, and Test Environment

15. iguazio © 2016
15
Function Example – Sentiment Analysis
Dependencies
(in YAML, UI or Code comments)
Call API
(Trigger independent)
Structured
Logging
Structured or Unstructured Response

16. iguazio © 2016
16
Dealer
Processor
Function
Workers
Partitioned data or
stream shards
Job Spec:
- Functions (selector)
- Task num/list
- Max tasks per processor
- Min/Max processors
- ..
- Job Metadata
Function
Workers
Job X (w 5 tasks)Job Y (w 4 tasks)
POD up/down events
Deployment scale changes
Auto-scale based on
CPU load or Q delay
Allocate or re-distribute tasks to
processors (1 task per worker)
Nuclio
controller
nuclio Dealer
• Enables real-time stream processing, batch and
interactive jobs on auto-scaling serverless functions
• Dynamically allocates tasks to workers and handles
task lifecycle, checkpoint and completion
Every job or stream is
partitioned to N smaller tasks
Processor

17. iguazio © 2016
17
Serverless is Enabling Intelligent and Continuous Applications
ACTIONS, INTERACTIVE
UI & DASHBOARDS
EXTERNAL
SOURCES
OPERATIONAL
DATA
CONTAINERIZED AI & ANALYTICS TOOLS
EXTERNAL DATA
SOURCES / LAKES
DATA SERVICES
Productivity | Faster Insights | Invisible Infrastructure | Lower TCO
SyncIngest / Alert
APIs &
Triggers
Analyze

18. DEMO
Type the following command and browse to <host-ip>:8070
docker run -p 8070:8070 -v /var/run/docker.sock:/var/run/docker.sock
-v /tmp:/tmp nuclio/playground:0.2.0-amd64

19. iguazio © 2016
19
Real-time Maps
& Dashboards
Extract Metadata:
Geo location, time,
device, thumbnail
Deep Learning:
Detect people,
gender, age, etc.
Queries :
Find people in
places, relations
Missing person
notification
Trigger
Update
SMS
Query/Scan
Trigger
Update
Trigger
• Pictures + Metadata
• People DB, Locations
and relations
Ingest
DATA SERVICES
Real-time Surveillance Example

20. iguazio © 2016
20
Streaming vehicle
data via web APIs
Real-time Analytics for Alerts and Proactive Maintenance
Real-time enrichment with
weather & road data
Finds correlations
between weather
conditions and vehicle
components
Weather
Microservice for
data ingestion
Customer use case #1
Road info
Real-time alerts
Real-time
Dashboard
DATA SERVICES

21. iguazio © 2016
21
Real-time Fleet Management Dashboard
Real-time
enriched data
Fuel consumption
and weather
correlation
Vehicle location
heatmap
Filter by bad
weather alerts
Customer use case #1

22. iguazio © 2016
22
Real Example: Event-Driven Analytics for Connected Cars
Geo Data
Weather/Road info
Vehicles Data
State
Changed?
Identify
Violation?
Drivers
Violations
Stream
State
Changes
Geo
Aggregate
Map
Process
Alerts
Process
Violations
External Sources
Import
service
Enriched
Events
Parallel
Enrichment
ML Processing
Complex Events + Data processed in real-time without the infrastructure hassle
real-time, auto-Scaling
serverless functions
Model Update
Stats
Update
* See code in the
UI/Playground slide

23. DEMO

24. Thank You
nuclio is easy: try it and give it a
https://github.com/nuclio
Sign up for our online hackathon and build the greatest serverless
application on nuclio. You just might win a Phantom 4 drone…

25. iguazio © 2016
25
nuclio
Function Spec
Support Kubernetes CRD:
Functions can be created and
deleted using kubectl
tags/labels used for search and
event sources (Label Selectors)
Control Min/Max Replicas
for controlled auto-scale
Pass text or secret
environment variables
(k8s convention)
Flex resource allocation,
GPUs are coming
Pluggable Data
Sources
Various src code options*: inline code, path
(local/http/git), or local/remote pre-built image
namespaced
For the full spec see: https://github.com/nuclio/nuclio/blob/master/docs/configuring-a-function.md

26. iguazio © 2016
26
CLI (Run Command Example)
$ nuctl run --help
Build, deploy and run a function
Usage:
nuctl deploy function-name [flags]
Flags:
--base-image string Name of base image. If empty, per-runtime default is used
--build-command String Commands to run on build of processor image
--data-bindings string JSON encoded data bindings for the function (default "{}")
--desc string Function description
-d, --disabled Start function disabled (don't run yet)
-e, --env string Environment variables (name1=val1,name2=val2..)
-f, --file string Function configuration file
--handler string Name of handler
-h, --help help for deploy
-i, --image string Docker image name, will use function name if not specified
-l, --labels string Additional function labels (lbl1=val1,lbl2=val2..)
--max-replicas int Maximum number of function replicas
--min-replicas int Minimum number of function replicas
--no-pull Don't pull base images - use local versions
--nuclio-src-dir string Local directory with nuclio sources (avoid cloning)
--nuclio-src-url string nuclio sources url for git clone (default "https://github.com/nuclio/nuclio.git")
-o, --output string Build output type - docker|binary (default "docker")
-p, --path string Function source code path
--port int Public HTTP port (node port)
--publish Publish the function
-r, --registry string URL of container registry (env: NUCTL_REGISTRY)
--replicas int If set, number of replicas is static (default 1)
--run-image string If specified, this is the image that the deploy will use, rather than try to build one
--run-registry string The registry URL to pull the image from, if differs from -r (env: NUCTL_RUN_REGISTRY)
--runtime string Runtime (e.g. golang, golang:1.8, python:2.7)
--triggers string JSON encoded triggers for the function (default "{}")
--version string Docker image version (default "latest")
Global Flags:
-k, --kubeconfig string Path to Kubernetes config (admin.conf)
-n, --namespace string Kubernetes namespace (default "default")
--platform string One of kube/local/auto (default "auto")
-v, --verbose verbose output
See more in: https://github.com/nuclio/nuclio/blob/master/docs/nuctl/nuctl.md