Spark Streaming the Industrial IoT

Spark Streaming the Industrial IoT
Washington DC Area Spark Interactive
Jim Haughwout, Chief Architect & VP of Software
May 24, 2016

© 2016 Savi Technology • May 25, 2016 • Page 2
Today’s Talk
• Discuss challenges of streaming in general with tips on doing this with Spark
• Special focus: IoT’s complexities of immediately tying together physical and data world
• Our talks are in three parts:
- Part I: Top-level POV of using Spark Streaming for Industrial IoT (Jim)
- Part II: Spark Streaming and Expert Systems – Spark + Drools (James)
- Part III: Overcoming Deficiencies in Streams (Anderson of MetiStream)

Savi Technology
• Sensor analytics solutions for Industrial IoT
• Focus areas are risk and performance
• Customers are Fortune-1000 and government
• Real-time visibility using complex event
processing and machine learning algorithms
• Strategic insights using batch analytics
• Hardware Engineers, Data Engineers,
Software Engineers, and Data Scientists
• HQ in Alexandria; offices across world
Some examples of what we do… HARDWAREAPPLICATIONS SERVICESANALYTICS

Our version of Google Now: Parking -> Stationary

Progressive streaming analysis of IoT data: Rules + ML
Times in UTC

Alerting with predictive analytics: Commercial ETA
• 22 hours out
we predicted
driver would be
late (giving
advanced notice)
• That prediction
was < 5 minutes
vs. actual (on a
68-hour trip)
Times in America/New_York

Batch discovery and prescriptive analytics: reducing theft
Third largest transport firm had
2x the median suspect issues

We have fully embraced Apache Spark
Spark is the core of our tech stack:
• Using Spark for batch processing since Spark 1.0, for streaming since Spark 1.2.1
- We use discretized streams (DStreams); our fastest batch interval is 1 second
• 24x7 production operation, with full monitoring and high levels of test coverage
• Supporting Fortune-500 customers, managing billions of dollars of stuff in near real-time
• Fully-automated CI & CD with SOC II certification
• We launch new Spark software several times every week—
Push-button with no visible downtime to customers
• Gives use enormous scale and cost advantages vs. traditional enterprise technologies
• Uptime in last 12 months has been 100%—knock on wood
13 months ago we had a brief outage due to a DNS outage in AWS US-West-2

Spark is at the core of our “hybrid” Lambda architecture
In-house
Analytic
Tools
Sensor
Readers
Mobile
Apps
Enterprise
Data
Open
Data
Partner
Data
Sensor
Meshes
I N T E G R A T I O N L A Y E RA M Q
P
C o A P
F T P
H T T P
M Q T T
S O A P
T C P
U D P
X M P P
S E R V I N G
L A Y E R
B A T C H L A Y E R
S P E E D L A Y E R
Savi IoT
Adapter
Batch
Processing
Domain
Specific
CEP
Sensor
Agnostic
CEP
Modeling,
Machine
Learning
R S - 2 3 2
U S B
p R F I D
B l u e t o o t
h
Z i g B e e
8 0 2 . 1 1
6 L o W P A
N
a R F I D
G S M
G P R S
3 G
4 G / L T E
S A T C O M
Data
Serving
Layer
Notification
s
Savi Apps
Immutable
Data Store
Customer
Export
REST APIs

The Details: Tech stack distributions and versions
Data Applications Tools
 Jetty 9.3.9
 Kafka (0.8.2) via CDH 5.3.3
 Spark (1.4.1 -> 1.6.1)
 Scikit-learn 0.15.2
 Cassandra 2.1.8 via DSE 4.7
 GlusterFS 3.7
 PostgreSQL 9.3.3 with PostGIS
 Hadoop 2.5.0 via CDH 5.3.3
 Hive 0.13.1 via CDH 5.3.3
 Hue 3.7.0 via CDH 5.3.3
 Parquet-format 2.2.0 via Spark
 Parquet-mr 1.6.0 via Spark
 Gobblin 0.7.0
 Drools 6.3.0
 ZooKeeper 3.4.5 via CDH 5.3.3
 Nginx
 Bootstrap
 D3.js, AmCharts, Flot
 WildFly
 Flask
 Shibboleth
 PostgreSQL
 DSE Cassandra
 DSE Solr
 Also mobile on iOS, Android
 Github (Github Flow)
 Ansible
 Docker
 Jenkins
 Maven
 Bower
 Slack
 Fluentd
 Graylog
 Sentry
 Jupyter (PySpark, Folium, Pandas,
Matplotlib, Scikit-learn, etc.)
We program in Scala 2.10, Java 8, Python 2.7, HTML5, LESS.css, and JavaScript
We are hosted in AWS but are not using any AWS-specific solutions (e.g., EMR)

Why we chose Spark
• We started on Apache Storm and MapReduce (we use a Lambda architecture)
• Moved to 100% Spark over the last 18 months (finished last Summer)
• Spark is NOT the best at everything
• However, it is advancing quickly
• We are an analytics company: Spark provides a single unified framework
- Speed layer and Batch Layer
- Use by Engineering and Data Science
- Product apps and ad-hoc analytics
• Ultimately this gives us better agility and cost (development + operations)
For more on our journey see: http://bit.do/savi-spark

Spark Streaming @ Savi
Tips and lessons streaming data 24x7

Spark Streaming is a different animal
20 seconds

 Time is much more precious (and important) in the Streaming world time
- Seconds vs. minutes or hours
- Down-time or interruption is immediately visible to end users—in IoT this can lead to missing key events
- Need to avoid breakdown in stream due to surges or failures—both of which are more common in IoT
 Streaming resource utilization is different than batch
- CPU is rarely the limiting factor
- Memory is less of a limitation than typical for Spark
- I/O is a much more common limiting factor
Some tips and lessons learned managing these differences…
Spark Streaming is a different animal

 Tip 1: Leverage Kafka
- Faster than HDFS, more durable than in-memory
- Supports parallel, independent consumption from multiple processing streams
- Supports FIFO within partitions
 Tip 2: DAG of DAGs (DAG of Streaming Apps and Kafka topics)
- Break down process graph—even near real-time—into critical and non-critical paths
- Route non-critical processing to separate streams, with their own persisted queues
- Do same for interactions with lower-durability sources and targets
 Tip 3: (Caveat to Tip 2) Avoid over-complicating your DAG
- Every time you re-queue: you create opportunities to get data out-of-order
- Instead rely on at-least-once processing and add non-more-than-once protection to non-idempotent processing
Tips to defensively architect Spark Streaming

 Tip 4: Offload bad data to non-blocking paths
- Bad data will happen
- Design your apps to offload this to non-blocking paths (vs. failing)—keeps the stream alive
 Tip 5: (Caveat to Tip 4) Wind-down if infrastructure fails
- Running a streaming process with broken infrastructure will create lots of bad issues
- Instead wind-down (and alert) and allow Kafka to help you recover
- Wind-down and re-start will often “clear up” network or memory bottlenecks
 Tip 6: Preserve data lineage (and immutability)
- Preserve full data lineage of each stage of processing – will save you when dealing with real-world issues
- Keep everything, even failures – this allows you to replay data for analysis, recovery (you will need it)
Tips to defensively architect Spark Streaming (cont.)

 Tip 1: Over-subscribe your cores
- Minimum core count needed is NsourceTopics + 2
- For efficiency over-subscribe your cores.
- High multiples are fine
 Tip 2: Use broadcast variables to
persist shared ephemeral rules
 Tip 3: Limit Kafka topics per App
- Counter-intuitive for defensive programming
- Avoids starvation due to imbalanced loads
 Tip 4: Avoid the shuffle
- Shuffle is tough on I/O,
with streaming it is worse
- Instead rely on Kafka partitioning
- However, Kafka offset partitioning
is still a work-in-progress
Tips: performance tuning Spark Streaming

Streaming Real-world Industrial IoT Data:
“It’s very different than the Canonical Twitter stream analysis teaching example”

All the normal “dirty data” issues plus
Streaming means you have to handle much of this in near-real time
IoT + Spark Streaming = Physical + Data (in near real-time)

 The “IoT Menagerie”
- Millions of source IPs: white listing is impossible
- Many transport protocols and standard: HTTP, FTP, CoAP, MQTT, TCP, UDP, X12, GPRS
 Several tools available to ingest IoT transactions into your platform
- Even some directly to Kafka for processing by Spark, Storm and Flink
 However, not everything is a simple transaction – most is not
 The “obvious” solution: increasing MAX KAFKA SIZE does not work:
- Bottlenecking and serialization issues
- Ultimately will not be able to increase enough
 Lessons Learned: Use hybrid ingestion
- Append critical metadata immediately at point of ingestion
- Includes transaction ID and digital signature
- Split metadata from payload for complex and large data types
- Keeps memory low and is fully scalable
Challenge 1: Ingesting IoT data
S T R E A M I N G D A T A T Y P E S
Micro-
batches
Simple
transactions
Loggers
Sensor
constellations
30% of xacs
10% of data
20% of xacs
35% of data
5% of xacs
10% of data
45% of xacs
15% of data
MIME media
transactions
<1% of xacs
30% of data

Challenge 2: Handing stream interruptions and surges
• Massive increase in stream interruptions
(vs. normal server flows)
- Loss of power
- Movement in and out of coverage
- Bad OTA updates (can cause false DDoS events)
• Often undetected by anyone but Spark
• Overcoming these
- Monitor and alarm on anomalous values
- Tune your fetch rates to avoid overwhelming I/O
- Our Hope: New Spark back-pressure (still in beta)

 Transmission, authentication, and formatting errors are
much more frequent in IoT
- Ever had a cellphone call dropped or duplicate text?
- Data is rarely self-describing
- Firmware configuration management issues
- Standards non-compliance
 Duplication is much more common (and complex)
than traditional Lambda
- Duplicate data can “hide” in unique wrappers
- Duplicate data can be obscured by transaction IDs
- Duplicates can come beyond viably sustainable window durations
 Lessons Learned:
- Accept everything—even authentication errors
- Capture entire lineage of processing (metadata and payload)
- Route failures away from DAG—but preserve to replay and recover
- Map data to based atomic unit THEN digitally sign and de-duplicate data
Challenge 3: Cleansing and transforming IoT data
U N I Q U E T R A N S A C T I O N S E T
Duplicate Facts
(from prior set)
Unique Header
Unique Facts
(to this set)
Unique Header
Incomplete Facts
(in this set)
Unique Header

Transformation and Cleansing example: “Simple” raw IoT data
$690300SR86506702256878020160321155058-
16060a34ST-663-
000p00105300090008000030b74a67db4333102660000470
00a67fb4333102660000330009cd9b433310266000025000
10g81-
00077.09254000038.8064970066.7000002016032115514
2000010006.926480003.01000020e210000000000000001
00e21000000000000000200e2200000000000000055246a8

Transformation and Cleansing: Canonical format for analytics
Turning machine data into cleaned,
self-describing, agnostic data that can
be readily used for analytics and
machine learning
Sensor Message
Universal Read Format

 Streaming data is many, many small files
- 100s or 1000s per second
 Adding to HDFS creates the small file problem
- Many files (Name node swamping)
- Much smaller than HDFS block size (inefficient)
 Delaying too long makes batch analysis stale
- Kafka dues not support complex queries
 Lots of back and forth on this; current best practice:
- Organize streams by volume and type into Kafka topics
- Batch extract by topic based on volume AND time
- Ultimately convert to parquet-format for batch analytics
Challenge 4: The small files problem
And now, the hardest challenge: Streaming CEP…

Challenge 5: CEP processing messy physical realities
Spark Streaming needs to make decisions quick enough to matter…
In the physical world, real-time gets stale very quickly

Sometimes, data just gets lost (or significantly delayed)
0
1
2
3
When streaming the IoT, the time lag of information is ever-present

Also, people have been known to “contradict” sensors
 Sometimes legitimate
 Sometimes mistaken
 Sometimes malicious
People will argue that the
sensors (or rules) are wrong

Finally, once I alert you to something, I cannot undo it
Human memory is not a batch layer: it’s hard to forget Type I errors

 Prioritize Type I bias vs. Type II based on context
 Windowing can be helpful, but not always
- Data can be delayed hours or days (windowing is not cost-effective)
 Use self-healing rule sets (and algorithms)
- Immutable journal data models for state management
- Keep track of multiple time dimensions: latest, most recent
- Keep track of multiple signal dimensions: detected, reported
 Use batch layer to assist with self-healing
- Re-order on review
- Auto-resolve based on new data
 Add a human signals (to build trust)
- Do not hide corrections, make them clear
- Show full time lineage
- Allow human to re-order to understand effects of outages and delays
CEP in IoT: (Timeliness + Good Enough) > (Late + Perfect)
James will dive into this deeper…

 Some challenges to overcome streaming IoT
 Once you overcome these—and share insights with customers—
the the real fun begins. There is lots you can do with Spark
 Questions, ideas, comments:
jhaughwout@savi.com
 Starting to open source some tools at:
https://github.com/sensoranalytics/
 Visit us at 3601 Eisenhower Avenue
Thank you!

Spark Streaming the Industrial IoT

Spark Streaming the Industrial IoT

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Spark Streaming the Industrial IoT (20)

Recently uploaded (20)

Spark Streaming the Industrial IoT