Apache Pulsar for ML
Tim Spann | Developer Advocate

AGENDA
● Apache Pulsar
● ML Functions
● Demo

Tim Spann
Developer Advocate
Tim Spann
Developer Advocate at StreamNative
● FLiP(N) Stack = Flink, Pulsar and NiFi Stack
● Streaming Systems & Data Architecture Expert
● Experience:
○ 15+ years of experience with streaming technologies including Pulsar, Flink, Spark, NiFi, Big
Data, Cloud, MXNet, IoT, Python and more.
○ Today, he helps to grow the Pulsar community sharing rich technical knowledge and experience
at both global conferences and through individual conversations.

This week in Apache Flink, Apache
Pulsar, Apache NiFi, Apache Spark
and open source friends.
https://bit.ly/32dAJft
FLiP Stack Weekly

Apache Pulsar
https://pulsar.apache.org/docs/en/reference-terminology/

What are the Benefits of Pulsar?
Data Durability
Scalability Geo-Replication
Multi-Tenancy
Unified Messaging
Model

Apache Pulsar is a Cloud-Native
Messaging and Event-Streaming Platform.

CREATED
Originally
developed inside
Yahoo! as Cloud
Messaging
Service
GROWTH
10x Contributors
10MM+ Downloads
Ecosystem Expands
Kafka on Pulsar
AMQP on Pulsar
Functions
. . .
2012 2016 2018 TODAY
APACHE TLP
Pulsar
becomes
Apache top
level project.
OPEN SOURCE
Pulsar
committed
to open source.
Apache Pulsar Timeline

Evolution of Pulsar Growth

Pulsar Has a Built-in Super Set of OSS
Features
Durability
Scalability Geo-Replication
Multi-Tenancy
Unified Messaging
Model
Reduced Vendor Dependency
Functions
Open-Source Features

Apache Pulsar is built to support legacy applications, handle the
needs of modern apps, and supports NextGen applications
Support legacy workloads.
Compatible with popular
messaging and streaming tools.
Legacy
Built for today's real-time
event driven applications.
Modern
Scalable, adaptive architecture
ready for the future of real-time
streaming.
NextGen

Apache Pulsar has a vibrant community
560+
Contributors
10,000+
Commits
7,000+
Slack Members
1,000+
Organizations
Using Pulsar

It is often assumed that Pulsar and Kafka have equal capabilities. In reality,
Pulsar offers a superset of Kafka.
● Pulsar is streaming and queuing together
● Pulsar is cloud-native with stateless brokers
● Natively includes geo-replication, multi-tenancy, and end-to-end
security out of the box
● Pulsar provides automated rebalancing
● Pulsar offers 100X lower latency w/ 2.5 greater throughput than Kafka
Advantages of Apache Pulsar

Apache Pulsar features
Cloud native with decoupled
storage and compute layers.
Built-in compatibility with your
existing code and messaging
infrastructure.
Geographic redundancy and high
availability included.
Centralized cluster management
and oversight.
Elastic horizontal and vertical
scalability.
Seamless and instant partitioning
rebalancing with no downtime.
Flexible subscription model
supports a wide array of use cases.
Compatible with the tools you use
to store, analyze, and process data.

● “Bookies”
● Stores messages and cursors
● Messages are grouped in
segments/ledgers
● A group of bookies form an
“ensemble” to store a ledger
● “Brokers”
● Handles message routing and
connections
● Stateless, but with caches
● Automatic load-balancing
● Topics are composed of
multiple segments
●
● Stores metadata for both
Pulsar and BookKeeper
● Service discovery
Store
Messages
Metadata &
Service Discovery
Metadata &
Service Discovery
Pulsar Cluster
Metadata
Storage
Pulsar Cluster

Component Description
Value /
Data payload
The data carried by the message. All Pulsar messages contain raw bytes,
although message data can also conform to data schemas.
Key Messages are optionally tagged with keys, used in partitioning and also is useful
for things like topic compaction.
Properties An optional key/value map of user-defined properties.
Producer name The name of the producer who produces the message. If you do not specify a
producer name, the default name is used.
Sequence ID Each Pulsar message belongs to an ordered sequence on its topic. The
sequence ID of the message is its order in that sequence.
Messages - the Basic Unit of Apache Pulsar

Different subscription modes
have different semantics:
Exclusive/Failover -
guaranteed order, single active
consumer
Shared - multiple active
consumers, no order
Key_Shared - multiple active
consumers, order for given key
Producer 1
Producer 2
Pulsar Topic
Subscription D
Consumer D-1
Consumer D-2
Key-Shared
<
K
1,
V
10
>
<
K
1,
V
11
>
<
K
1,
V
12
>
<
K
2
,V
2
0
>
<
K
2
,V
2
1>
<
K
2
,V
2
2
>
Subscription C
Consumer C-1
Consumer C-2
Shared
<
K
1,
V
10
>
<
K
2
,V
2
1>
<
K
1,
V
12
>
<
K
2
,V
2
0
>
<
K
1,
V
11
>
<
K
2
,V
2
2
>
Subscription A Consumer A
Exclusive
Subscription B
Consumer B-1
Consumer B-2
In case of failure in
Consumer B-1
Failover
Apache Pulsar Subscription Modes

Streaming
Consumer
Consumer
Consumer
Subscription
Shared
Failover
Consumer
Consumer
Subscription
In case of failure in
Consumer B-0
Consumer
Consumer
Subscription
Exclusive
X
Consumer
Consumer
Key-Shared
Subscription
Pulsar
Topic/Partition
Messaging

Unified Messaging Model
Simplify your data infrastructure and
enable new use cases with queuing and
streaming capabilities in one platform.
Multi-tenancy
Enable multiple user groups to share the
same cluster, either via access control, or
in entirely different namespaces.
Scalability
Decoupled data computing and storage
enable horizontal scaling to handle data
scale and management complexity.
Geo-replication
Support for multi-datacenter replication
with both asynchronous and
synchronous replication for built-in
disaster recovery.
Tiered storage
Enable historical data to be offloaded to
cloud-native storage and store event
streams for indefinite periods of time.
Apache Pulsar Benefits

Messaging Use Cases Streaming Use Cases
Service x commands service y to make some
change.
Example: order service removing item from
inventory service
Moving large amounts of data to another service
(real-time ETL).
Example: logs to elasticsearch
Distributing messages that represent work
among n workers.
Example: order processing not in main “thread”
Periodic jobs moving large amounts of data and
aggregating to more traditional stores.
Example: logs to s3
Sending “scheduled” messages.
Example: notification service for marketing emails
or push notifications
Computing a near real-time aggregate of a message
stream, split among n workers, with order being
important.
Example: real-time analytics over page views
Messaging vs Streaming

Messaging Use Case Streaming Use Case
Retention The amount of data retained is
relatively small - typically only a day
or two of data at most.
Large amounts of data are retained,
with higher ingest volumes and
longer retention periods.
Throughput Messaging systems are not designed
to manage big “catch-up” reads.
Streaming systems are designed to
scale and can handle use cases
such as catch-up reads.
Differences in Consumption

byte[] msgIdBytes = // Some byte
array
MessageId id =
MessageId.fromByteArray(msgIdBytes);
Reader<byte[]> reader =
pulsarClient.newReader()
.topic(topic)
.startMessageId(id)
.create();
Create a reader that will read from
some message between earliest and
latest.
Reader
Apache Pulsar Reader Interface

● New Consumer type added in Pulsar 2.10 that provides a
continuously updated key-value map view of compacted topic data.
● An abstraction of a changelog stream from a primary-keyed table,
where each record in the changelog stream is an update on the
primary-keyed table with the record key as the primary key.
● READ ONLY DATA STRUCTURE!
Apache Pulsar TableView

Schema Registry
schema-1 (value=Avro/Protobuf/JSON) schema-2 (value=Avro/Protobuf/JSON) schema-3
(value=Avro/Protobuf/JSON)
Schema
Data
ID
Local Cache
for Schemas
+
Schema
Data
ID +
Local Cache
for Schemas
Send schema-1
(value=Avro/Protobuf/JSON) data
serialized per schema ID
Send (register)
schema (if not in
local cache)
Read schema-1
(value=Avro/Protobuf/JSON) data
deserialized per schema ID
Get schema by ID (if
not in local cache)
Producers Consumers
Schema Registry

● Utilizing JSON Data with a JSON Schema
● Consistency, Contracts, Clean Data
● This enables easy SQL:
○ Pulsar SQL (Presto SQL)
○ Flink SQL
○ Spark Structured Streaming
Use Schemas

• Functions - Lightweight Stream
Processing (Java, Python, Go)
• Connectors - Sources & Sinks
(Cassandra, Kafka, …)
• Protocol Handlers - AoP (AMQP), KoP
(Kafka), MoP (MQTT)
• Processing Engines - Flink, Spark,
Presto/Trino via Pulsar SQL
• Data Offloaders - Tiered Storage - (S3)
Sources, Sinks and Processing

Kafka on Pulsar (KoP)

MQTT on Pulsar (MoP)

AMQP on Pulsar (AoP)

Use Apache Pulsar For Ingest

Use Apache Pulsar To Stream to Lakehouses

Pulsar SQL
Presto/Trino workers can
read segments directly
from bookies (or
offloaded storage) in
parallel.
Bookie
1
Segment 1
Producer Consumer
Broker 1
Topic1-Part1
Broker 2
Topic1-Part2
Broker 3
Topic1-Part3
Segment 2 Segment 3 Segment 4 Segment X
Segment 1
Segment 1 Segment 1
Segment 3 Segment 3
Segment 3
Segment 2
Segment 2
Segment 2
Segment 4
Segment 4
Segment 4
Segment X
Segment X
Segment X
Bookie
2
Bookie
3
Query
Coordinator
...
...
SQL Worker SQL Worker SQL Worker
SQL Worker
Query
Topic
Metadata

● Event-Driven services that are interested in events from other
services can use Pulsar subscriptions to receive new events in
the stream.
● Pulsar supports both linear (exclusive) and filtered subscriptions
(key-shared).
● You can also use StreamNative’s filtering & routing service to
provide filtered subscriptions.
Subscriptions

Functions

● Visual Question and Answer
● Natural Language Processing
● Sentiment Analysis
● Text Classification
● Named Entity Recognition
● Content-based
Recommendations
• Predictive
Maintenance
• Fault Detection
• Fraud Detection
• Time-Series
Predictions
• Naive Bayes
Apache Pulsar Functions for ML Models

Benefits of Pulsar Functions
• Allows you to focus on the business logic.
• Sets up producer/consumer for you, eliminating the need for
boilerplate code.
• Handles message consumption and publication for you.
• Run as threads, processes or Kubernetes Pods inside Pulsar.
No need for another processing framework. Can scale up
automatically.

Event-Based Microservices Application
• A basic order entry use case
for a food delivery service
will involve several different
microservices.
• The communicate via
messages sent to Pulsar
topics.
• A service can also subscribe
to the events published by
another service.

Function Mesh
Pulsar Functions, along with Pulsar
IO/Connectors, provide a powerful API for
ingesting, transforming, and outputting data.
Function Mesh, another StreamNative
project, makes it easier for developers to
create entire applications built from sources,
functions, and sinks all through a declarative
API.

ML
• Visual Question and Answer
• NLP (Natural Language Processing)
• Sentiment Analysis
• Text Classification
• Named Entity Recognition
• Content-based Recommendations
• Predictive Maintenance
• Fault Detection
• Fraud Detection
• Time-Series Predictions
• Naive Bayes

Why Pulsar Functions for Microservices?
Desired Characteristic Pulsar Functions…
Highly maintainable and testable Are small pieces of code written in popular
languages such as Java, Python, or Go. They can be
easily maintained in source control repositories and
tested with existing frameworks automatically.
Loosely coupled with other
services
Are not directly linked to one another and
communicate via messages.
Independently deployable Are designed to be deployed independently
Can be developed by a small team Are often developed by a single developer.
Inter-service Communication Support all message patterns using Pulsar as the
underlying message bus.
Deployment & Composition Can run as individual threads, processes, or K8s
pods. The Function Mesh allows you to deploy
multiple Pulsar Functions as a single unit.

Built-in
Back
Pressure
Producer<String> producer = client.newProducer(Schema.STRING)
.topic("hellotopic")
.blockIfQueueFull(true) // enable blocking
// when queue is full
.maxPendingMessages(10) // max queue size
.create();
// During Back Pressure: the sendAsync call blocks
// with no room in queues
producer.newMessage()
.key("mykey")
.value("myvalue")
.sendAsync(); // can be a blocking call

ML Java Coding (Deep Java Library)

● Lightweight computation similar
to AWS Lambda.
● Specifically designed to use
Apache Pulsar as a message
bus.
● Function runtime can be
located within Pulsar Broker.
● Java Functions
A serverless event
streaming framework
Pulsar Functions

● Consume messages from one or
more Pulsar topics.
● Apply user-supplied processing
logic to each message.
● Publish the results of the
computation to another topic.
● Support multiple programming
languages (Java, Python, Go)
● Can leverage 3rd-party libraries
to support the execution of ML
models on the edge.
Pulsar Functions

Simple Native Java Function

Simple Function with Pulsar SDK

from pulsar import Function
from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
import json
class Chat(Function):
def __init__(self):
pass
def process(self, input, context):
fields = json.loads(input)
sid = SentimentIntensityAnalyzer()
ss = sid.polarity_scores(fields["comment"])
row = { }
row['id'] = str(msg_id)
if ss['compound'] < 0.00:
row['sentiment'] = 'Negative'
else:
row['sentiment'] = 'Positive'
row['comment'] = str(fields["comment"])
json_string = json.dumps(row)
return json_string
Entire Function
Pulsar Python NLP Function

Starting a Function - Distributed Cluster
Once compiled into a JAR, start a Pulsar Function in a distributed cluster:

Pulsar Functions
● Route
● Enrich
● Convert
● Lookups
● Run
Machine Learning
● Logging
● Auditing
● Parse
● Split
● Convert

Pulsar Functions
● Consume messages from one
or more Pulsar topics.
● Apply user-supplied
processing logic to each
message.
● Publish the results of the
computation to another topic.
● Support multiple programming
languages (Java, Python, Go)
● Can leverage 3rd-party
libraries to support the
execution of ML models on the
edge.

Building Real-Time Apps Requires a Team

NLP Streaming Architecture

Learn More

Deploying AI With an
Event-Driven
Platform
https://dzone.com/trendreports/enterprise-ai-1

Apache Pulsar in Action
http://tinyurl.com/bdha5p4r
Please enjoy David’s complete book which is the ultimate guide to Pulsar.

Now Available
On-Demand
Pulsar Training
Academy.StreamNative.io

https://streamnative.io/blog/engineering/2021-11-17-building-edge-applications-with-apache-pulsar/

Github
● https://github.com/tspannhw/FLiP-Py-ADS-B
● https://github.com/tspannhw/pulsar-adsb-function
● https://github.com/tspannhw/pulsar-weather-function
● https://github.com/tspannhw/spring-pulsar-airquality
● https://github.com/tspannhw/pulsar-airquality-function
● https://github.com/tspannhw/FLiP-Current22-LetsMonitorAllTheThings
● https://github.com/tspannhw/FLiP-Transit
● https://github.com/tspannhw/SmartTransit
● https://github.com/tspannhw/airquality-kafka-consumer

Scan the QR code
to learn more about
Apache Pulsar and
StreamNative.

Scan the QR code
to build your own
apps today.

Tim Spann
Developer Advocate
@PaaSDev
https://www.linkedin.com/in/timothyspann
https://github.com/tspannhw
Let’s Keep in Touch