Showing posts with label FLANK. Show all posts
Showing posts with label FLANK. Show all posts

Did the user really ask for Exactly Once? Fault Tolerance

Exactly Once Requirements

It is very tricky and can cause performance degradation, if your user could just use at least once, then always go with that.    Having data sinks like Kudu where you can do an upsert makes exactly once less needed.

Apache Flink, Apache NiFi Stateless and Apache Kafka can participate in that.

For CDF Stream Processing and Analytics with Apache Flink 1.10 Streaming:

Both Kafka sources and sinks can be used with exactly once processing guarantees when checkpointing is enabled.

End-to-End Guaranteed Exactly-Once Record Delivery

The Data Source and Data Sink to need to support exactly-once state semantics and take part in checkpointing.

Data Sources
  • Apache Kafka - must have Exactly-Once selected, transactions enabled and correct driver.

Select:  Semantic.EXACTLY_ONCE

Data Sinks
  • HDFS BucketingSink
  • Apache Kafka


The Rise of the Mega Edge (FLaNK)

At one point edge devices were cheap, low energy and low powered.   They may have some old WiFi and a single core CPU running pretty slow.    Now power, memory, GPUs, custom processors and substantial power has come to the edge.

Sitting on my desk is the NVidia Xaver NX which is the massively powerful machine that can easily be used for edge computing while sporting 8GB of fast RAM, a 384 NVIDIA CUDA® cores and 48 Tensor cores GPU, a 6 core 64-bit ARM CPU and is fast.   This edge device would make a great workstation and is now something that can be affordably deployed in trucks, plants, sensors and other Edge and IoT applications.  

Next that titan device is the inexpensive hobby device, the Raspberry Pi 4 that now sports 8 GB of LPDDR4 RAM, 4 core 64-bit ARM CPU and is speedy!   It can also be augmented with a Google Coral TPU or Intel Movidius 2 Neural Compute Stick.   

These boxes come with fast networking, bluetooth and the modern hardware running in small edge devices that can now deployed en masse.    Enabling edge computing, fast data capture, smart processing and integration with servers and cloud services.    By adding Apache NiFi's subproject MiNiFi C++ and Java agents we can easily integrate these powerful devices into a Streaming Data Pipeline.   We can now build very powerful flows from edge to cloud with Apache NiFi, Apache Flink, Apache Kafka  (FLaNK) and Apache NiFi - MiNiFi.    I can run AI, Deep Learning, Machine Learning including Apache MXNet, DJL, H2O, TensorFlow, Apache OpenNLP and more at any and all parts of my data pipeline.   I can push models to my edge device that now has a powerful GPU/TPU and adequate CPU, networking and RAM to do more than simple classification.    The NVIDIA Jetson Xavier NX will run multiple real-time inference streams at 60 fps on multiple cameras.  

I can run live SQL against these events at every segment of the data pipeline and combine with machine learning, alert checks and flow programming.   It's now easy to build and deploy applications from edge to cloud.

I'll be posting some examples in my next article showing some simple examples.

By next year, 12 or 16 GB of RAM may be a common edge device RAM, perhaps 2 CPUs with 8 cores, multiple GPUs and large fast SSD storage.   My edge swarm may be running much of my computing power as my flows running elastically on public and private cloud scale up and down based on demand in real-time.

Explore Enterprise Apache Flink with Cloudera Streaming Analytics - CSA 1.2

Explore Enterprise Apache Flink with Cloudera Streaming Analytics - CSA 1.2

What's New in Cloudera Streaming Analytics

Try out the tutorials now:

So let's get our Apache Flink on, as part of my FLaNK Stack series I'll show you some fun things we can do with Apache Flink + Apache Kafka + Apache NiFi.

We will look at some of updates in Apache Flink 1.10 including the SQL Client and API.

We are working with Apache Flink 1.10, Apache NiFi 1.11.4 and Apache Kafka 2.4.1.

The SQL features are strong and we will take a look at what we can do.

Table connectors
  • Kafka
  • Kudu
  • Hive (through catalog)

Data formats (Kafka)
  • JSON
  • Avro
  • CSV

Building a DataStream Application in Flink

Build A Flink Project

mvn archetype:generate                               \
      -DarchetypeGroupId=org.apache.flink              \
      -DarchetypeArtifactId=flink-quickstart-java      \


No More Spaghetti Flows

Spaghetti Flows

You may have heard of:   For Apache NiFi, I have seen some (and have done some of them in the past), I call them Spaghetti Flows.

Let's avoid them.   When you are first building a flow it often meanders and has lots of extra steps and extra UpdateAttributes and random routes. This applies if you are running on-premise, in CDP or in other stateful NiFi clusters (or single nodes). The following video from Mark Payne is a must watch before you write any NiFi flows.

Apache NiFi Anti-Patterns with Mark Payne 

Do Not:

  • Do not Put 1,000 Flows on one workspace.

  • If your flow has hundreds of steps, this is a Flow Smell.   Investigate why.

  • Do not Use ExecuteProcess, ExecuteScripts or a lot of Groovy scripts as a default, look for existing processors

  • Do not Use Random Custom Processors you find that have no documentation or are unknown.

  • Do not forget to upgrade, if you are running anything before Apache NiFi 1.10, upgrade now!

  • Do not run on default 512M RAM.

  • Do not run one node and think you have a highly available cluster.

  • Do not split a file with millions of records to individual records in one shot without checking available space/memory and back pressure.

  • Use Split processors only as an absolute last resort. Many processors are designed to work on FlowFiles that contain many records or many lines of text. Keeping the FlowFiles together instead of splitting them apart can often yield performance that is improved by 1-2 orders of magnitude.


  • Reduce, Reuse, Recycle.    Use Parameters to reuse common modules.

  • Put flows, reusable chunks (write to Slack, Database, Kafka) into separate Process Groups.

  • Write custom processors if you need new or specialized features

  • Use Cloudera supported NiFi Processors

  • Use RecordProcessors everywhere

  • Read the Docs!

  • Use the NiFi Registry for version control.

  • Use NiFi CLI and DevOps for Migrations.

  • Run a CDP NiFi Datahub or CFM managed 3 or more node cluster.

  • Walk through your flow and make sure you understand every step and it’s easy to read and follow.   Is every processor used?   Are there dead ends?

  • Do run Zookeeper on different nodes from Apache NiFi.

  • For Cloud Hosted Apache NiFi - go with the "high cpu" instances, such as 8 cores, 7 GB ram.

  • same flow 'templatized' and deployed many many times with different params in the same instance

  • Use routing based on content and attributes to allow one flow to handle multiple nearly identical flows is better than deploying the same flow many times with tweaks to parameters in same cluster.

  • Use the correct driver for your database.   There's usually a couple different JDBC drivers.

  • Make sure you match your Hive version to the NiFi processor for it.   There are ones out there for Hive 1 and Hive 3!   HiveStreaming needs Hive3 with ACID, ORC.

Let's revisit some Best Practices:

Get your Apache NiFi for Dummies.   My own NiFi 101.

Here are a few things you should have read and tried before building your first Apache NiFi flow:

Also when in doubt, use Records!  Use Record Processors and use pre-defined schemas, this will be easier to develop, cleaner and more performant. Easier, Faster, Better!!!

There are record processors for Logs (Grok), JSON, AVRO, XML, CSV, Parquet and more.

Look for a processor that has “Record” in the name like PutDatabaseRecord or QueryRecord.

Use the best DevOps processes, testing and tools.

Some newer features in 1.8, 1.9, 1.10, 1.11 that you need to use.

Advanced Articles:

Spaghetti is for eating, not for real-time data streams.   Let's keep it that way.

If you are not sure what to do check out the Cloudera Community, NiFi Slack or the NiFi docs.   Also I may have a helpful article here. Join me and my NiFi friends at virtual meetups for more in-depth NiFi, Flink, Kafka and more. We keep it interactive so you can feel free to ask questions.

Note:   In this picture I am in Italy doing spaghetti research.

Commonly Used TCP/IP Ports in Streaming

Cloudera CDF and HDF Ports
NiFi and Friends
FLaNK Extended Stack


All of these ports can be changed by administrators or in version updates.   Also if you are running Apache Knox like in Cloudera Data Platform Public Cloud, these ports may be changed or hidden.   This is just based on a version of CDF I am running and defaults in.   This does not include standard Cloudera ports for Cloudera Manager, Hadoop, Atlas, Ranger and other necessary and fun services.

Cloudera Flow Management (CFM Powered by Apache NiFi)
  • Cloudera NiFi HTTP:    8080 or 9090
  • Cloudera NiFi HTTPS:  8443 or 9443
  • Cloudera NiFi RIP Socket: 10443 or 50999
  • Cloudera NiFi Node Protocol: 11443
  • Cloudera NiFi Load Balancing:  6342
  • Cloudera NiFi Registry: 18080
  • Cloudera NiFi Registry SSL: 18433
  • Cloudera NiFi Certificate Authority:  10443

Cloudera Edge Flow Management (CEM Powered by Apache NiFi - MiNiFi)

  • Cloudera EFM HTTP:  10080
  • Cloudera EFM CoAP:  8989

Cloudera Stream Processing (CSP Powered by Apache Kafka)
  • Cloudera Kafka: 9092
  • Cloudera Kafka SSL:  9093
  • Cloudera Kafka Connect:  38083
  • Cloudera Kafka Connect SSL:  38085
  • Cloudera Kafka Jetty Metrics: 38084
  • Cloudera Kafka JMX: 9393
  • Cloudera Kafka MirrorMaker JMX: 9394
  • Cloudera Kafka HTTP Metric: 24042
  • Cloudera Schema Registry Registry: 7788
  • Cloudera Schema Registry Admin: 7789
  • Cloudera Schema Registry SSL:  7790
  • Cloudera Schema Registry Admin SSL:  7791
  • Cloudera Schema Registry Database (Postgresql):  5432
  • Cloudera SRM:  6669
  • Cloudera RPC: 8081
  • Cloudera SRM Rest: 6670
  • Cloudera SRM Rest SSL:  6671
  • Cloudera SMM Rest / UI: 9991
  • Cloudera SMM Manager:  8585
  • Cloudera SMM Manager SSL:  8587
  • Cloudera SMM Manager Admin:  8586
  • Cloudera SMM Manager Admin SSL: 8588
  • Cloudera SMM Service Monitor:  9997
  • Cloudera SMM Kafka Connect:  38083
  • Cloudera SMM Database (Postgresql):  5432

Cloudera Streaming Analytics (CSA Powered by Apache Flink)
  • Cloudera Flink Dashboard:  8082


FLaNK: Low Code Streaming: Populating Kafka Topics with FlinkSQL Joins in Real-Time

FLaNK:  Low Code Streaming:  Populating Kafka Topics with FlinkSQL Joins in Real-Time 


Then I can create my 3 tables.   Two are the source ones to join and the third is the destination for my insert.

INSERT INTO global_sensor_events 
 scada.systemtime ,  
scada.temperaturef , 
scada.pressure , 
scada.humidity , 
scada.lux , 
scada.proximity , 
scada.oxidising , 
scada.reducing , 
scada.nh3 , 
energy.voltage ,
energy.`power` ,

FROM energy,
    scada.systemtime = energy.systemtime;


Assets / Scripts / DDL / SQL

Flink Guide to SQL Joins


Article on Joins


Time Series Analysis - Dataflow

In a first, we joined together for the forces of NYC, New Jersey and Philly to power this meetup.   A huge thanks to John Kuchmek, Amol Thacker and Paul Vidal for promoting and cross running a sweet meetup.   John was an amazing meetup lead and made sure we kept moving.  A giant thanks to Cloudera marketing for helping with logistics and some awesome giveaways!   Hopefully next year's we can do a Cinco De Mayo Taco Feast!  Bill Brooks and Robert Hryniewicz were great help!   And thanks for Cloudera for providing CDP Public Cloud on AWS and CDP-DC on OpenStack for demos, development and general data fun.   And thanks for the initial meetup suggestion and speaker to Bethann Noble and her awesome machine learning people.

Philly - NJ - NYC

To quote, John Kuchmek:

The Internet of Things (IoT) is growing in popularity but it isn’t new. Connected devices have existed in manufacturing and utilities with Supervisory Control and Data Acquisition (SCADA) systems. Time series data has been looked at for sometime in these industries as well as the stock market. Time series analysis can bring valuable insight to businesses and individuals with smart homes. There are many parts and components to be able to collect data at the edge, store in a central location for initial analysis, model build, train and eventually deploy. Time series forecasting is one of the more challenging problems to solve in data science. Important factors in time series analysis and forecasting are seasonality, stationary nature of data and autocorrelation of target variables. We show you a platform, built on open source technology, that has this potential. Sensor data will be collected at the edge, off a Raspberry Pi, using Cloudera’s Edge Flow Manager (powered by MiNiFi). The data will then be pushed to a cluster containing Cloudera Flow Manager (powered by NiFi) so it can be manipulated, routed, and then be stored in Kudu on Cloudera’s Data Platform. Initial inspection can be done in Hue using Impala. The time series data will be analyzed with potential forecasting using an ARIMA model in CML (Cloudera Machine Learning). Time series analysis and forecasting can be applied to but not limited to stock market analysis, forecasting electricity loads, inventory studies, weather conditions, census analysis and sales forecasting.

The main portion of our meetup was an amazing talk by Data Scientist - Victor Dibia.

Analyzing Time Series Data with an ARIMA model

His talk comes right after mine and is about an hour of in-depth Data Science with many hard questions answered.   Also a cool demo.   Thanks again Victor.

We also had some really great attendees who asked some tough question.  My favorite question was by a Flink expert who joined from the West Coast who asked for a FLaNK sticker.

Time Series Analysis - Dataflow

For my small part I did a demo of ingesting data from MiNiFi to NiFi to CML and Kafka.   Flink reads from two Kafka topics, joins them and inserts into a third Kafka topic.   We call the ML model for classification as part of our ingest flow.   This is an example of my FLaNK Stack.

MiNiFi sends the data it reads from sensors and a camera and sends them to a local NiFi gateway.   That NiFi gateway sends a stream to my CDP hosted CFM NiFi cluster for processing.  This cluster splits the data based on which set of sensors (energy or scada) and then publishes to Kafka topics and populates Kudu tables with an UPSERT.

We have great options for monitoring, querying and analyzing our data with the tools from CDP and CDP-DC.   These include Cloudera DAS, Apache Hue, Cloudera SMM for Kafka, Flink SQL console, Flink Dashboard, CML Notebooks, Jupyter Notebooks from CML and Apache Zeppelin.

As a separate way to investigate Kafka, I have created a Hive external table in beeline and connected that to a Kafka topic.  I can know query the current state of that topic.

Video Walkthrough of FlinkSQL Application (and awesome Machine Learning Talk on Time Series)

Slides From Talk

Related Articles

Flink SQL Preview

FLaNK:  Flink SQL Preview

From our Web Flink Dashboard, we can see how our insert is doing and view the joins and records passing quickly through our tiny cluster.

As part of the May 7th, 2020 Virtual Meetup, I was doing some work with Flink SQL to show for a quick demo as the introduction to the meetup and I found out how easy it was to do some cool stuff.   This was inspired by my Streaming Hero, Abdelkrim, who wrote this amazing article on Flink SQL use cases:

As part of our time series meetup, I have a few streams of data coming from one device from a MiNiFi Java agent to NiFi for some transformation, routing and processing and then sent to Apache Flink for final processing.   I decided to join Kafka topics with Flink SQL.   

Let's create Flink Tables:

This table will be used to insert the joined events from both source Kafka topics.

CREATE TABLE global_sensor_events (
 uuid STRING, 
systemtime STRING ,  
temperaturef STRING , 
pressure DOUBLE, 
humidity DOUBLE, 
lux DOUBLE, 
proximity int, 
oxidising DOUBLE , 
reducing DOUBLE, 
nh3 DOUBLE , 
gasko STRING,
`current` INT, 
voltage INT ,
`power` INT,
`total` INT,
fanstatus STRING
) WITH (
'connector.type'    = 'kafka',
'connector.version' = 'universal',
'connector.topic'    = 'global_sensor_events',
'connector.startup-mode' = 'earliest-offset',
'' = '',
'' = 'flink-sql-global-sensor_join',
'format.type' = 'json'

This table will hold Kafka topic messages from our energy reader.

uuid STRING, 
systemtime STRING,  
        `current` INT, 
voltage INT, 
`power` INT, 
`total` INT, 
swver STRING, 
hwver STRING,
type STRING, 
model STRING, 
mac STRING, 
deviceId STRING, 
alias STRING, 
devname STRING, 
iconhash STRING, 
relaystate INT, 
ontime INT, 
activemode STRING, 
feature STRING, 
updating INT, 
rssi INT, 
ledoff INT, 
latitude INT, 
longitude INT, 
`day` INT, 
`index` INT, 
zonestr STRING, 
tzstr STRING, 
dstoffset INT, 
host STRING, 
currentconsumption INT, 
devicetime STRING, 
ledon STRING, 
fanstatus STRING, 
`end` STRING, 
cpu INT, 
memory INT, 
diskusage STRING
) WITH (
'connector.type'    = 'kafka',
'connector.version' = 'universal',
'connector.topic'    = 'energy',
'connector.startup-mode' = 'earliest-offset',
'' = '',
'' = 'flink-sql-energy-consumer',
'format.type' = 'json'

The scada table holds events from our sensors.

uuid STRING, 
systemtime STRING,  
amplitude100 DOUBLE, 
        amplitude500 DOUBLE, 
amplitude1000 DOUBLE, 
lownoise DOUBLE, 
midnoise DOUBLE,
        highnoise DOUBLE, 
amps DOUBLE, 
ipaddress STRING, 
host STRING, 
host_name STRING,
        macaddress STRING, 
endtime STRING, 
runtime STRING, 
starttime STRING, 
        cpu DOUBLE, 
cpu_temp STRING, 
diskusage STRING, 
memory DOUBLE, 
temperature STRING, 
adjtemp STRING, 
adjtempf STRING, 
temperaturef STRING, 
pressure DOUBLE, 
humidity DOUBLE, 
lux DOUBLE, 
proximity INT, 
oxidising DOUBLE, 
reducing DOUBLE, 
nh3 DOUBLE, 
gasko STRING
) WITH (
'connector.type'    = 'kafka',
'connector.version' = 'universal',
'connector.topic'    = 'scada',
'connector.startup-mode' = 'earliest-offset',
'' = '',
'' = 'flink-sql-scada-consumer',
'format.type' = 'json'

This is the magic part:

INSERT INTO global_sensor_events 
scada.systemtime ,  
scada.temperaturef , 
scada.pressure , 
scada.humidity , 
scada.lux , 
scada.proximity , 
scada.oxidising , 
scada.reducing , 
scada.nh3 , 
energy.voltage ,
energy.`power` ,

FROM energy,
    scada.systemtime = energy.systemtime;

So we join two Kafka topics and use some of their fields to populate a third Kafka topic that we defined above.

With Cloudera, it is so easy to monitor our streaming Kafka events with SMM.

For context, this is where the data comes from: