Cloud Native Computing Foundation KubeCon + CloudNativeCon Europe 2021, 15 May 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Sponsored Session: Dell Technologies - Real-Time Object Detection with Pravega and Flink

Description

Don’t miss out! Join us at our upcoming event: KubeCon + CloudNativeCon North America 2021 in Los Angeles, CA from October 12-15. Learn more at https://kubecon.io The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.

Sponsored Session: Dell Technologies - Real-Time Object Detection with Pravega and Flink

Speakers: David Adams

See a real time object detection application with streaming data built on Pravega and Apache Flink. It’s built with streaming data from enterprise-grade Dell EMC Streaming Data Platform with long term storage on the recently announced Kubernetes-based object storage, Dell EMC ObjectScale. The entire solution is deployed on vSphere with Tanzu on a single turnkey cloud solution, VMware Cloud Foundation on VxRail.

A

Hi, I'm david adams, a senior principal tech, marketing engineer with dell technologies cloud and with me are ashish batwara engineering head streaming data platform and andre keady senior principal software engineer today, we'll be demoing, a real-time object, detection application, which utilizes data streams from dell emc streaming data platform and tiers long-term storage to s3, compatible object storage with the recently announced dell emc object scale currently in early access all deployed on dell technologies cloud platform with vsphere with tonzu enabled this demo solution is made up of four key technologies.

A

Dell technologies cloud platform: dtcp delivers a turnkey experience. That's easy to deploy and manage due to the tight integration between vmware cloud foundation and vxrail with dtcp organizations can now develop test and run cloud native applications alongside virtualized applications in a single platform.

A

Vsphere with tonzu is the re-architecting of vsphere to embed kubernetes into the control plane of vsphere to unify access to compute storage and networking. It allows a user to deploy pods directly into vsphere, known as a vsphere pod or create developer, managed tanzu kubernetes grid clusters, which run in virtual machines. We will be deploying both of these. In this demo, dell emc object scale is a re-engineered object.

A

Storage platform, which takes advantage of kubernetes native automation for deployment scaling and management capabilities with rich s3 compatibility and self-service apis developer can quickly spin up object, storage containers to fuel everything from big data and analytics applications to ephemeral, devtest sandboxes in this case, we'll be using it as the long-term storage for our streaming application.

A

The dell emc streaming data platform takes the best of open source streaming data software in prevega and apache flink and pulls it together in a production grade supported platform sdp provides the reliable repeatable platform for edge and core solutions.

A

Next, let's take a look at the basic architectural view for what we've deployed for this demo solution. First, we've deployed dtcp then enabled vsphere with tonsu.

A

Next we deploy object scale as vsphere pods in its own namespace in the supervisor cluster and create an object store for streaming data platform. Then we created a namespace and tkg clusters for streaming data platform and jupyterhub.

A

Then we deployed sdp, jupiter hub and then our streaming application for object, detection.

A

Let's quickly review the infrastructure deployment in vsphere before we move on here, we can see that our dell emc object scale service has been enabled under supervisor services, and so our vsphere plugin has been enabled. We can look here under object stores to see that we've already created an object store. We can either create new stores here or view our existing ones.

A

This is this object store that we're using for our streaming data platform, long term storage. You can see here that we have several buckets and we can also manage users kubernetes resources for this deployment certificates, events and health checks. On the left hand, side you can see that it's all deployed in a dell emc object scale system namespace as vsphere pods, taking a look at our streaming data platform namespace, where we've deployed prevega, apache, flink and all the other resources you'll see that we have multiple tkg clusters that we've deployed here.

A

We have a cluster for our jupiter hub deployment, as well as a streaming data platform cluster, which has highly available masternodes and multiple worker nodes where we've deployed the streaming data platform, application which will be reviewed later on in the demo.

A

Now I'm going to hand it over to ashish to give some more detail on the streaming data platform architecture.

B

Thank you david. uh So let me start with the dell emc streaming data platform, so dell emc streaming data platform is a modern, analytic platform that solves the problem of ingesting storing and analyzing real time and historical streaming data all with enterprise, skill and production support. Streaming. Data platform is built using community developed, open source software components, for example, it uses kubernetes as the orchestration layer pravega, which is a core key component for streaming.

B

Data platform is a streaming storage that simplifies the development of streaming applications by unifying the concept of historical and real-time data, while providing powerful production abilities such as exactly once, consistency and injection, auto scaling, etc.

B

The main the main other component is a plug-in architecture to support modern real-time analytics engines such as apache flink, also available in the upcoming release, apache spark and provider search, which is kind of an elastic search like engine that also allows for real-time streaming queries of your unstructured log data all on the same stp platform.

B

Hence reducing the need for separate pipeline and hardware by taking the best of open source and pulling it together in a production grid, supported platform. Sdp provides the reliable, secure, manageable and repeatable platform for edge and core solutions streaming. Data platform has a few key features: number one unified data analytics for both real time and historical data, so that data scientists and developers write code once that deals with all types of data without worrying about independent batch and stream processing.

B

In other words, developers don't have to code differently against the live of the wired data versus historical files of data. Now batch iot events and pure byte stream use cases all can co-exist on the same platform. The second key feature is a dvr-like injection and playback capabilities.

B

That stp has that allows ingestion of historical and real-time data using the single injection pipeline, unlike two existing solutions that require one solution for real-time data and another one for different batch processing for the historical data for sdp users can go back in time and play the historical stream alongside the real time stream.

B

The third key feature is that the support for two tier long term stories, an architecture that allows industry standard storage systems such as ice salon and ecs, to be used for the long term storage. The configuration allows users to configure data retention period or until the size grows, to a specific limit.

B

The fourth key feature is the enterprise grade security, which is of prime importance when multiple business units leverages on the same platform, instead of every business unit, building their own independent solutions and, last but not least, to the secure and scalable multi-tenant development platform, a platform that can be used by multiple independent business units in comparison to existing solutions that require each business unit to stay, stand their own system for the analytics.

B

So that means sdp has necessary access control to ensure that data is secure. With streaming data platform in place, organizations can now ingest real-time streaming data and work with it in ways that they have never even imagined. We are seeing. Customers solve some amazing problems just by having this access to the depth and breadth of their sea of data and being able to analyze it. However, they need to with each. Let's take a look at our next section where andre is demonstrating. Real-Time object, detection using streaming data platform.

C

Hi everyone today, I'm going to show you the object: detection demo using streaming data platform running on vmware tenzoo.

C

My persona in this demo is data scientist, I'm a part of a team development, advanced driver assistant system. We continuously collect video and other sensor data from the fleet of test vehicles.

C

There is a three stage of this experiment. First, we use the proviga grpc connector and jupyter notebook to push the video frames into a prevega stream. Proviga provides a python and java client for the grpc connector. Here in jupiter hub, we are using the python client.

C

The data will be stored in the proviga stream that I call the raw video for object. Detection. I created a flink application.

C

The object detection model runs as a fling job in stp it uses the yolo object, detection model which stands for you only look ones. The fling job uses java binding for tensorflow as each video frame is ingested. The object detection model is used to detect objects such as cars, bosses and people, and it will add labeled boxes to the video frames around detected objects.

C

It will also enrich the metadata with the detected object list and the detection accuracy and finally, using the same connector, the grpc.

C

I can pull the events from the object: detector, output, video stream in prevega and play the video in jupiter notebook and now, let's move to the actual demo, we're starting at the sdp ui first going to log in.

C

And here, under the analytics tab, I can create a project.

C

This will create a namespace for my provega streams in my flink cluster.

C

Note that, for this example, we are using the object, object store available on ttcp, which was reviewed earlier in this demo.

C

To the left side, I can create a flink cluster for my application. This can be created using this ui or the helm to deploy the flink cluster size that we need for the flink application.

C

Stp provides a local maven repo, where we can upload the application needed for this project.

C

Now I can run the flink application for object, detection that reads from the raw video stream process. The data and push the processed video to a new stream called object: detector output, video on our vega tab. Here we can see all the work all the streams to build this demo. First, we can start with the raw video and the output of the processing goes into the stream.

C

From the jupiter notebook from the jupiter hub, I created two notebooks, one called the adas ingest and that will push the data into provega stream.

C

I'm going to go ahead and start it.

C

The second notebook will use the grpc to read the video and display it here with the object, detection.

C

Both those notebooks are needed for this demo, one for the ingestion and one to read the data processed by.

B

C

Flink application and now let's go back to our sdp ui from here- we're going to look at the proviga streams.

C

We should be able to see the data start to get in the stream here and, as the data is coming into the vega stream, we're seeing the flink application reading the data and processing it.

C

So below that here is the flink job reading from the stream processing and writing this data to the object: detector output, video stream.

C

If you go back to the scope, we should see the data again written to this stream here, as you can see,.

C

On the screen here.

C

Here is the enrich raw video after running the object detection job back to the jupiter hub.

C

Now, let's test the new object detection model on the video collected that video that raw video has been continuously ingested into the scp, and now I'm going to play the demo.

C

As you can see in this picture, there's frames around the detected objects, people, buses, cars.

C

Bicycles and it's running live as we are ingesting the data, the object, detection runs, the model, detect the objects and we can process it and display it, as you can see here with this. This concludes our demo for the object detection using the streaming data platform running on vmware 10 zoom. Thank.

C

C