Real-Time Stream Processing, Monitoring, and Analytics With Apache Kafka


Manage episode 341243534 series 2510642
By Confluent, founded by the original creators of Apache Kafka® and Founded by the original creators of Apache Kafka®. Discovered by Player FM and our community — copyright is owned by the publisher, not Player FM, and audio is streamed directly from their servers. Hit the Subscribe button to track updates in Player FM, or paste the feed URL into other podcast apps.

Processing real-time event streams enables countless use cases big and small. With a day job designing and building highly available distributed data systems, Simon Aubury (Principal Data Engineer, Thoughtworks) believes stream-processing thinking can be applied to any stream of events.
In this episode, Simon shares his Confluent Hackathon ’22 winning project—a wildlife monitoring system to observe population trends over time using a Raspberry Pi, along with Apache Kafka®, Kafka Connect, ksqlDB, TensorFlow Lite, and Kibana. He used the system to count animals in his Australian backyard and perform trend analysis on the results. Simon also shares ideas on how you can use these same technologies to help with other real-world challenges.
Open-source, object detection models for TensorFlow, which appropriately are collected into "model zoos," meant that Simon didn't have to provide his own object identification as part of the project, which would have made it untenable. Instead, he was able to utilize the open-source models, which are essentially neural nets pretrained on relevant data sets—in his case, backyard animals.

Simon's system, which consists of around 200 lines of code, employs a Kafka producer running a while loop, which connects to a camera feed using a Python library. For each frame brought down, object masking is applied in order to crop and reduce pixel density, and then the frame is compared to the models mentioned above. A Python dictionary containing probable found objects is sent to a Kafka broker for processing; the images themselves aren't sent. (Note that Simon's system is also capable of alerting if a specific, rare animal is detected.)

On the broker, Simon uses ksqlDB and windowing to smooth the data in case the frames were inconsistent for some reason (it may look back over thirty seconds, for example, and find the highest number of animals per type). Finally, the data is sent to a Kibana dashboard for analysis, through a Kafka Connect sink connector.

Simon’s system is an extremely low-cost system that can simulate the behaviors of more expensive, proprietary systems. And the concepts can easily be applied to many other use cases. For example, you could use it to estimate traffic at a shopping mall to gauge optimal opening hours, or you could use it to monitor the queue at a coffee shop, counting both queued patrons as well as impatient patrons who decide to leave because the queue is too long.

264 episodes