わずか5日間で Kafka スキルをレベルアップ | ストリーミングシーズンに参加
This blog post is part three of a three-part series on how businesses can drive value using real-time data streaming and processing technologies. Check out parts one and two for more insights.
From fraud prevention, autonomous vehicles, voice assistants, and intelligent cybersecurity systems that protect our networks, to recommendation engines, dynamic pricing, and predictive maintenance, streaming analytics infused with AI and machine learning (ML) is driving the world around us and helping us make time sensitive decisions with greater situational awareness.
Today, AI and ML have become mainstream in the business world. Streaming AI/ML leverages dynamic features — “what’s happening right NOW” — on flowing data to make contextually relevant predictions. Businesses across every domain are leveraging AI technology for operational efficiency, competitive advantage, and better user experience. AI/ML applied on streaming data becomes powerful when applied to the right data at the right time and the right place.
Check out our one-stop destination for real-time AI resources.
Streaming AI/ML is the application of AI/ML models to continuously flowing data. Next-gen AI/ML-based use cases on streaming data become competitive differentiators for businesses, allowing them to analyze data, discover patterns and similarities, and continually adapt, adjust, and counter data and model drift (Note: model drift occurs when an ML model’s predictive power gradually decays as input data and relationships between variables change over time).
“Data streaming is the central nervous system for data, while AI/ML algorithms are the brain.”
Data streaming platforms are essential to modern AI/ML models, as they are the best way to power these models with real-time, trusted data across the enterprise. They are capable of decoupling data producers and consumers with abstractions and removing underlying complexities. This is a boon for AI/ML engineers because it allows for scaling of development across teams, while freeing them from wrestling with associated subtleties and complexities of streaming. While models provide logic to generate features from streaming data and compute predictions, the streaming pipeline helps with data ingestion, processing, and delivering results to business processes and applications to incorporate the predictions, classifications, and recommendations.
Today, every enterprise needs to proactively make critical business decisions before it’s too late by leveraging real-time streaming data. Systems fueled by streaming data can detect fraud before it occurs, predict customer churn, and detect critical infrastructure failures before they happen. Fusing data streaming and AI models removes end-to-end latency to respond quickly to changing events. There are two types of streaming AI/ML systems.
Systems that make predictions in real time, aka ‘Online Predictions’
Systems that incorporate new data to update models in real time, aka ‘Continual Learning’ (essential for managing drifts)
Such systems typically have two main capabilities:
They possess a streaming pipeline to process streaming data and compute features before feeding it to the AI/ML model. This includes data ingestion from sources like IoT devices, mobile and web apps, stream processing engines like Flink to invoke AI/ML models and compute insights with low latency, and data streaming platforms to enable incremental model training by incorporating new data patterns as detected. This allows ML algorithms to adapt to changes in data distribution in production.
They are capable of online inferencing and decisioning by applying algorithms to the flowing data to generate real-time updates, reports, and alerts
The application of streaming data to AI/ML is particularly useful in time-sensitive use cases spanning domains such as finance, healthcare, and transportation, where time is of the essence. Some notable use cases include:
Analyzing customer interactions in real time, to improve customer experience and provide personalized recommendations
Monitoring and analyzing the performance of industrial equipment to improve operational efficiency and reduce downtime
Preventing fraudulent activities in real time to minimize financial losses by predicting market prices and powering algorithmic trading
Detecting anomalous patterns from edge IOT devices that provide preventive and predictive maintenance with efficient decisioning and forecasting
Allowing the supply chain and transportation industries to monitor traffic and weather to recommend route changes based on real-time data and optimize driving time to reduce fuel consumption
Enhancing customer 360 with real-time data from phone call records, emails, texts, social media posts, clickstreams, POS, and geospatial technology
“The results of AI/ML models need to be available for consumption less than a minute after the event is received.”
This is a common expectation of predictive use cases fueled by data streaming and AI, because businesses often need to make critical, time-sensitive decisions. For certain use cases, it doesn’t matter how good the models’ predictions are; if it takes even a few milliseconds to generate results, the insights become useless — this is known as the “time value of data.” Applying AI/ML to streaming data for training and inference has multiple challenges, including:
Ensuring end-to-end latency to maintain system responsiveness, calculate real-time features, and serve up near-instant insights
Managing high throughput of end user-facing applications honoring SLA latencies
Controlling infrastructure costs to overcome the above two challenges and keep costs low to achieve a healthy ROI
Maturing the skill sets of data teams to build, deploy, and maintain streaming systems with AI/ML capabilities in production
Data streaming systems are complex to design, implement, test, maintain, and troubleshoot, especially with high throughput, high-volume data, and real-time SLAs. Combine that with the complexities of developing, testing, and deploying AI/ML applications and you end up with a scenario that most organizations would back away from due to the enormous challenges involved. Data streaming systems are also distributed and almost always stateful, which exponentially increases the complexity of development, testing, and deployment. Real-time AI/ML pipelines need fast event processing capable of executing complex computations in real time with quick serving.
Confluent Cloud is Confluent’s fully managed, cloud-native data streaming platform that simplifies the integration of streaming data into AI/ML models by abstracting away technical complexity and allowing businesses to:
Embed ML models in event streams to reduce latency by containerizing the compute for stream processing and inferencing, and removing external dependency for a model server. Models can be deployed and/or embedded into an Apache Kafka® application built on top of Confluent Cloud with Kafka-native stream processing, KSQL (user defined functions), or Kafka API in any language binding.
Perform simulation and analysis by incorporating real-time data into ML models
Enable asynchronous communication with streaming, serving as the backbone for integrating AI/ML applications on data streams
Support deployment of AI/ML models at the edge with a lightweight C/C++ edge library — librdkafka. Confluent also supports embedding AI/ML models on mobile applications with JavaScript and REST Proxy, or WebSocket integration.
Empower engineering teams to incorporate streaming into product features with client-side libraries, without needing to reason about how the infrastructure works
Monitor infrastructure and model-specific performance and accuracy metrics of models
Support autoscaling, observability, and automatic model updates
Develop a real-time pipeline in a notebook and test it on streaming data
Confluent Cloud, with its fully managed Apache Flink® service, provides high performance, serverless functions for processing streaming events, supporting aggregations, grouping, joins, sliding window calculation, and predefined libraries for common tasks such as geocoding and date time conversion. Flink optimizes resource consumption and provides low latency, and robust fault tolerance with incremental checkpointing.
Confluent Cloud provides a robust messaging and data streaming platform for incorporating AI/ML capabilities into streaming data, simplifying the end-to-end process. It is the central nervous system for data integration, stream processing, model deployment, monitoring, and embedding models into a real-time stream. It lowers the operational barriers that can arise when maturing AI/ML projects and makes it considerably easier for AI and ML engineers to embed state of the art algorithms and frameworks into their cloud service to build solutions that are simpler, secure, scalable, and performant.
The ability to provide real-time predictions for user-facing applications is becoming an ever-important requirement and will become increasingly critical as more and more AI/ML models are applied closer to user products. Real-time AI/ML is becoming a strategic differentiator, as it increasingly becomes important to incorporate real-time context from streaming data into the decision-making process. Data streaming platforms provide the necessary foundations and scaffolding to enable AI/ML developers to efficiently leverage and train existing models for use in production.
Resources
Head over to Confluent’s data streaming resources hub for the latest explainer videos, case studies, and industry reports on data streaming.
Interested in learning more about how AI is better with data streaming? Check out our one-stop destination for real-time AI resources.
This blog explores how cloud service providers (CSPs) and managed service providers (MSPs) increasingly recognize the advantages of leveraging Confluent to deliver fully managed Kafka services to their clients. Confluent enables these service providers to deliver higher value offerings to wider...
With Confluent sitting at the core of their data infrastructure, Atomic Tessellator provides a powerful platform for molecular research backed by computational methods, focusing on catalyst discovery. Read on to learn how data streaming plays a central role in their technology.