DevConf.US 2019

As IT operations become more agile and complex, at the same time the need to enhance operational efficiency and intelligence grows. Monitoring applications and kubernetes clusters with Prometheus has become quite common. Yet identifying relevant metrics and thresholds for your setup is getting harder.

In this talk, Marcel will show the tooling used to collect and store metrics gathered by Prometheus for the long term. Then analyze those on a large scale for extracting trends and seasonality but also forecasting of expected values for a given metric. Finally, he will integrate the predicted metrics back into the Prometheus monitoring and alerting stack to enable dynamic thresholding and anomaly detection.

The overwhelming majority of data collected by enterprises is unlabeled data. Understanding hidden structure in data is therefore a central task for machine learning. While the field is still not nearly as mature and structured as supervised learning ("predictive analytics"), there are multiple conceptual frameworks and techniques that can be used to attack such problems. This talk introduces some deep-learning based techniques that have shown promising performance.

Used by Facebook, Netflix, Twitter, Uber, Lyft, and many others, Presto has become a ubiquitous solution for running fast SQL analytics across disparate data sources. Presto is an open source distributed SQL query engine widely recognized for its low-latency queries, high concurrency, and native ability to query multiple data sources. These data sources may include Ceph, S3, Google Cloud Storage, Azure Storage, Hadoop/HDFS, relational database systems such as PostgreSQL, and non-relational systems such as Apache Kafka. Presto’s connector based architecture allow you to query virtually anything.
In the first part of the talk we will focus on what Presto is, its background, and its architecture. In the second part of the talk we will learn about Presto’s cloud native capabilities using Red Hat OpenShift and Kubernetes. Kubernetes reduces the burden and complexity of configuring, deploying, managing, and monitoring containerized applications. To achieve these capabilities with Presto, Red Hat and Starburst partnered to provide the Presto Kubernetes Operator and the Presto Container on OpenShift.
The talk will also include a live demo of using Presto on Kubernetes and showing SQL federation of data between Ceph and other data sources. After the talk, you will have a good understanding of Presto basics and ready to participate in the Presto open source community as well as try Presto on your own.

As data is exponentially growing in organizations, there is an increasing need to consolidate silos of information into a single source of truth, a 'Data Lake' to feed hungry Analytics and Machine Learning Engines that can gather insight at scale. In this talk, we will provide an overview of the Open Data Hub architecture. Then we will highlight a current data science use case leveraging Red Hat OpenShift, Ceph Storage, and analytics with Spark and SKLearn on JupyterHub.

Data scientists and machine learning (ML) engineers rely heavily on creating workflows to train, verify, and deploy ML models. Argo is a cloud-native workflow management tool that enables the creation of sophisticated native Kubernetes ML workflows. Kubeflow is an ML toolkit that incorporates best-of-breed ML projects (for example, TensorFlow) as well as providing important infrastructure such as hyperparameter tuning. The Open Data Hub (ODH) is a scalable data lake platform that provides tools such as distributed Spark and Ceph data store. In this presentation we will explore workflow features available using Argo running as a component of Kubeflow and integrated with ODH. Presentation will include a live demonstration of ML workflows using Kubeflow, Spark and Ceph.

Machine learning is more complex than traditional software development. We all know that fetching the right parameters for tuning a machine learning model is quintessential in finding the optimal model.

Have multiple machine learning models with different parameters running, but do not have a platform to track each of these models? MLFlow is the new open source platform which allows data scientists, product engineers, developers and large organizations to manage the ML lifecycle, including experimentation, reproducibility and deployment.

Attend this talk to learn how to:
1. Use MLFlow for experiment tracking
2. Perform hyperparameter model tuning with MLFlow
3. Run, compare and visualize the metrics/parameters of multiple runs of your machine learning model with MLFlow

SEUSS is custom operating system we spliced into the backend of a high throughput distributed Serverless platform, Apache OpenWhisk. SEUSS uses an alternative isolation mechanism to containers, Library Operating Systems (LibOSs). LibOSs enable a lightweight snapshotting technique. Snapshotting LibOSs enables two counterintuitive results: 1) although LibOSs inherently replicate system state, SEUSS can cache multiplicatively more functions on a node; 2) although LibOSs can suffer bad "first run" performance, SEUSS is able to reduce cold start times by orders of magnitude. By increasing sharing and decreasing deterministic bringup, SEUSS radically reduces the amount of hardware and cycles required to run a FaaS platform.

As log data continues to grow, is it possible to automate analyzing logs to find anomalies and errors quickly with machine learning? Log anomaly detection is at its core an unsupervised natural language processing (NLP) Machine Learning problem that can be difficult to validate and tune in a production setting. However, if there was a system that could incorporate minimal user feedback during model training, we could offset some of these challenges.

In this talk, you will learn about the lessons learned from building log anomaly detection system in production giving specific emphasis to the technical challenges faced with, scalability, implementing a human-in-the-loop ML system and Integrating ETL with unsupervised ML to detect anomalies in application logs

The Open Data Hub (ODH) is a scalable data lake platform for data scientists and their machine learning (ML) activities. We have an internal deployment of ODH that processes RHEL build logs and test results. It aggregates and replicates the data from internal and upstream CI systems. AIOps then uses ML techniques to analyze the data, specifically for anomaly detection. The central component is Apache Kafka deployed across multi data centers on a common OpenShift platform which ensures high availability of the core of the ODH system. The presenters will discuss this use case where Kafka is deployed as a caching tool in a highly available and horizontally scalable architecture running on OpenShift, processing hundreds of gigabytes of log data per minute and thousands of messages per second.

A Developer or Quality Engineer (QE) performs a multitude of software engineering tasks like searching lengthy documentations such as functional specs in order to get condensed information, sifting through logs to triage bugs or to find anomalies, analyzing bugs to determine if they are duplicate or not, analyzing test failure logs to determine false positives, etc. In this talk, will detail AI-Library, an open source machine learning framework built on OpenShift, that contains well known machine learning models and solutions to common software engineering use cases. We will show how developers or QE can leverage this framework to solve software engineering problems in an intelligent and automated manner thereby increasing their productivity.

Streaming platforms have emerged as a popular, new trend, but what exactly is a streaming platform? Part messaging system, part Hadoop made fast, part fast ETL and scalable data integration, with Apache Kafka at the core, streaming platforms offer an entirely new perspective on managing the flow of data. This talk will explain what a streaming platform such as Apache Kafka is and some of the use cases and design patterns around its use. Moreover, this talk will also present and answer a set of random -- but recurring -- questions from the community about Apache Kafka.

Elegant solutions often synthesize known facts with a small amount of new effort. To date, we've struggled to get computing machines involved in much elegant problem solving. Considering constraints like budget caps and the polar ice caps, this lack of elegance becomes more than an aesthetic issue.

In this talk I'll present ASC and SEUSS, two systems designed to reduce this new effort. ASC, a Harvard/BU collaboration, attempts to auto-parallelize single threaded workloads, reducing new effort required from programmers to achieve wall clock speedup.

SEUSS is custom operating system we spliced into the backend of a high throughput distributed Serverless platform, Apache OpenWhisk. SEUSS uses an alternative isolation mechanism to containers, Library Operating Systems (LibOSs). LibOSs enable a lightweight snapshotting technique. Snapshotting LibOSs enables two counterintuitive results: 1) although LibOSs inherently replicate system state, SEUSS can cache multiplicatively more functions on a node; 2) although LibOSs can suffer bad ""first run"" performance, SEUSS is able to reduce cold start times by orders of magnitude. By increasing sharing and decreasing deterministic bringup, SEUSS radically reduces the amount of hardware and cycles required to run a FaaS platform.

In this talk we will learn how OpenShift can support data science and ML workflows for Tensorflow application developers.We will learn how to use OpenShift to be more productive while training and deploying tensorflow models.We will look at various model formats, tools available for developers and patterns useful for both Data scientists and devops engineers.

ML systems are well-positioned to analyze natural language workloads and discover actionable insights in a DevOps environment. However, this presents a challenge when it comes to narrowing down to a system which can generalize to a wide variety of use cases. We discuss the limitations encountered in use of sentiment analysis across various artifacts in a DevOps environment. Also, we show how we improved the service through continuous learning and evolved the system to learn, adapt and produce desirable outcomes for a multitude of use cases. We share the lessons learned while going through the process and demonstrate usage of the framework with examples. Through the presentation, developers will quickly learn how to leverage and implement sentiment analysis into their existing environment.

Have a great idea for a data science experiment but don't have the hardware to run it? The MOC and Red Hat have partnered to deploy the Open Data Hub into the MOC giving you access to hardware and support required for leading edge experiments.

The MOC IaaS platform combined with OpenShift and current data science development tools provides you with an alternative to using public clouds to execute your experiments.

Attend this talk to learn about:
- What the Massachusetts Open Cloud and Open Cloud Exchange is
- Current projects running in the MOC
- Running your project in the MOC

Project ChRIS is an open source initiative of Boston Children's Hospital, Boston University, Mass Open Cloud and Red Hat to democratize access to and development of medical image processing software and leverage OpenShift based containers to optimize performance. Today, processing a single set of images takes about ten hours. Our goal is to reduce image-processing time by enabling GPU hardware acceleration and comparing the performance between CPU and GPU based computing environments. We'll share some strategies for tuning the performance of workloads.
Learn about Project ChRIS here:
https://red.ht/2D6XNx7
And code here:
https://bit.ly/2FXMSGJ
No limit on attendees
Familiarity with OpenShift, Tensorflow helpful but not required

The human brain has many capabilities thanks to its network structure that allows transferring information among neurons in order to perform a specific action or multiple actions. Using Machine Learning (ML) we are focusing on the learning capability, but the brain uses its network also to store information in the memory.
Nowadays many fields (e.g. medicine, biology, security, space) rely on graph structures to store data with semantic and context specific to that domain. Graph neural networks (GNNs) allow machines to learn from this kind of structure, taking a little step closer to mimic the behavior and architecture of the human brain.
In this talk, we will explore this type of neural networks that take graphs as input, showing their capabilities, issues and applications.