Schedule

We are witnessing a rapid growth in the number of mission-critical data pipelines that leaders of data products are responsible for. “Are your data pipelines healthy?” This question was posed to more than 200 leaders of data products from various industries. The answers ranged from “unfortunately, no” to “they are mostly fine, but I am always afraid that something or the other will cause a pipeline to break”.

This talk presents the concept of Pipeline HealthCheck (PHC) which enables leaders of data products to have high confidence in the correctness, performance, and cost efficiency of their data pipelines. More importantly, PHC enables leaders of data products as well as their development and operations teams to have high confidence in their ability to quickly detect, troubleshoot, and fix problems that make data pipelines unhealthy. The talk also includes a demo of how PHC helps handle common problems in data pipelines like incorrect results, missing SLAs, and overshooting cost budgets.

Apache Airflow is known to be a great orchestration tool that enables use cases that would not be possible otherwise. One of the great features that Airflow has is the possibility to “glue” together totally separate services to establish bigger functionalities. In this talk you will learn about various Airflow usages that let Airflow users to automate their critical company processes and even establish businesses. The examples provided will be based on Airflow used in the context of Cloud Composer which is a managed service to provision and manage Airflow instances.

Airflow has a lot of moving parts, and it can be a little overwhelming as a new user - as I was not too long ago.

Join me as we go though Airflow’s architecture at a high level, explore how DAGs work and run, and look at some of the good, the bad, and the unexpected things lurking inside.

We will describe how we were able to build a system in Airflow for MySQL to Redshift ETL pipelines defined in pure Python using dataclasses. These dataclasses are then used to dynamically generate DAGs depending on pipeline type. This setup allows us to implement robust testing, validation, alerts, and documentation for our pipelines. We will also describe the performance improvements we achieved by upgrading to Airflow 2.0.

This presentation will detail how Elyra creates Jupyter Notebook, Python and R script- based pipelines without having to leave your web browser. The goal of using Elyra is to help construct data pipelines by surfacing concepts and patterns common in pipeline construction into a familiar, easy to navigate interface for Data Scientists and Engineers so they can create pipelines on their own. In Elyra’s Pipeline Editor UI, portions of Apache Airflow’s domain language are surfaced to the user and either made transparent or understandable through the use of tooltips or helpful notes in the proper context during pipeline construction. With these features, Elyra can rapidly prototype data workflows without the need to know or write any pipeline code. Lastly, we will look at what features we have planned on our roadmap for Airflow, including more robust Kubernetes integration and support for runtime specific components/operators. Project Home: https://github.com/elyra-ai/elyra

As the Apache Airflow project grows, we seek both ways to incorporate rising technologies and novel ways to expose them to our users. Ray is one of the fastest-growing distributed computation systems on the market today. In this talk, we will introduce the Ray decorator and Ray backend. These features, built with the help of the Ray maintainers at Anyscale, will allow Data Scientists to natively integrate their distributed pandas, XGBoost, and TensorFlow jobs to their airflow pipelines with a single decorator. By merging the orchestration of Airflow and the distributed computation of Ray, this coordination of technologies opens Airflow users to a whole host of new possibilities when designing their pipelines.

Airflow scheduler uses DAG definitions to monitor the state of tasks in the metadata database, and triggers the task instances whose dependencies have been met. It is based on state of dependencies scheduling. The idea of event based scheduling is to let the operators send events to the scheduler to trigger a scheduling action, such as starting jobs, stopping jobs and restarting jobs. Event based scheduling allows potential support for richer scheduling semantics such as periodic execution and manual trigger at per operator granularity.

Cloudflare’s network keeps growing, and that growth doesn’t just come from building new data centers in new cities. We’re also upgrading the capacity of existing data centers by adding newer generations of servers — a process that makes our network safer, faster, and more reliable for our users.

In this talk, I’ll share how we’re leveraging Apache Airflow to build our own Provision-as-a-Service (PraaS) platform and cut by 90% the amount of time our team spent on mundane operational tasks.

In this talk, we present Viewflow, an open-source Airflow-based framework that allows data scientists to create materialized views in SQL, R, and Python without writing Airflow code.

We will start by explaining what problem does Viewflow solve: writing and maintaining complex Airflow code instead of focusing on data science. Then we will see how Viewflow solves that problem. We will continue by showing how to use VIewflow with several real-world examples. Finally, we will see what the upcoming features of Viewflow are!

Resources:

• Announcement blog post: https://medium.com/datacamp-engineering/viewflow-fe07353fa068
• GitHub repo: https://github.com/datacamp/viewflow

This talk aims to share how Airflow’s secrets backend works, and how users can create their custom secret backends for their specific use cases & technology stack.

At Fivetran, we are seeing many organizations adopt the Modern Data Stack to suit the breadth of their data needs. However, as incoming data sources begin to scale, it can be hard to manage and maintain the environment, with more time spent repairing and reengineering old data pipelines than building new ones. This talk will introduce a number of new Airflow Providers, including the airflow-provider-fivetran, and discuss some of the benefits and considerations we are seeing data engineers, data analysts, and data scientist experience in doing so.

Airflow 2.0 was a big milestone for the Airflow community. However, companies and enterprises are still facing difficulties in upgrading to 2.0.

In this talk, I would like to focus and highlight the ideal upgrade path and talk about

• upgrade_check CLI tool
• separation of providers
• registering connections types
• important 2.0 Airflow configs
• DB Migration
• deprecated feature around Airflow Plugins

We’ve all heard the phrase “data is the new oil.” But really imagine a world where this analogy is more real, where problems in the flow of data - delays, low quality, high volatility - could bring down whole economies? When data is the new oil with people and businesses similarly reliant on it, how do you avoid the fires, spills, and crises?

As data products become central to companies’ bottom line, data engineering teams need to create higher standards for the availability, completeness, and fidelity of their data.

In this session we’ll demonstrate how Databand helps organizations guarantee the health of their Airflow pipelines. Databand is a data pipeline observability system that monitors SLAs and data quality issues, and proactively alerts users on problems to avoid data downtime.

The session will be led by Josh Benamram, CEO and Cofounder of Databand.ai. Josh will be joined by Vinoo Ganesh, an experienced software engineer, system architect, and current CTO of Veraset, a data-as-a-service startup focused on understanding the world from a geospatial perspective.

Join to see how Databand.ai can help you create stable, reliable pipelines that your business can depend on!

The scheduler is the core of Airflow, and it’s a complex beast.

In this session we will go through the scheduler in some detail; how it works; what the communication paths are and what processing is done where.

In this talk we’ll see some real world examples from Firebolt customers demonstrating how Airflow is used to orchestrate operational data analytics applications with large data volumes, while keeping query latency low.

Engineering teams leverage the factory coding pattern to write easy-to-read and repeatable code. In this talk, we’ll outline how data engineering teams can do the same with Airflow by separating DAG declarations from business logic, abstracting task declarations from task dependencies, and creating a code architecture that is simple to understand for new team members. This approach will set analytics teams up for success as team and Airflow DAG sizes grow exponentially.

Data quality has become a much discussed topic in the fields of data engineering and data science, and it has become clear that data validation is absolutely crucial to ensuring the reliability of any data products and insights produced by an organization’s data pipelines. This session will outline patterns for combining three popular open source tools in the data ecosystem - dbt, Airflow, and Great Expectations - and use them to build a robust data pipeline with data validation at each critical step.

Using Airflow as our scheduling framework, we ETL data generated by tens of millions of transactions every day to build the backbone for our reports, dashboards, and training data for our machine learning models. There are over 500 (and growing) such ingested and aggregated tables owned by multiple teams that contain intricate dependencies between one another. Given this level of complexity, it can become extremely cumbersome to coordinate backfills for any given table, when also taking into account all its downstream dependencies, aggregation intervals, and data availability. This talk will focus on how we customized and extended Airflow at Adyen to streamline our backfilling operations. This allows us to prevent mistakes and enable our product teams to keep launching fast and iterating.

In this talk, I’ll describe how you can leverage 3 open-source standards - workflow management with Airflow, EL with Airbyte, transformation with DBT - to build your next modern data stack. I’ll explain how to configure your Airflow DAG to trigger Airbyte’s data replication jobs and DBT’s transformation one with a concrete use case.

At Near we work on TBs of Location data with close to real time modelling to generate key consumer insights and estimates for our clients across the globe. We have hundreds of country specific models deployed and managed through airflow to achieve this goal. Some of the workflows that we have deployed our schedule based, some are dynamic and some are trigger based. In this session I would be discussing some of the workflows that are being scheduled and monitored using airflow and the key benefits and also the challenges that we have faced in our production systems.

Multi-tenant Airflow instances can help save costs for an organization. This talk will walk through how we dynamically assigned roles to users based on groups in Active Directory so that teams would have access to DAGs they created in the UI on our multi-tenant Airflow instance. We created our own custom AirflowSecurityManager class in order to achieve this that ultimately ties LDAP and RBAC together.

In this session we will discuss Amazon Managed Workflows for Apache Airflow (MWAA), how Apache Airflow (and specifically version 2.0) is implemented in the service, best practices for deployment and operations, and the Amazon MWAA team’s commitment to open source usage and contributions.

Apache Airflow aims to speed the development of workflows, but developers are always ready to add bugs here and there.

This talk illustrates a few pitfalls faced while developing workflows at the BBC to build machine learning models. The objective is to share some lessons learned and, hopefully, save others time.

Some of the topics covered, with code examples:

• Tasks unsuitable to be run from within Airflow executors
• Plugins misusage
• Inconsistency while using an operator
• (Mis)configuration
• What to avoid during a workflow deployment
• Consequences of non-idempotent tasks

Learn how to use Airflow’s robust ecosystem of providers to construct secure, high-quality DAGs.

At Snowflake you can imagine we do a lot of data pipelines and tables curating metrics metrics for all parts of the business. These are the lifeline of Snowflake’s business decisions. We also have a lot of source systems that display and make these metrics accessible to end users. So what happens when your data model does not match your system? For example your bookings numbers in salesforce do not match your data model that curates bookings metrics. At snowflake we continued to run into this problem over and over again.

Having this problem we set out to build an infrastructure that would allow users to effortlessly sync the results of their data pipelines with any downstream / upstream system. Allowing us to have a central source of truth in our warehouse. This infrastructure was built on snowflake using airflow and allows a user to begin syncing data with a few details such as model and system to update.

In this presentation we will show you how using airflow and snowflake we are able to use our data pipelines as the source of truth for all systems involved in the business. With this infrastructure we are able to use snowflake models as a central source of truth for all applications used throughout the company. This ensures that any number synced in this way seen by two users is always the same.

Machine Learning models can add value and insight to many projects, but they can be challenging to put into production due to problems like lack of reproducibility, difficulty maintaining integrations, and sneaky data quality issues. Kedro, a framework for creating reproducible, maintainable, and modular data science code, and Great Expectations, a framework for data validations, are two great open-source Python tools that can address some of these problems. Both integrate seamlessly with Airflow for flexible and powerful ML pipeline orchestration. In this talk we’ll discuss how you can leverage existing Airflow provider packages to integrate these tools to create sustainable, production-ready ML models.

If you manage a lot of data, and you’re attending this summit, you likely rely on Apache Airflow to do a lot of the heavy lifting. Like any powerful tool, Apache Airflow allows you to accomplish what you couldn’t before… but also creates new challenges. As DAGs pile up, complexity layers on top of complexity and it becomes hard to grasp how a failed or delayed DAG will affect everything downstream. In this session we will provide a crash course on OpenLineage, an open platform for metadata management and data lineage analysis. We’ll show how capturing metadata with OpenLineage can help you maintain inter-DAG dependencies, capture data on historical runs, and minimize data quality issues.

In recent years, the bioinformatics world has seen an explosion in genomic analysis as gene sequencing technologies have become exponentially cheaper. Tests that previously would have cost tens of thousands of dollars will soon run at pennies per sequence. This glut of data has exposed a notable bottleneck in the current suite of technologies available to bioinformaticians. At Drift Biotechnologies, we use Apache Airflow to transition traditionally on-premise large scale data and deep learning workflows for bioinformatics to the cloud, with an emphasis on workflows and data from next generation sequencing technologies.

The two most common user questions at Pinterest are: 1) why is my workflow running so long? 2) why did my workflow fail - is it my issue, or a platform issue? As with any big data organization, the workflow platform is just the orchestrator but the “real” work is done on another layer, managed by another platform. There can be plenty of these, and the challenges of figuring out the root cause of an issue can be mundane and time consuming. At Pinterest, we set out to provide additional tooling in our Airflow webserver to make it a quicker inspection process and provide smart tips such as increased runtime analysis, bottleneck identifying, rca, and an easy way for backfilling. We explore deeper the tooling provided to reduce the admin load, and empower our users.

Autoscaling in Airflow - what we learnt based on Cloud Composer case. We would like to present how we approach the autoscaling problem for Airflow running in Kubernetes in Cloud Composer: how we calculate our autoscaling metric, what problem we had for scaling down and how did we solve it. Also we share an ideas on what and how we could improve the current solution

After performing several experiments with Airflow, we reached the best architectural design for processing text medical records in scale. Our hybrid solution uses Kubernetes, Apache Airflow, Apache Livy, and Apache cTAKES. Using Kubernetes’ containers has the benefit of having a consistent, portable, and isolated environment for each component of the pipeline. With Apache Livy, you can run tasks in a Spark Cluster at scale. Additionally, Apache cTAKES helps with the extraction of information from electronic medical records clinical free-text by using natural language processing techniques to identify codable entities, temporal events, properties, and relations.

Apache Superset is a modern, open-source data exploration & visualization platform originally created by Maxime Beauchemin. In this talk, I will showcase advanced technical Superset features like the rich Superset API, how to version control dashboards using Github, embedding Superset charts in other applications, and more. This talk will be technical and hands-on, and I will share all code examples I use so you can play with them yourself afterwards!

As people define and publish a DAG, it can be really useful to make it clear how this DAG should behave under different “operating contexts”. Common operating contexts may match your different environments (dev / staging / prod) and/or match your operating needs (quick run, full backfill, test run, …).

Over the years, patterns have emerged around workflow authors, teams and organizations, and little has been shared as to how to approach this. In this talk, we’ll talk about what an “operating context” is, why it’s useful, and describe common patterns and best practices around this topic.

In this talk Jarek and Kaxil will talk about official, community support for running Airflow in the Kubernetes environment.

The full support for Kubernetes deployments was developed by the community for quite a while and in the past users of Airflow had to rely on 3rd-party images and helm-charts to run Airflow on Kubernetes. Over the last year community members made an enormous effort to provide robust, simple and versatile support for those deployments that would respond to all kinds of Airflow users. Starting from official container image, through quick-start docker-compose configuration, culminating in April with release of the official Helm Chart for Airflow.

This talk is aimed for Airflow users who would like to make use of all the effort. The users will learn how to:

• Extend or customize Airflow Official Docker Image to adapt it to their needs
• Run quickstart docker-compose environment where they can quickly verify their images
• Configure and deploy Airflow on Kubernetes using the Official Airflow Helm chart

In Apache Airflow, Xcom is the default mechanism for passing data between tasks in a DAG. In practice, this has been restricted to small data elements, since the Xcom data is persisted in the Airflow metadatabase and is constrained by database and performance limitations.

With the new TaskFlow API introduced in Airflow 2.0, it is seamless to pass data between tasks and the use of Xcom is invisible. However, the ability to pass data is restricted to a relatively small set of data types which can be natively converted in JSON.

This tutorial describes how to go beyond these limitations by developing and deploying a Custom Xcom backend within Airflow to enable the sharing of large and varied data elements such as Pandas data frames between tasks in a data pipeline, using a cloud storage such as Google Storage or Amazon S3.

An informal and fun chat about the journey that we took and the decisions that we made in building Amazon Managed Workflows for Apache Airflow. We will talk about

• Our first tryst with understanding Airflow
• Talking to Amazon Data Engineers and how they ran workflows at scale
• Key design decisions and the reasons behind them
• Road ahead, and what we dream about for future of Apache Airflow.
• Open-Source tenets and commitment from the team

We will leave time at the end for a short AMA/Questions.