At Apple, data scientists and engineers are running enormous Spark workloads to deliver amazing cloud services. Apple Cloud Service supports the ever-increasing scale of Spark workloads and resource requirements with great user experience: from code to deployment management, one interface for all compute backends.

In this talk, Aaruna and Zhou would walk through the lessons we learnt and pitfalls encountered for supporting the service at Apple scale - we would share how Apple Cloud Services effectively orchestrate Spark applications, as well as the seamless switchover among different resource managers - be it in Mesos or Kubernetes, private or on-premise infrastructure. We will also cover the monitoring system and how it helps tuning Spark resource requirements with actual execution analysis.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Aggregate (group-by) is one of most important SQL operations in data warehouses. It is required when we want to get aggregated insights from input datasets. Over the last year, we added a series of aggregate optimizations internally at Facebook Spark SQL, and we started to contribute back to Apache Spark recently.

(1).sort aggregate (SPARK-32461): add code generation to improve query performance, replace hash with sort aggregate when child is sorted, etc. (2).object hash aggregate (SPARK-34286): adaptive sort-based fallback based on JVM heap memory usage during query execution. (3).hash aggregate (SPARK-31973): adaptive bypass partial aggregate when aggregate reduction ratio is low. (4).data source aggregate push down (SPARK-34960): aggregate push down to ORC data source by utilizing column statistics (5).files statistics aggregate: aggregate output files (and all columns) statistics distributively when writing query output

we’ll take deep dive of above features and lessons learned.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

"Why did our dashboard break?" "What happened to my data?" "Why is this column missing?" If you've been on the receiving end of these messages (and many others!) from downstream stakeholders, you're not alone. Data engineering teams spend 40 percent or more of their time tackling data downtime, or periods of time when data is missing, erroneous, or otherwise inaccurate, and as data systems become increasingly complex and distributed, this number will only increase. To address this problem, data observability is becoming an increasingly important part of the cloud data stack, helping engineers and analysts reduce time to detection and resolution for data incidents caused by faulty data, code, and operational environments. But what does data observability actually look like in practice? During this presentation, Barr Moses, CEO and co-founder of Monte Carlo, will present on how some of today's best data leaders implement observability across their data lake ecosystem and share best practices for data teams seeking to achieve end-to-end visibility into their data at scale. Topics addressed will include: building automated lineage for Apache Spark, applying data reliability workflows, and extending beyond testing and monitoring to solve for unknown unknowns in your data pipelines.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Presto was originally designed to run interactive queries against data warehouses, but now it has evolved into a unified SQL engine on top of open data lake analytics for both interactive and batch workloads. However, Presto doesn't scale to very large and complex batch pipelines. Presto Unlimited was designed to address such scalability challenges but it didn’t fully solve fault tolerance, isolation, and resource management.

Spark is the tool of choice across the industry for running large scale complex batch ETL pipelines. This motivated the development of Presto On Spark. Presto on Spark runs Presto as a library that is submitted with spark-submit to a Spark cluster. It leverages Spark for scaling shuffle, worker execution, and resource management. It thereby eliminates any query conversion between interactive and batch use cases. This solution helps enable a performant and scalable platform with seamless end-to-end experience to explore and process data.

Many analysts at Intuit use Presto to explore data in the Data Lake/S3 and use Spark for batch processing. These analysts would earlier spend several hours converting these exploration SQLs written for Presto to Spark SQL to operationalize/schedule them as data pipelines. Presto On Spark is now used by analysts at Intuit to run thousands of critical jobs. No query conversion is required here, improved analysts' productivity and empowered them to deliver insights at high speed.

Benefits from session: Attendees will learn about Presto On Spark architecture Attendees will learn when To Use Spark's Execution Engine With Presto Attendees will learn how Intuit runs thousands of presto jobs daily leveraging databricks platform which they can apply to their own work

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

PySpark has rapidly evolved with the momentum of Project Zen introduced in Apache Spark 3.0. We improved error messages, added type hints for autocompletion, implemented visualization, etc. Most importantly, Pandas API on Spark was introduced from Apache Spark 3.2 which exposes the pandas API that runs on Apache Spark, and the Pandas API on Spark has gained a lot of popularity.

In Apache Spark 3.3, the effort of Project Zen continued and PySpark has many cool changes such as more API coverage & faster default index in Pandas API on Spark, datetime.timedelta support, new PyArrow batch interface, better autocompletion, Python & Pandas UDF profiler and new error classification.

In this talk, we will introduce what is new in PySpark at Apache Spark 3.3, and what is next beyond Apache Spark 3.3 with the current effort and roadmap in PySpark.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Using Tensorflow with big datasets has been an impediment for building deep learning models due to the added complexities of running it in a distributed setting and complicated MLOps code, recent advancements in tensorflow 2, and some extension libraries for Spark has now simplified a lot of this. This talk focuses on how we can leverage the best of both Spark and tensorflow to build machine learning and deep learning models using minimal MLOps code letting Spark handle the grunt of work, enabling us to focus more on feature engineering and building the model itself. This design also enables us to use any of the libraries in the tensorflow ecosystem (like tensorflow recommenders) with the same boilerplate code. For businesses like ours, fast prototyping and quick experimentations are key to building completely new experiences in an efficient and iterative way. It is always preferable to have tangible results before putting more resources into a certain project. This design provides us with that capability and lets us spend more time on research, building models, testing quickly, and rapidly iterating. It also provides us with the flexibility to use our choice of framework at any stage of the machine learning lifecycle. In this talk, we will go through some of the best and new features of both spark and tensorflow, how to go from single node training to distributed training with very few extra lines of code, how to leverage MLFlow as a central model store, and finally, using these models for batch and real-time inference.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Many organizations are standardizing on the lakehouse, however, this new architecture poses challenges with an underlying query execution engine for accessing structured and unstructured data. The execution engine needs to provide the performance of a data warehouse and the scalability of data lakes. To ensure optimum performance, the Databricks Lakehouse Platform offers Photon. This next-gen vectorized query execution engine outperforms existing data warehouses in SQL workloads and implements a more general execution framework for efficient processing of data with support of the Apache Spark™ API. With Photon, analytical queries are seeing a 3 to 5x speed increase, with a 40% reduction in compute hours for ETL workloads. In this session, we will dive into Photon, describe its integration with the Databricks Platform and Apache Spark™ runtimes, talk through customer use cases, and show how your SQL and DataFrame workloads can benefit from the performance of Photon.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

The dominant paradigm today for real-time personalized recommendations and personalized search is the retrieval and ranking architecture based on embeddings. It is a fan-out architecture where a single query produces a storm of requests on the backend. A single query will search through millions of items to retrieve hundreds of candidates that are then enriched by a feature store and ranked so only a few recommended items are presented to the user. A search should return in much less than 1 second. Retrieval and ranking architectures need significant infrastructure - an embeddings store and a feature store - to provide both the required scale and real-time performance. In this talk, we will introduce an open-source, scalable retrieval and ranking serving architecture based on open-source technology: Hopsworks Feature Store, OpenSearch, and KServe. We will describe how to build and operate personalized search and recommendation systems using a retrieval model based on a two tower embedding model, and a ranking model gradient boosted trees. We will also show how you can train your embeddings and build your embeddings store index using Hopsworks and Apache Spark.

Attend this session to learn:

• how to to build a scalable, real-time retrieval and ranking recommender system using open-source platforms;
• how to train item/user embedding models and ranking models;
• how to put all these pieces together in an end-to-end solution for training and operating a scalable recommender/search engine.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Data is the key component of Analytics, AI or ML platform. Organizations may not be successful without having a Platform that can Source, Transform, Quality check and present data in a reportable format that can drive actionable insights.

This session will focus on how Capital One HR Team built a Low Cost Data movement Ecosystem that can source data, transform at scale and build the data storage (Redshift) at a level that can be easily consumed by AI/ML programs - by using AWS Services with combination of Open source software(Spark) and Enterprise Edition Hydrograph (UI Based ETL tool with Spark as backend) This presentation is mainly to demonstrate the flexibility that Apache Spark provides for various types ETL Data Pipelines when we code in Spark.

We have been running 3 types of pipelines over 6+ years , over 400+ nightly batch jobs for $1000/mo. (1) Spark on EC2 (2) UI Based ETL tool with Spark backend (on the same EC2) (3) Spark on EMR. We have a CI/CD pipeline that supports easy integration and code deployment in all non-prod and prod regions ( even supports automated unit testing). We will also demonstrate how this ecosystem can failover to a different region in less than 15 minutes , making our application highly resilient.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Training modern Deep Learning models in a timely fashion requires leveraging GPUs to accelerate the process. Ensuring that this expensive hardware is properly utilised and scales efficiently is complex however. All the steps, from data storage and loading through to preprocessing and finally distributing the model training process requires careful thought.

To reduce the cost of training a model, we need to ensure that we are making best use of our hardware resources. Typically, the GPUs that we rely on are memory constrained with much smaller amounts of VRAM being available relative to CPU RAM. As such we will need to leverage a variety of libraries to help ensure that we can keep our GPUs running.

Through the use of libraries like Petastorm to handle the data loading side, PyTorch Lightning and Horovod to handle the model distribution side we can accelerate can leverage commodity spark clusters to accelerate the training process for our Deep Learning Models.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

In general , in-memory pipelines would scale quite well in Spark if we apply the same processing logic to all records. But for Salesforce the major challenge is, we need to apply custom logic specific to a Log Record Type (LRT). The custom logic includes applying different schemas while processing each event. So performing such custom logic specific to LRT , we need to have a mechanism to collect LRT specific data In-Memory such that we can apply custom logic to each collection. We normally get around 50K files in S3 every 5 minutes and there are around 4 billion log events there in 50K files. Creating a DataFrame from 50K files, then group events by LRTs and applying filters per LRT to create a child DataFrame is one approach. One major challenge is that LRT data distribution is very skewed , so we need an efficient in-memory partitioning strategy to distribute the data. Also just applying filters on parent DataFrame will have many child Data frames with empty partitions due to large skew in data distribution and this creates too many empty tasks while processing child DataFrames. So we need to have a Partitioning schema to distribute data and filter by Log Type but not create unnecessary empty partitions in child DataFrames. We also need a scheduling algorithm to process all child DataFrames to utilize cluster efficiency. We have implemented a custom Spark Streaming for reading SQS notifications and then reading new files in S3 which is designed to scale with ingestion volume . This talk will cover how we performed a Spark RangePartition based on Size distribution of the incoming data and applying schema specific transformation logic. This talk will explain various optimizations at various stages of the processing to meet our latency goal.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Apache Kafka in conjunction with Apache Spark became the de facto standard for processing and analyzing data. Both frameworks are open, flexible, and scalable. Unfortunately, the latter makes operations a challenge for many teams. Ideally, teams can use serverless SaaS offerings to focus on business logic. However, hybrid and multi-cloud scenarios require a cloud-native platform that provides automated and elastic tooling to reduce the operations burden.

This post explores different architecture to build serverless Kafka and Spark multi-cloud architectures across regions and continents. We start from the analytics perspective of a data lake and explore its relation to a fully integrated data streaming layer with Kafka to build a modern data lakehouse. Real-world use cases show the joint value and explore the benefit of the "delta lake" integration.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Scribd's data architecture was originally batch-oriented, but in the last couple years, we introduced streaming data ingestion to provide near-real-time ad hoc query capability, mitigate the need for more batch processing tasks, and set the foundation for building real-time data applications.

Kafka and Delta Lake are the two key components of our streaming ingestion pipeline. Various applications and services write messages to Kafka as events are happening. We were tasked with getting these messages into Delta Lake quickly and efficiently.

Our first solution was to deploy Spark Structured Streaming jobs. This got us off the ground quickly, but had some downsides.

Since Delta Lake and the Delta transaction protocol are open source, we kicked off a project to implement our own Rust ingestion daemon. We were confident we could deliver a Rust implementation since our ingestion jobs are append only. Rust offers high performance with a focus on code safety and modern syntax.

In this talk I will describe Scribd's unique approach to ingesting messages from Kafka topics into Delta Lake tables. I will describe the architecture, deployment model, and performance of our solution, which leverages the kafka-delta-ingest Rust daemon and the delta-rs crate hosted in auto-scaling ECS services. I will discuss foundational design aspects for achieving data integrity such as distributed locking with DynamoDb to overcome S3's lack of "PutIfAbsent" semantics, and avoiding duplicates or data loss when multiple concurrent tasks are handling the same stream. I'll highlight the reliability and performance characteristics we've observed so far. I'll also describe the Terraform deployment model we use to deliver our 70-and-growing production ingestion streams into AWS.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

CARTO’s Spatial Extension provides the fundamental building blocks for Location Intelligence in Databricks. Many of the largest organizations using CARTO leverage Databricks for their analytics. Customers very often build custom spatial applications that simplify either a spatial analysis use case or provide a more direct interface to access business intelligence or information. CARTO facilitates the creation of these apps with a complete set of development libraries and APIs. For visualization, CARTO makes use of the powerful deck.gl visualization library. You utilize CARTO Builder to design your maps and perform analytics using Spatial SQL similar to PostGIS, but with the scalability of Apache Spark and then you reference them in your code. CARTO will handle visualizing large datasets, updating the maps, and everything in between. In this talk we will walk you through the process to build spatial applications with CARTO hosted in Apache Spark.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

In this talk, we present two open source projects, Coral and Transport, that enable deep SQL and UDF interoperability between Spark and other engines, such as Trino and Hive. Coral is a SQL analysis, rewrite, and translation engine that enables compute engines to interoperate and analyze different SQL dialects and plans, through the conversion to a common relational algebraic intermediate representation. Transport is a UDF framework that enables users to write UDFs against a single API but execute them as native UDFs of multiple engines, such as Spark, Trino, and Hive. Further, we discuss how LinkedIn leverages Coral and Transport, and present a production use case for accessing views of other engines in Spark as well as enhancing Spark DataFrame and Dataset view schema. We discuss other potential applications such as automatic data governance and data obfuscation, query optimization, materialized view selection, incremental compute, and data source SQL and UDF communication.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

In the past, stream processing over data lakes required a lot of development efforts from data engineering teams, as Itai has shown in his talk at Spark+AI Summit 2019 (https://tinyurl.com/2s3az5td). Today, with Delta Lake and Databricks Auto Loader, this becomes a few minutes' work! Not only that, it unlocks a new set of ways to efficiently leverage your data.

Nexar, a leading provider of dynamic mapping solutions, utilizes Delta Lake and advanced features such as Auto Loader to map 150 million miles of roads a month and provide meaningful insights to cities, mobility companies, driving apps, and insurers. Nexar’s growing dataset contains trillions of images that are used to build and maintain a digital twin of the world. Nexar uses state-of-the-art technologies to detect road furniture (like road signs and traffic lights), surface markings, and road works.

In this talk, we will describe how you can efficiently ingest, process, and maintain a robust Data Lake, whether you’re a mapping solutions provider, a media measurement company, or a social media network. Topics include: * Incremental & efficient streaming over cloud storage such as S3 * Storage optimizations using Delta Lake * Supporting mutable data use-cases with Delta Lake

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Spark applications often need to query external data sources such as file-based data sources or relational data sources. In order to do this, Spark provides Data Source APIs to access structured data through Spark SQL.

Data Source APIs have optimization rules such as filter push down and column pruning to reduce the amount of data that needs to be processed to improve query performance. As part of our ongoing project to provide generic Data Source V2 push down APIs, we have introduced partial aggregate push down, which significantly speeds up spark jobs by dramatically reducing the amount of data transferred between data sources and Spark. We have implemented aggregate push down in both JDBC and parquet.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Join Scott Haines (Databricks Beacon) as he teaches you to write your own Notebook style service (like Jupyter / Zeppelin / Databricks) for both fun (and profit?). Cause haven't we all just been a little curious how Notebook environments work? From the outside things probably seem magical, however just below the surface there is a literal world of possibilities waiting to be exploited (both figuratively and literally) to assist in the building of unimaginable new creations. Curiosity is of course the foundation for creativity and novel ideation, and when armed with the knowledge you'll pick up in this session, you'll have gained an additional perspective and way of thinking (mental model) for solving complex problems using dynamic procedural (on-the-fly) code compilation.

Did I mention you'll use Spark Structured Streaming in order to generate a "live" communication channel between your Notebook service and the "outside world"?

Overview During this session you'll learn to build your own Notebook-style service on top of Apache Spark & the Scala ILoop. Along the way, you'll uncover how to harness the SparkContext to manage, drive, and scale your own procedurally defined Apache Spark applications by mixing core configuration and other "magic". As we move through the steps necessary to achieve this end result, you'll learn to run individual paragraphs, or the entire synchronous waterfall of paragraphs, leading to the dynamic generation of applications.

Deep dive into the world of possibilities that fork from a solid understanding of procedurally generated, on-the-fly, code compilation (live injection), the security ramifications (cause of course this is unsafe!), but come away with a new mental model focused on architecting composite applications, or auto-generated

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

Talk about a challenge of building a scalable framework for data scientists and ML engineers, that could accommodate hundreds of generic or customer specific ML models, running both in streaming and batch, capable of processing 100+ million records per day from social media networks.

The goal has been archived using Spark and Delta. Our framework is built on clever usage of delta features such as change data feed, selective merge and spark structure streaming from and into delta tables. Saving the data in multiple delta tables, where the structure of these tables are reflecting the particular step in the whole flow. This brings great efficiency, as the downstream processing does very little transformations and thus even people without extensive experience of writing ML pipelines and jobs can use the framework easily. At the heart of the framework there is a series of Spark structure streaming jobs continuously evaluating rules and looking for what social media content should be processed by which model. These rules could be updated by the users anytime and the framework needs to automatically adjust the processing. In an environment like this, the ability to track the records throughout the whole process and the atomicity of operations is of utmost importance and delta tables are providing all of this out of the box.

In the talk we are going to focus on the ideas behind the framework and efficient combining of structured streaming and delta tables. Key takeaways would be exploring some of the lesser known delta table features and real-life experiences from building a ML framework solution based on scalable big data technologies, showing how capable and fast such a solution can be, even with minimal hardware resources.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/

From the organization of the tree of life, to the tissues and structures of living organisms: trees and graphs are a recurring data structure in biology. Given the tree-like relationships between biological entities, Knowledge Graphs are emerging as the ideal way to store and retrieve biological data.

In our first Data + AI talk (https://www.youtube.com/watch?v=Kj5bZ2afWSU), we presented the Bellman open source library (https://github.com/gsk-aiops/bellman). Bellman was developed to translate SPARQL queries into Apache Spark Dataset operations so that scientists can submit graph queries in familiar environments like Jupyter and Databricks notebooks.

In this talk, we present the new logical inferencing capabilities we've built into the Bellman OSS library. We will demonstrate how connections between biological entities that are not explicitly connected in the data are deduced from ontologies. These inferred connections are returned to the scientist to aid in the discovery of new connections with the intent on accelerating gene to disease research. To demonstrate these capabilities, we will take a deep dive into the "subclassOf" logical entailment to retrieve all subclasses of a biological entity. The performance characteristics of inference algorithms like forward and backward chaining will also be compared.

Connect with us: Website: https://databricks.com Facebook: https://www.facebook.com/databricksinc Twitter: https://twitter.com/databricks LinkedIn: https://www.linkedin.com/company/data... Instagram: https://www.instagram.com/databricksinc/