[Paper Reading]Orca: A Modular Query Optimizer Architecture for Big Data

Dec 20, 20210 likes147 views

The performance of analytical query processing in data management systems depends primarily on the capabilities of the system's query optimizer. Increased data volumes and heightened interest in processing complex analytical queries have prompted Pivotal to build a new query optimizer. In this paper we present the architecture of Orca, the new query optimizer for all Pivotal data management products, including Pivotal Greenplum Database and Pivotal HAWQ. Orca is a comprehensive development uniting state-of-the-art query optimization technology with own original research resulting in a modular and portable optimizer architecture. In addition to describing the overall architecture, we highlight several unique features and present performance comparisons against other systems.

Paper Reading
Orca: A Modular Query Optimizer Architecture for Big Data
Presented by Weizhen Wang

Background
In this paper we present the architecture of Orca, the new query
optimizer for all Pivotal data management products, including
Pivotal Greenplum Database and Pivotal HAWQ.

ORCA = GPORCA = Greenplum ORCA = ...
Jul 2015, PQO: Pivotal Query Optimizer (Released in GPDB 4.3.5.0)
Jun 2014, ORCA: SIGMOD 2014 Paper
https://github.com/greenplum-db/gporca
In source code GPOPT
src/backend/gpopt
psql: select gp_opt_version();

High level GPDB architecture
• Storage and processing of large amounts of data are handledby distributing the load across several
servers or hosts to create an array of individual databases. all working together to present a single
database image.
• The master is the entry point to GPDB, where clients connect and submit SQL statements. The
master coordinates work with other database instances, called segments, to handle data processing
and storage.
• When a query is submitted to the master, it is optimized and broken into smaller components
dispatched to segments to work together on delivering the final results.
• The interconnect is the networking layer responsible for inter-process communication between the
segments. The interconnect uses a standard Gigabit Ethernet switching fabric.

GPDB Using Planner (`set optimizer=off`)

Orca architecture
DXL： Data exchange
Language

Interaction of Orca with database system

Five Optimization Steps in Orca
● [Step 0] Pre-Process: e.g., predicates pushdown
● [Step 1] Exploration: All equivalent logical plans
● [Step 2] Statistics Derivation: histograms
● [Step 3] Implementation: Logical to Physical
● [Step 4] MPP Optimization: Enforcing and Costing

Step 0 Pre-Process with 25 iterations (TL, DR)

Step 1 Exploration 1/2: Memoization
Compact in-memory data
structure
capturing plan space:
Group: Container of equivalent
expressions
Group Expression: Operator
that has
other groups as its children

Processing optimization requests in the Memo

Parallel Query Optimization
Exp(g): Generate logically equivalent expressions of all group expressions in group g.
Exp(gexpr): Generate logically equivalent expressions of a group expression gexpr.
Imp(g): Generate implementations of all group expressions in group g.
Imp(gexpr): Generate implementation alternatives of a group expression gexpr.
Opt(g, req): Return the plan with the least estimated cost that is rooted by an operator in
group g and satises optimization request req.
Opt(gexpr, req): Return the plan with the least estimated cost that is rooted by gexpr and
satises optimization request req.
Xform(gexpr, t): Transform group expression gexpr using rule t.
Optimization process is broken to small work units called optimization jobs.
Orca currently has seven different types of optimization jobs:

Parallel Query Optimization
Optimization jobs dependency graph

Minimal Repros
AMPERe is a tool for Automatic capture of
Minimal
Portable and Executable Repros.

MEASURING ACCURACY
• Discordance of plan pairs.
• Relevance of plan
• Pairwise distance

PingCAP.com
TAQO components
● Configuration Generator
● Execution Tracker
● Plan Deduplicator
● Ranker

Great Query Performance of Orca vs Planner
(TPC-DS 10TB)

This paper proposes interleaving with coroutines for any type of index join. It showcases the proposal on SAP HANA by implementing binary search and CSB+-tree traversal for an instance of index join related to dictionary compression. Coroutine implementations not only perform similarly to prior interleaving techniques, but also resemble the original code closely, while supporting both interleaved and non-interleaved execution. Thus, this paper claims that coroutines make interleaving practical for use in real DBMS codebases. Paper: http://www.vldb.org/pvldb/vol11/p230-psaropoulos.pdf Follow PingCAP on Twitter: https://twitter.com/PingCAP Follow PingCAP on LinkedIn: https://www.linkedin.com/company/13205484/

[Paper Reading] Steering Query Optimizers: A Practical Take on Big Data Workl...PingCAP

This document discusses methods for optimizing query performance in a query optimizer called Scope by selecting alternative rule configurations. It proposes using rule signatures to group similar queries and generate candidate rule configurations to execute for each group. A learning model is then trained on execution results to select the best configuration for future queries in each group. The goal is to improve upon the default configuration by adapting to workloads and addressing inaccuracies in cardinality estimation that can lead to suboptimal plans.

Generalized Linear Models with H2O Sri Ambati

1) The document discusses generalized linear models (GLM) using H2O. GLM is a well-known statistical method that fits a linear model to predict outcomes. 2) H2O enables distributed, parallel GLM on large datasets with billions of data points. It supports standard GLM features like regularization to prevent overfitting. 3) An example demonstrates predicting flight delays using airline data with 116 million rows. GLM and deep learning models are fit in seconds on H2O using an 8-node cluster.

ADMM-Based Scalable Machine Learning on Apache Spark with Sauptik Dhar and Mo...Databricks

Apache Spark is rapidly becoming the de facto framework for big-data analytics. Spark’s built-in, large-scale Machine Learning Library (MLlib) uses traditional stochastic gradient descent (SGD) to solve standard ML algorithms. However, MlLib currently provides limited coverage of ML algorithms. Further, the convergence of the adopted SGD approach is heavily dictated by issues such as step-size selection, conditioning of the problem and so on, making it difficult for adoption by non-expert end users. In this session, the speakers introduce a large-scale ML tool built on the Alternating Direction Method of Multipliers (ADMM) on Spark to solve a gamut of ML algorithms. The proposed approach decomposes most ML problems into smaller sub-problems suitable for distributed computation in Spark. Learn how this toolkit provides a wider range of ML algorithms, better accuracy compared to MLlib, robust convergence criteria and a simple python API suitable for data scientists – making it easy for end users to develop advanced ML algorithms at scale, without worrying about the underlying intricacies of the optimization solver. It’s a useful arsenal for data scientists’ ML ecosystem on Spark.

Sketching Data with T-Digest In Apache Spark: Spark Summit East talk by Erik ...Spark Summit

Algorithms for sketching probability distributions from large data sets are a fundamental building block of modern data science. Sketching plays a role in diverse applications ranging from visualization, optimizing data encodings, estimating quantiles, data synthesis and imputation. The T-Digest is a versatile sketching data structure. It operates on any numeric data, models tricky distribution tails with high fidelity, and most crucially it works smoothly with aggregators and map-reduce. T-Digest is a perfect fit for Apache Spark; it is single-pass and intermediate results can be aggregated across partitions in batch jobs or aggregated across windows in streaming jobs. In this talk I will describe a native Scala implementation of the T-Digest sketching algorithm and demonstrate its use in Spark applications for visualization, quantile estimations and data synthesis. Attendees of this talk will leave with an understanding of data sketching with T-Digest sketches, and insights about how to apply T-Digest to their own data analysis applications.

Time-evolving Graph Processing on Commodity Clusters: Spark Summit East talk ...Spark Summit

Real-world graphs are seldom static. Applications that generate graph-structured data today do so continuously, giving rise to an underlying graph whose structure evolves over time. Mining these time-evolving graphs can be insightful, both from research and business perspectives. While several works have focused on some individual aspects, there exists no general purpose time-evolving graph processing engine. We present Tegra, a time-evolving graph processing system built on a general-purpose dataflow framework. We introduce Timelapse, a flexible abstraction that enables efficient analytics on evolving graphs by allowing graph-parallel stages to iterate over complete history of nodes. We use Timelapse to present two computational models, a temporal analysis model for performing computations on multiple snapshots of an evolving graph, and a generalized incremental computation model for efficiently updating results of computations.

Ernest: Efficient Performance Prediction for Advanced Analytics on Apache Spa...Spark Summit

Recent workload trends indicate rapid growth in the deployment of machine learning, genomics and scientific workloads using Apache Spark. However, efficiently running these applications on cloud computing infrastructure like Amazon EC2 is challenging and we find that choosing the right hardware configuration can significantly improve performance and cost. The key to address the above challenge is having the ability to predict performance of applications under various resource configurations so that we can automatically choose the optimal configuration. We present Ernest, a performance prediction framework for large scale analytics. Ernest builds performance models based on the behavior of the job on small samples of data and then predicts its performance on larger datasets and cluster sizes. Our evaluation on Amazon EC2 using several workloads shows that our prediction error is low while having a training overhead of less than 5% for long-running jobs.

Exploiting GPU's for Columnar DataFrrames by Kiran LonikarSpark Summit

Kiran Lonikar proposes extending Project Tungsten in Spark SQL to enable parallel execution of DataFrame operations on GPUs. The proposal involves refactoring DataFrames to use a columnar layout and generating OpenCL code for batched execution across columns. Initial results show speedups from GPU execution. Future work includes supporting multi-GPU execution and adapting additional systems like Impala that may be better suited than Spark for GPU integration.

Implementing Near-Realtime Datacenter Health Analytics using Model-driven Ver...Spark Summit

This document discusses using Spark Streaming and GraphX to perform near-realtime analytics on large distributed systems. The authors present a model-driven approach to implement Pregel-style graph processing to handle heterogeneous graphs. They were able to achieve over 100,000 messages per second on a 4 node cluster by using sufficient batch sizes. Implementation challenges included scaling graph processing across nodes, dealing with graph heterogeneity, and hidden memory costs from intermediate RDDs. Lessons learned include the importance of partitioning, testing high availability, and addressing memory sinks.

Inside Apache SystemML by Frederick ReissSpark Summit

1. The document describes the origins and goals of the SystemML project for scalable machine learning. 2. SystemML was created to allow data scientists to write machine learning algorithms in R and automatically compile and optimize them to run efficiently on large datasets in parallel. 3. An example alternating least squares algorithm is shown written concisely in R, while traditional approaches required translating algorithms to other languages like Scala which was error-prone and slowed iteration. SystemML aims to allow the same algorithm to run fast at large scale with the same answer.

Implementation of linear regression and logistic regression on SparkDalei Li

H2O World - GLM - Tomas NykodymSri Ambati

Generalized linear models (GLMs) are a well-known statistical method for fitting linear models to different types of prediction problems. GLMs allow modeling of non-normal distributions and non-identity link functions. Regularization techniques like lasso, ridge, and elastic net are commonly used with GLMs to prevent overfitting. H2O's GLM implementation handles large datasets with many predictors efficiently using distributed and parallel processing. It supports automatic handling of categorical variables, multiple solvers, and regularization including elastic net. GLMs scaled to "big data" problems can be fit in seconds on H2O using all CPU cores.

Tokyo Webmining Talk1Kenta Oono

This document provides an overview and agenda for a tutorial on deep learning implementations and frameworks. The tutorial is split into two sessions. The first session will cover basics of neural networks, common design aspects of neural network implementations, and differences between deep learning frameworks. The second session will include coding examples of different frameworks and a conclusion. Slide decks and resources will be provided on topics including basics of neural networks, common design of frameworks, and differences between frameworks. The tutorial aims to introduce fundamentals of deep learning and compare popular frameworks.

A Scalable Hierarchical Clustering Algorithm Using Spark: Spark Summit East t...Spark Summit

Clustering is often an essential first step in datamining intended to reduce redundancy, or define data categories. Hierarchical clustering, a widely used clustering technique, can offer a richer representation by suggesting the potential group structures. However, parallelization of such an algorithm is challenging as it exhibits inherent data dependency during the hierarchical tree construction. In this paper, we design a parallel implementation of Single-linkage Hierarchical Clustering by formulating it as a Minimum Spanning Tree problem. We further show that Spark is a natural fit for the parallelization of single-linkage clustering algorithm due to its natural expression of iterative process. Our algorithm can be deployed easily in Amazon’s cloud environment. And a thorough performance evaluation in Amazon’s EC2 verifies that the scalability of our algorithm sustains when the datasets scale up.

Surge: Rise of Scalable Machine Learning at Yahoo!DataWorks Summit

Andy Feng discusses Yahoo's use of scalable machine learning for search and advertisement applications with massive datasets and features. Three machine learning algorithms - gradient boosted decision trees, logistic regression, and ad-query vectors - presented challenges of scale that were addressed using Hadoop and YARN across hundreds of servers. Approximate computing techniques like streaming, distributed training, and in-memory processing enabled speedups of 30x to 1000x and scaling to billions of examples and terabytes of data, allowing daily model training. Hadoop and distributed processing on CPU and GPU resources were critical to solving Yahoo's needs for scalable machine learning on big data.

How Machine Learning and AI Can Support the Fight Against COVID-19Databricks

In this session, we show how to leverage CORD dataset, containing more than 400000 scientific papers on COVID and related topics, and recent advances in natural language processing and other AI techniques to generate new insights in support of the ongoing fight against this infectious disease. The idea explored in our talk is to apply modern NLP methods, such and named entity recognition (NER) and relation extraction to article’s abstracts (and, possibly, full text), to extract some meaningful insights from the text, and to enable semantically rich search over the paper corpus. We first investigate how to train NER model using Medical NER dataset from Kaggle, and specialized version of BERT (PubMedBERT) as a feature extractor, to allow automatic extraction of such entities as medical condition names, medicine names and pathogens. Entity extraction alone can provide us with some interesting findings, such as how approaches to COVID treatment evolved with time, in terms of mentioned medicines. We demonstrate how to use Azure Machine Learning for training the model. To take this investigation one step further, we also investigate the usage of pre-trained medical models, available as Text Analytics for Health service on the Microsoft Azure cloud. In addition to many entity types, it can also extract relations (such as the dosage of medicine provisioned), entity negation, and entity mapping to some well-known medical ontologies. We investigate the best way to use Azure ML at scale to score large paper collection, and to store the results.

Enhancements on Spark SQL optimizer by Min QiuSpark Summit

This document summarizes enhancements made to Spark SQL's optimizer including rule-based optimizations and cost-based optimizations. Rule-based optimizations included join condition push down through predicate rewrite and join order adjustment, as well as data volume reduction through column pruning enhancements. Cost-based optimizations leveraged statistics and histograms to select optimal join types, partitions, and join orders. Future work focuses on enumerating the full space of query plans and improving estimation accuracy.

Lazy Join Optimizations Without Upfront Statistics with Matteo InterlandiDatabricks

Modern Data-Intensive Scalable Computing (DISC) systems such as Apache Spark do not support sophisticated cost-based query optimizers because they are specifically designed to process data that resides in external storage systems (e.g. HDFS), or they lack the necessary data statistics. Consequently, many crucial optimizations, such as join order and plan selection, are presently out-of-scope in these DISC system optimizers. Yet, join order is one of the most important decisions a cost-optimizer can make because wrong orders can result in a query response time that can become more than an order-of-magnitude slower compared to the better order.

Google's DremelMaria Stylianou

The document discusses Dremel, an interactive query system for analyzing large-scale datasets. Dremel uses a columnar data storage format and a multi-level query execution tree to enable fast querying. It evaluates Dremel's performance on interactive queries, showing it can count terms in a field within seconds using 3000 workers, while MapReduce takes hours. Dremel also scales linearly and handles stragglers well. Today, similar systems like Google BigQuery and Apache Drill use Dremel-like techniques for interactive analysis of web-scale data.

Deep Dive into Project Tungsten: Bringing Spark Closer to Bare Metal-(Josh Ro...Spark Summit

This document summarizes Project Tungsten, an effort by Databricks to substantially improve the memory and CPU efficiency of Spark applications. It discusses how Tungsten optimizes memory and CPU usage through techniques like explicit memory management, cache-aware algorithms, and code generation. It provides examples of how these optimizations improve performance for aggregation queries and record sorting. The roadmap outlines expanding Tungsten's optimizations in Spark 1.4 through 1.6 to support more workloads and achieve end-to-end processing using binary data representations.

Optimizing Terascale Machine Learning Pipelines with Keystone MLSpark Summit

The document describes KeystoneML, an open source software framework for building scalable machine learning pipelines on Apache Spark. It discusses standard machine learning pipelines and examples of more complex pipelines for image classification, text classification, and recommender systems. It covers features of KeystoneML like transformers, estimators, and chaining estimators and transformers. It also discusses optimizing pipelines by choosing solvers, caching intermediate data, and operator selection. Benchmark results show KeystoneML achieves state-of-the-art accuracy on large datasets faster than other systems through end-to-end pipeline optimizations.

Block Sampling: Efficient Accurate Online Aggregation in MapReduceVasia Kalavri

Accumulo Summit 2016: GeoMesa: Using Accumulo for Optimized Spatio-Temporal P...Accumulo Summit

LocationTech GeoMesa is a project that builds on open-source, distributed databases like Accumulo, HBase, and Cassandra to scale up indexing, querying, and analyzing billions of spatio-temporal data points. GeoMesa uses space-filling curves to index multi-dimensional data in Accumulo, and we'll discuss recent improvements for non-point geometries. Over the two and a half years GeoMesa has been an open-source project, GeoMesa's Accumulo schemas have evolved and our team has had a chance to work through creating and optimizing custom Accumulo iterators. These custom iterators allow for better query performance and interesting aggregations. GeoMesa provides support for distributed processing in Spark via MapReduce input and output formats that extend their Accumulo counterparts. We will discuss the performance benefit gained by reducing the number of default map/Spark tasks created for complex query patterns. The talk will conclude with updates about GeoMesa's integration with Jupyter notebook and improvements to GeoMesa's Spark integration. – Speaker – Dr. James Hughes Mathematician, Commonwealth Computer Research, Inc (CCRi) Dr. James Hughes is a mathematician at Commonwealth Computer Research, Inc. in Charlottesville, Virginia. He is a core committer for GeoMesa which leverages Accumulo and other distributed database systems to provide distributed computation and query engines. He is a LocationTech committer for GeoMesa, SFCurve, and GeoBench. He serves on the LocationTech Project Management Committee and Steering Committee. Through work with LocationTech and OSGeo projects like GeoTools and GeoServer, he works to build end-to-end solutions for big spatio-temporal problems. He holds a PhD in algebraic topology from the University of Virginia. — More Information — For more information see http://www.accumulosummit.com/

Overview of Apache SystemML by Berthold Reinwald and Nakul JindalArvind Surve

Generalized Linear Models in Spark MLlib and SparkRDatabricks

Generalized linear models (GLMs) unify various statistical models such as linear regression and logistic regression through the specification of a model family and link function. They are widely used in modeling, inference, and prediction with applications in numerous fields. In this talk, we will summarize recent community efforts in supporting GLMs in Spark MLlib and SparkR. We will review supported model families, link functions, and regularization types, as well as their use cases, e.g., logistic regression for classification and log-linear model for survival analysis. Then we discuss the choices of solvers and their pros and cons given training datasets of different sizes, and implementation details in order to match R’s model output and summary statistics. We will also demonstrate the APIs in MLlib and SparkR, including R model formula support, which make building linear models a simple task in Spark. This is a joint work with Eric Liang, Yanbo Liang, and some other Spark contributors.

Some experiences for porting application to Intel Xeon PhiMaho Nakata

The document discusses experiences porting applications to Intel Xeon Phi. It provides tips for compiling applications with Intel Composer XE 2013 using the -mmic flag. While some applications like DGEMM require tuning to achieve peak performance, others like Gaussian09 and Povray require patches and multi-step configurations to build for Xeon Phi. There is also an effort underway to port the pkgsrc packaging system to help bring more software packages to the Xeon Phi.

Accumulo and the Convergence of Machine Learning, Big Data, and SupercomputingAccumulo Summit

Machine learning, big data, and simulation challenges have led to a proliferation of computing hardware and software solutions. Hyperscale data centers, accelerators, and programmable logic can deliver enormous performance via a wide range of analytic environments and data storage technologies. Apache Accumulo is a unique technology with the potential to enable all of these fields. Effectively exploiting Accumulo in these fields requires mathematically rigorous interfaces that allow users to focus on their domains. Mathematically rigorous interfaces are at the core MIT Lincoln Laboratory Supercomputing Center (LLSC) and enable the LLSC to deliver Apache Accumulo o thousands of scientists and engineers. This talk discusses the rapidly evolving computing landscape and how mathematically rigorous interfaces are the key to exploiting Apache Accumulo's advanced capabilities. – Speaker – Jeremy Kepner Fellow, MIT Dr. Jeremy Kepner is a MIT Lincoln Laboratory Fellow. He founded the Lincoln Laboratory Supercomputing Center and pioneered the establishment of the Massachusetts Green High Performance Computing Center. He has developed novel big data and parallel computing software used by thousands of scientists and engineers worldwide. He has led several embedded computing efforts, which earned him a 2011 R&D 100 Award. Dr. Kepner has chaired SIAM Data Mining, the IEEE Big Data conference, and the IEEE High Performance Extreme Computing conference. Dr. Kepner is the author of two bestselling books, Parallel MATLAB and Graph Algorithms in the Language of Linear Algebra. His peer-reviewed publications include works on abstract algebra, astronomy, cloud computing, cybersecurity, data mining, databases, graph algorithms, health sciences, signal processing, and visualization. Dr. Kepner holds a BA degree in astrophysics from Pomona College and a PhD degree in astrophysics from Princeton University. — More Information — For more information see http://www.accumulosummit.com/

Java/Scala Lab 2016. Сергей Моренец: Способы повышения эффективности в Java 8.GeeksLab Odessa

16.4.16 Java/Scala Lab Upcoming events: goo.gl/I2gJ4H Зачастую эффективности и оптимизации кода уделяется мало времени при разработке проекта. В результате проблемы с быстродействием возникают уже на стадии финального тестирования или внедрения, когда исправить это не так просто. Приложение или потребляет слишком много ресурсов или работает слишком медленно. Мы рассмотрим основные ошибки, которые допускают Java разработчики при написании кода, а также я дам советы, как сделать ваш код более эффективным. Все примеры в презентации будут подробно прокомментированы.

PRETZEL: Opening the Black Box of Machine Learning Prediction Serving SystemsNECST Lab @ Politecnico di Milano

Machine Learning (ML) models are often composed as pipelines of operators, from “classical” ML operators to pre-processing and featurization operators. Current systems deploy pipelines as "black boxes”, where the same implementation of training is run for inference. This solution is convenient but leaves large room to improve performance and resource usage. This talk presents Pretzel, a framework for deployment of ML pipelines that is inspired to Database Systems: Pretzel inspects and optimizes pipelines end-to-end much like queries, and manages resources common to multiple pipelines such as operators' state. Pretzel is joint work with University of Seoul and Microsoft Research and has recently been presented at OSDI ’18. After the overview, this talk also shows experimental results of Pretzel against state-of-art ML solutions and discusses limitations and extensions.

Pivotal Data Labs - Technology and Tools in our Data Scientist's Arsenal Srivatsan Ramanujam

These slides give an overview of the technology and the tools used by Data Scientists at Pivotal Data Labs. This includes Procedural Languages like PL/Python, PL/R, PL/Java, PL/Perl and the parallel, in-database machine learning library MADlib. The slides also highlight the power and flexibility of the Pivotal platform from embracing open source libraries in Python, R or Java to using new computing paradigms such as Spark on Pivotal HD.

More Related Content

What's hot (20)