Introduction​

As an execution engine for big data processing, MR3 is a distributed system consisting of a single master (called DAGAppMaster) and multiple workers (called ContainerWorkers) running across the network. The master orchestrates the execution of workers and implements all the features required of a distributed system. Workers receive commands from the master and communicate with each other in order to transfer intermediate data. In this way, MR3 tries to maximize the utilization of cluster resources.

Introduction​

In our previous article published in October 2018, we use the TPC-DS benchmark to compare the performance of Hive-LLAP and SparkSQL 2.3.1 included in HDP 3.0.1 along with Hive 3.1.0 on MR3 0.4. In this article, we update the result by testing SparkSQL 2.3.2 included in HDP 3.1.4. As in the previous experiment, we use the TPC-DS benchmark.

Introduction​

In our previous article published in October 2018, we use the TPC-DS benchmark to compare the performance of Hive-LLAP in HDP 3.0.1 (as well as HDP 2.6.4) and Hive 3 on MR3 0.4. We have shown that Hive 3 on MR3 yields consistently higher throughput than Hive-LLAP in concurrency tests, but since then, the performance of Hive-LLAP has improved considerably for concurrent queries. Thus we are interested in the question of how Hive on MR3 compares with Hive-LLAP in the latest lease of HDP.

Introduction​

In our previous article, we use the TPC-DS benchmark to compare the performance of three SQL-on-Hadoop systems: Impala 2.12.0+cdh5.15.2+0, Presto 0.217, and Hive 3.1.1 on MR3 0.6. It uses sequential tests to draw the following conclusion:

• Impala runs faster than Hive on MR3 on short-running queries that take less than 10 seconds.
• For long-running queries, Hive on MR3 runs slightly faster than Impala.
• For most queries, Hive on MR3 runs faster than Presto, sometimes an order of magnitude faster.

Introduction​

Do you trust Hive? Do you trust Presto? Do you trust Impala? Do you trust your SQL system?

Introduction​

In our previous article, we use the TPC-DS benchmark to compare the performance of five SQL-on-Hadoop systems: Hive-LLAP, Presto, SparkSQL, Hive on Tez, and Hive on MR3. As it uses both sequential tests and concurrency tests across three separate clusters, we believe that the performance evaluation is thorough and comprehensive enough to closely reflect the current state in the SQL-on-Hadoop landscape.

Introduction​

We often ask questions on the performance of SQL-on-Hadoop systems:

• How fast is Hive-LLAP in comparison with Presto, SparkSQL, or Hive on Tez?
• As it is an MPP-style system, does Presto run the fastest if it successfully executes a query?
• As it stores intermediate data in memory, does SparkSQL run much faster than Hive on Tez in general?
• What is the best system for running concurrent queries?
• ...

NOTE: This article is superceded by a new article expanded with results of concurrency tests on newer versions of SQL-on-Hadoop systems.​

Introduction​

Hive running on top of MR3 0.2, or Hive-MR3 henceforth, supports LLAP (Low Latency Analytical Processing) I/O. In conjunction with the ability to execute multiple TaskAttempts concurrently inside a single ContainerWorker, the support for LLAP I/O makes Hive-MR3 functionally equivalent to Hive-LLAP. Hence Hive-MR3 can now serve as a substitute for Hive-LLAP in typical use cases.

Introduction​

In order to check if Hive running on top of MR3, or Hive-MR3 henceforth, is ready for production environments, we should test it for performance, stability, and scalability in multi-user environments in which many queries run concurrently. While Hive-on-Tez does a good job in multi-user environments, an analysis of the architecture of Tez reveals that we can further improve its support for multi-user environments by allowing a single DAGAppMaster to manage multiple concurrent DAGs. One of the design goals of MR3 is to overcome this limitation of Tez so as to better support multi-user environments as a new execution engine of Hive.