War of the Hadoop SQL engines. And the winner is …?
You may have wondered why we were quiet over the last couple of weeks? Well, we locked ourselves into the basement and did some research and a couple of projects and PoCs on Hadoop, Big Data, and distributed processing frameworks in general. We were also looking at Clickstream data and Web Analytics solutions. Over the next couple of weeks we will update our website with our new offerings, products, and services. The article below summarises some of our research on using SQL on Hadoop.
I believe that one of the pre-requisites for Hadoop to make inroads into the Enterprise Data Warehouse space is to have the following three items in place: (1) Subsecond response times for SQL queries (often refered to as interactive or real time queries). Performance similar to existing MPP RDBMS such as Teradata. (2) Support for a rich SQL feature set (3) Support for Update and Delete DML operations. Currently, I don’t see any of the existing solutions ticking all of these boxes. However, we are getting closer and closer. The post will shed some light on the current status of SQL on Hadoop and my own recommendations, which of these solutions you should bet your house on.
Initially developed by Facebook, Hive is the original SQL framework on Hadoop. The motivation to develop Hive was to provide an abstraction layer on top of Map Reduce (M/R) to make it easier for analysts and data scientists to query data on the Hadoop File System. Rather than write hundreds of lines of Java code to get answers to relatively simple questions the objective was to offer SQL, the natural choice of the data analyst. While this approach works well in a batch oriented environment it does not perform well for interactive workloads in near real time. The problem with the original M/R framework was that it works in stages and at each stage the data is set down to disk and then again read from disk in the next phase. In addition the various stages can not be parallelized. This is highly inefficent and the rationale for the Apache Tez project. Similar to M/R, Tez is a Hive execution engine developed by Hortonworks (also committers from Facebook, Microsoft, and Yahoo).
Hive on Apache Tez
Tez is part of the Stinger initiative led by Hortonworks to make Hive enterprise ready and suitable for realtime SQL queries. The two main objectives of the initiative were to increase performance and offer a rich set of SQL features such as analytic functions, query optimization, and standard data types such as timestamp etc. Tez is the underlying engine that creates more efficient execution plans in comparison to Map Reduce. The Tez design is based on research done by Microsoft on parallel and distributed computing. The two main objectives were delivered as part of the recent Hive 0.13 release. The roadmap for release 0.14 includes DML functionality such as Updates and Inserts for lookup tables.
Hive on Spark
Recently, Cloudera together with MapR, Intel, and Databricks spearheaded a new initiative to add a third execution engine to the mix. They propose to add Spark as a third Hive execution engine. Developers then will be able to choose between Map Reduce, Tez, and Spark as their execution engine for Hive. Based on the design document the three engines will be fully interchangeable and compatible. Cloudera see Spark as the next generation distributed processing engine, which has various advantages over the Map Reduce paradigm, e.g. intermediate resultsets can be cached in memory. Going forward, Spark will underpin many of the components in the Cloudera platform. The rationale for Hive on Spark then is to make Spark available to the vast amount of Hive users and establish Hive on the Spark framework. It will also allow users to run faster Hive queries without having to install Tez. Contrary to Hortonworks, Cloudera don’t see Hive on Spark (or Hive on Tez) to be suitable as a realtime SQL query engine. Their flagship product for interactive SQL queries is Impala, while Databricks see Spark SQL as the tool of choice for realtime queries.
Impala is a massively parallel SQL query engine. It is based on Google Dremel and Google Big Query.
Based on their own benchmarks Cloudera conclude that Presto and Hive Tez are not fit for purpose for interactive query loads. Cloudera see Hive as a batch processing engine. Of course, Hortonworks see this differently and they believe that Hive on Tez is also useful for interactive queries. The jury is out on this one.
You can install and test Impala as part of the Cloudera distribution. Cloudera have been accused of using Impala as vehicle to lock customers into their own distribution. However, you can also download Impala from GitHub.
. Alternatively, you can add the Cloudera repository and download it from there. Impala has various per-requisites in terms of the libraries and respective versions it supports, e.g. it relies (as of this post’s writing) on Hive 0.12 and Hadoop 2.3. So if you want to install it on the latest Hortonworks distro, you are out of luck.
Facebook was and is a heavy user of Hive. However, for some of their workloads they required low latency response times in an interactive fashion. This is behind the rationale of Presto.
One of the advantages of Presto is that you can also query non HDFS data sources such as an RDBMS. It seems to be relatively easy to write your own connectors.
Spark SQL and Shark
Spark is the new darling on the Big Data scene and widely seen as the replacement for Map Reduce. Originally developed by AMPLab at UC Berkeley it is now developed by Databricks and also runs on Hadoop YARN. There are various components that ship with Spark. A micro batch near realtime processing module (Spark Streaming), a machine learning component (MLLIB), a graph database (alpha release), SparkR (alpha release), and what we are interested for the purpose of this article Spark SQL.Spark SQL to some extent borrows from Shark its predecssor, which was based on the Hive codebase but similar to Tez came with its own execution engine. The big advantage of Spark SQL over the other engines is that it is easy to mix machine learning with SQL. BTW, an alternative to this is to use the HiveMall machine learning library (unfortunately, there is very little documentation). This is similar to in database analytics as offered by vendors such as Oracle and has the advantage that you don’t have to move around the data between different tools and technologies. While you can write resultsets back into Hive, in my opinion Spark SQL is currently not really an option for doing SQL based ETL/batch as you would have to intermix it with Scala code to perform more complex transformations, which makes things somewhat ugly. So the primary use case is to use it as an interactive query tool and mix it with machine learning. It also does not offer a rich SQL feature set right now, e.g. analytic functions are missing. Like the other engines, Spark SQL also has its own optimizer named Catalyst. Based on performance benchmarks by Databricks, Spark SQL seems to be able to trump its predecessor Shark in terms of performance (last slide in deck).
Similar to Impala, Apache Drill is another MPP SQL query engine inspired by the Google Dremel paper. Apache Drill is mainly supported by MapR. At the moment it is in alpha release. Together with Spark SQL It is at the moment of this writing the least mature SQL solution on Hadoop. As outlined by MapR Apache Drill will be available Q2 2014.
Similar to Presto, Apache Drill will also support non Hadoop data sources.
InfiniDB is rather different to any of the other SQL engines on Hadoop. I want to include it here as it is an interesting product that we will hear about more in the future. InfiniDB is an open source MPP columnar RDBMS. As such it falls more into the category of the likes of Amazon Redshift, Teradata, or Vertica. Unlike its competitors, it allows for its data to sit on the Hadoop File System (HDFS). The only and pretty fundamental caveat, however, is that you would have to load the data into the InfinDB proprietary data format. It currently does not support popular data serialization formats such as Parquet or Sequence Files. They may add this feature in a future release. However, this will negatively impact performance. I will keep an eye on this product as it as an excellent and open source alternative to MPPs such as Teradata, Vertica, Netezza etc. However, the data duplication issue is a problem if you have subscribed to the paradigm of bringing the processing to the data rather than the other way around. On the other hand InfiniDB (as you would expect from an MPP RDBMS) supports Updates and Deletes.
There are various other SQL engines on top of Hadoop including Cascading Lingual built as a SQL abstraction layer on top of Cascading, Hadapt, and various other commercial products. Another open source solution is Apache Tajo.
Probably the most mature of the SQL Hadoop engines is BigSQL from IBM. I recently had the opportunity to attend a session by one of their chief architects and it looks quite impressive, e.g. it builds on the DB2 optimizer and as a result is built on decades of experience. As this post mainly deals with open source engines I won’t go into any more detail on BigSQL. However, if you want to find out more have a look at this video.
This benchmark by Cloudera compares Impala to Shark (disk and memory), Hive Tez (0.13), and Presto. Unsurprisingly, Cloudera Impala scores best here :-).
This benchmark by Cloudera compares Impala to Hive (0.12 and not 0.13) and to an unnamed MPP RDBMS. Surprise surprise, Cloudera Impala scores best here :-).
This benchmark by InfiniDB compares InfinDB to Presto, Impala, Hive on M/R. For all workloads InfinDB is the performance winner.
This benchmark by AMPLab at UC Berkley compares Redshift to Hive on M/R, Hive on Tez, Impala, and Shark. Performance winner for most workloads is Amazon Redshift
This benchmark by Hortonworks compares performance between Hive M/R (0.10) and Hive Tez (0.13). Interestingly there is no comparison to other Hadoop SQL engines. You can draw or own conclusions why this is.
Conclusion and Recommendation
As of this writing the most mature product with the richest feature set is Apache Hive (running on Tez). Crucially it offers analytic functions, support for the widest set of file formats, and ACID support (full support in release 0.14)
As of the current release, Impala lacks important SQL features. However, this is about to change in Impala 2.0.
Once it has matured Hive on Spark should be a very good alternative to Hive on Tez.
While Hortonworks claims that Hive can be used for interactive queries, Cloudera questions this. The various benchmarks are not conclusive. As always you should test yourself if Hive is suitable for realtime queries for your workload and use case.
All of the different solutions follow a similar approach in that they all first create a logical query plan in a Directed Acyclic Graph (DAG). This is then translated into a physical execution plan and the various components and operators of the explain plan are then executed in a distributed fashion.
There are various benchmarks out there, which suggest that Impala is the fastest for various workloads. However, I wouldn’t trust any of these too much and would suggest for you to perform your own benchmarks for your specific workload.
Spark SQL looks very promising for use cases where you want to use SQL to run machine learning algorithms (similar to in database analytics, e.g. in Oracle). As an alternative you could look at using HiveMall. It also looks promising for interactive SQL.
Performance benchmarks suggest that none of the Hadoop SQL execution frameworks currently match the performance of an MPP RDBMS such as InfiniDB, Amazon Redshift, or Teradata. One Cloudera benchmark suggests otherwise. However, this benchmark is criticised for not implementing the full set of the TPC-DS benchmark and various other items and as a result is somewhat questionable
. This does not come as a surprise really as decades of experience have gone into these relational engines.
So what to do
? Right now I would run both batch style queries (ETL) and interactive queries on Hive Tez as Hive offers the richest SQL feature set, especially analytic functions and supports a wide set of file formats. If you don’t get satisfactory query performance for your realtime queries you may want to look at some of the other engines. Impala is a mature solution. However, it lacks support for analytic functions, which are crucially important for data analysis tasks. Analytic functions will be added to the next release of Impala though. Another option is Presto, which offers this feature set. At this stage Spark SQL is only in alpha release and does not yet look very mature especially in terms of the SQL features. However, it is quite promising for in database style machine learning and predictive analytics (bring the processing to the data rather than data to processing). Apache Drill is also only in alpha release and may not be mature enough for your use case. If I had to bet my house on which of the solutions will prevail I would put it on a combination of Hive on Spark (for batch ETL) and Spark SQL (interactive queries and in database style machine learning and predictive analytics) to cover all use cases and workloads. If Spark SQL matures further in terms of the SQL feature set (analytic functions etc.) and allows for ETL based on the SQL paradigm I would exclusively put my money on it.