随笔 - 60  文章 - 0  评论 - 540  0


What’s new for Spark SQL in Apache Spark 1.3




The Apache Spark 1.3 release represents a major milestone for Spark SQL.  In addition to several major features, we are very excited to announce that the project has officially graduated from Alpha, after being introduced only a little under a year ago.  In this blog post we will discuss exactly what this step means for compatibility moving forward, as well as highlight some of the major features of the release.

  • exactly [ɪɡˈzæktli] 恰恰;确切地;精确地;完全地,全然
  • compatibility [kəmˌpætəˈbɪlətɪ] 适合;互换性; 通用性;和睦相处 兼容性
  • highlight [ˈhaɪˌlaɪt] 强调,突出;把…照亮;标示记号;为…中最突出的事物

1、Graduation from Alpha

While we know many organizations (including all of Databricks’ customers) have already begun using Spark SQL in production, the graduation from Alpha comes with a promise of stability for those building applications using this component.  Like the rest of the Spark stack, we now promise binary compatibility for all public interfaces through the Apache Spark 1.X release series.

  • rest [rest]  休息;剩余部分;支持物;宁静,安宁 休息;静止;停止;安心
  • the rest of  其余的;剩下的
  • binary compatibility [计] 二进制兼容性

Since the SQL language itself and our interaction with Apache Hive represent a very large interface, we also wanted to take this chance to articulate our vision for how the project will continue to evolve. A large number of Spark SQL users have data in Hive metastores and legacy workloads which rely on Hive QL. As a result, Hive compatibility will remain a major focus for Spark SQL moving forward.

  • interaction [ˌɪntɚˈækʃən] 互动;一起活动;合作;互相影响
  • articulate [ɑ:rˈtɪkjuleɪt]  清晰地发(音);言语表达;(用关节)连接 表达
  • vision [ˈvɪʒən] 视力,视觉;美景,绝妙的东西;幻影;想像力 视野
  • evolve  [iˈvɑ:lv] 使发展;使进化;设计,制订出;发出,散发 演变
  • legacy [ˈlɛɡəsi] 遗产;遗赠

More specifically, the HiveQL interface provided by the HiveContext remains the most complete dialect of SQL that we support and we are committed to continuing to maintain compatibility with this interface.  In places where our semantics differ in minor ways from Hive’s (i.e. SPARK-5680), we continue to aim to provide a superset of Hive’s functionality.  Additionally, while we are excited about all of the new data sources that are available through the improved native Data Sources API (see more below), we will continue to support reading tables from the Hive Metastore using Hive’s SerDes.

  • dialect [ˈdaɪəˌlɛkt] 方言,土语;语调;[语]语支;专业用语
  • in minor ways 在小的方面
  • minor [ˈmaɪnɚ] 较小的,少数的,小…;未成年的;[乐]小调的,小音阶的;

The new DataFrames API (also discussed below) is currently marked experimental.  Since this is the first release of this new interface, we wanted an opportunity to get feedback from users on the API before it is set in stone.  That said, we do not anticipate making any major breaking changes to DataFrames, and hope to remove the experimental tag from this part of Spark SQL in Apache Spark 1.4.  You can track progress and report any issues at SPARK-6116.

  • experimental [ɪkˌspɛrəˈmɛntl]  实验的;根据实验的;试验性的
  • it is set in stone 坚如磐石、坚定不移、绝不动摇、不可改变
  • anticipate  [ænˈtɪsəˌpet] 预见;预料;预感;先于…行动

2、Improved Data Sources API

The Data Sources API was another major focus for this release, and provides a single interface for loading and storing data using Spark SQL.  In addition to the sources that come prepackaged with the Apache Spark distribution, this API provides an integration point for external developers to add support for custom data sources.  At Databricks, we have already contributed libraries for reading data stored in Apache Avro or CSV and we look forward to contributions from others in the community (check out spark packages for a full list of sources that are currently available).

  • prepackaged  [pri:'pækɪdʒd] 出售以前先包装( prepackage的过去式和过去分词 ) 
  • integration point 集成点;积分点;整合点

3、Unified Load/Save Interface

In this release we added a unified interface to SQLContext and DataFrame for loading and storing data using both the built-in and external data sources.  These functions provide a simple way to load and store data, independent of whether you are writing in Python, Scala, Java, R or SQL.  The examples below show how easy it is to both load data from Avro and convert it into parquet in different languages.

  • independent of  [ˌɪndɪˈpɛndənt ʌv] 不依赖…,独立于… 

3.1 Scala

val df = sqlContext.load("/home/michael/data.avro", "com.databricks.spark.avro")
df.save("/home/michael/data.parquet", "parquet")

3.2 Python

df = sqlContext.load("/home/michael/data.avro", "com.databricks.spark.avro")
df.save("/home/michael/data.parquet", "parquet")

3.3 Java

DataFrame df = sqlContext.load("/home/michael/data.avro", "com.databricks.spark.avro")
df.save("/home/michael/data.parquet", "parquet")

3.4 SQL

  USING com.databricks.spark.avro
    path "/home/michael/data.avro"

> CREATE TABLE parquetData
  USING parquet
    path "/home/michael/data/parquet")
  AS SELECT * FROM avroData

4、Automatic Partition Discovery and Schema Migration for Parquet

Parquet has long been one of the fastest data sources supported by Spark SQL.  With its columnar format, queries against parquet tables can execute quickly by avoiding the cost of reading unneeded data.

In the Apache Spark 1.3 release we added two major features to this source.  First, organizations that store lots of data in parquet often find themselves evolving the schema over time by adding or removing columns.  With this release we add a new feature that will scan the metadata for all files, merging the schemas to come up with a unified representation of the data.  This functionality allows developers to read data where the schema has changed overtime, without the need to perform expensive manual conversions.

  • organizations [ˌɔ:ɡənaɪ'zeɪʃnz]  组织( organization的名词复数 );组织性;组织工作;安排
  • come up with 拿出来
  • representation 表示
  • perform [pərˈfɔ:rm] 执行;履行;表演;扮演
  • expensive [ɪkˈspɛnsɪv] 昂贵的,花钱多的;豪华的
  • manual [ˈmænjuəl] 用手的;手制的,手工的;[法]占有的;体力的
  • conversions [kən'vɜ:ʃnz] 变换( conversion的名词复数 );(宗教、信仰等)彻底改变;

Additionally, the parquet datasource now supports auto-discovering data that has been partitioned into folders, and then prunes which folders are scanned based on predicates in queries made against this data.  This optimization means that you can greatly speed up may queries simply by breaking up your data into folders.  For example:

  • prune  [prun] 删除;减少
  • predicate ['predɪkət] 断言,断定;宣布,宣讲;使基于
  • breaking up 破裂 分手 分解

SELECT * FROM table WHERE year = 2015

In Apache Spark 1.4, we plan to provide an interface that will allow other formats, such as ORC, JSON and CSV, to take advantage of this partitioning functionality.

5、Persistent Data Source Tables

Another feature that has been added in Apache Spark 1.3 is the ability to persist metadata about Spark SQL Data Source tables to the Hive metastore.  These tables allow multiple users to share the metadata about where data is located in a convenient manner.  Data Source tables can live alongside native Hive tables, which can also be read by Spark SQL.

  • alongside [əˌlɔ:ŋˈsaɪd]  在…的侧面;在…旁边;与…并排

6、Reading from JDBC Sources

Finally, a Data Source for reading from JDBC has been added as built-in source for Spark SQL.  Using this library, Spark SQL can extract data from any existing relational databases that supports JDBC.  Examples include mysql, postgres, H2, and more.  Reading data from one of these systems is as simple as creating a virtual table that points to the external table.  Data from this table can then be easily read in and joined with any of the other sources that Spark SQL supports.

  USING org.apache.spark.sql.jdbc
    url "jdbc:postgresql:dbserver",
    dbtable "impressions"

> SELECT COUNT(*) FROM impressions

This functionality is a great improvement over Spark’s earlier support for JDBC (i.e., JdbcRDD).  Unlike the pure RDD implementation, this new DataSource supports automatically pushing down predicates, converts the data into a DataFrame that can be easily joined, and is accessible from Python, Java, and SQL in addition to Scala.

  • pure [pjʊr] 纯的;单纯的;纯真的;干净的 纯粹的

7、Introducing DataFrames

While we have already talked about the DataFrames in other blog posts and talks at the Spark Summit East, any post about Apache Spark 1.3 would be remiss if it didn’t mention this important new API. DataFrames evolve Spark’s RDD model, making it faster and easier for Spark developers to work with structured data by providing simplified methods for filtering, aggregating, and projecting over large datasets. Our DataFrame implementation was inspired by Pandas’ and R’s data frames, and are fully interoperable with these implementations.  Additionally, Spark SQL DataFrames are available in Spark’s Java, Scala, and Python API’s as well as the upcoming (unreleased) R API.

  • remiss [rɪˈmɪs]  玩忽职守的,马虎 失职
  • inspired  [ɪnˈspaɪərd]  有创造力的;品质优秀的;有雄心壮志的; 受到启发的
  • available [əˈveləbəl] 可获得的;有空的;可购得的;能找到的
  • upcoming [ˈʌpˌkʌmɪŋ] 即将到来;即将来到的,即将出现的


Internally, DataFrames take advantage of the Catalyst query optimizer to intelligently plan the execution of your big data analyses. This planning permeates all the way into physical storage, where optimizations such as predicate pushdown are applied based on analysis of user programs.  Since this planning is happening at the logical level, optimizations can even occur across function calls, as shown in the example below.

  • advantage  [ədˈvæntɪdʒ]  有利条件;益处;优越(性);处于支配地位 优势
  • permeate  [ˈpɜ:rmieɪt]  渗透;渗入;弥漫;遍布
  • pushdown 下推

  • predicate pushdown 谓词下推,属于逻辑优化。优化器可以将谓词过滤下推到数据源,从而使物理执行跳过无关数据。在使用Parquet的情况下,更可能存在文件被整块跳过的情况,同时系统还通过字典编码把字符串对比转换为开销更小的整数对比。在关系型数据库中,谓词则被下推到外部数据库用以减少数据传输。
  • column pruning 列裁剪,在读数据的时候,只读取查询中需要用到的列,而忽略其他列。

In this example, Spark SQL is able to push the filtering of users by their location through the join, greatly reducing its cost to execute.  This optimization is possible even though the original author of the add_demographics function did not provide a parameter for specifying how to filter users!

  • specifying ['spesɪfaɪʃ]  指定( specify的现在分词 );详述;提出…的条件;使具有特性

This is only example of how Spark SQL DataFrames can make developers more efficient by providing a simple interface coupled with powerful optimization.

To learn more about Spark SQL, Dataframes, or Apache Spark 1.3, checkout the SQL programming guide on the Apache Spark website. Stay tuned to this blog for updates on other components of the Apache Spark 1.3 release!


  • https://databricks.com/blog/2015/03/24/spark-sql-graduates-from-alpha-in-spark-1-3.html
posted on 2018-04-08 21:12  虾皮  阅读(...)  评论(...编辑  收藏