官术网_书友最值得收藏!

Shared variables

Spark being an MPP environment generally does not provide a shared state as the code is executed in parallel on a remote cluster node. Separate copies of data and variables are generally used during the map() or reduce() phases, and providing an ability to have a read-write shared variable across multiple executing tasks would be grossly inefficient. Spark, however, provides two types of shared variables:

  • Broadcast variables - Read-only variables cached on each machine
  • Accumulators - Variables that can be added through associative and commutative property

Broadcast variables

Largescale data movement is often a major factor in negatively affecting performance in MPP environments and hence every care is taken to reduce data movement while working on a clustered environment. One of the ways to reduce data movement is to cache frequently accessed data objects on the machines, which is essentially what Spark's broadcast variables are about - keep read-only variables cached on each machine rather than shipping a copy of it with its tasks. This is often required when you need to have the same copy of a small data set (typically a dimension table) accessible to every node in the cluster. Spark will distribute the data to the worker nodes using a very efficient algorithm:

  • Broadcast variables are set by the calling program/driver program and will be retrieved by the workers across the cluster
  • Since the objective is to share the data across the cluster, they are read-only after they have been set
  • The value of a broadcast variable is retrieved and stored only on the first read

A very common example is processing weblogs, where the weblogs contain only the pageId, whereas the page titles are stored in a lookup table. During the analysis of the weblogs you might want to join the page Id from the weblog to the one in the lookup table to identify what particular page was being browsed, which page gets the most hits, which page loses the most customers, and so on. This can be done using the web page lookup table being broadcasted across the cluster. For an example of Broadcast variables, please visit Appendix, There's More with Spark.

Accumulators

Accumulators are variables that support associative and commutative properties, which are essential for parallel computations. They are often required to implement counters and are natively supported by Spark for numeric types. Accumulators are different from broadcast variables because:

  • They are not read-only
  • Executors across the cluster can add to the value of the accumulator variables
  • The driver program can access the value of the accumulator variables
  • For an example on Accumulators, please visit Appendix, There's More with Spark.
主站蜘蛛池模板: 宁都县| 广德县| 云南省| 尉氏县| 大邑县| 阿勒泰市| 上虞市| 包头市| 盈江县| 蓬莱市| 密山市| 巴青县| 内乡县| 五华县| 中山市| 濮阳县| 房产| 峡江县| 宜君县| 吉木萨尔县| 新乡县| 资源县| 青铜峡市| 四子王旗| 沂水县| 陇南市| 太康县| 北流市| 岳阳市| 韩城市| 中江县| 全州县| 龙游县| 芮城县| 吕梁市| 南投县| 江西省| 桦川县| 东城区| 格尔木市| 泾阳县|