Skip to main content

What is Shuffling in Spark

Shuffling in Spark is a mechanism that Re-Distributes the data across different executors or workers in the clusters. 

Why do we need to Re-Distribute the data? 

 A) Re-Distribution is needed when there is a need of increasing or decreasing the data partitions in the situations below:

  • When the partitions are not sufficient enough to process the data load in the cluster
  • When the partitions are too high in numbers that it creates task scheduling overhead and it becomes the bottleneck in the processing time.
Re-Distribution can also be achieved by executing the shuffling on existing distributed data collection like RDD, DataFrames, etc by using the "Repartition" and "Coalesce" APIs in Spark.

B) During Aggregation and Joins on data collection in Spark, all the data records belonging to aggregation or join should reside in the single partition and when the existing partitioning scheme doesn't satisfy this condition there is a need to re-distributing the data in input collections before performing the aggregations or joins.

Spark Shuffling is an expensive process as it is moving around data among different executors or workers in the cluster. Imagine, if you have 1000s of workers and have a huge volume of data, in this case, the Spark Shuffling can lead to expensive overheads.

Generally, Spark Shuffling involves
  • Disk I/O
  • Network I/O
  • Data Serialization & Deserialization
Some of the operations that might lead to the Shuffling are 

  • join
  • cogroup
  • groupWith
  • join
  • leftOuterJoin
  • rightOuterJoin
  • groupByKey
  • reduceByKey
  • repartition
  • coalesce
Data Engineers should always consider the possibility where they can avoid Shuffling in their process. For Example, they can choose to use reduceByKey instead of groupByKey due to the following reasons.

  • groupByKey shuffles all the data and therefore it is too slow.
  • reduceByKey shuffles only the sub-aggregation results in the partitions.

To summarize, Shuffling is one of the important operations of Spark that re-distributes the data to make the pipelines more efficient. But we are also aware of how expensive Shuffling could be especially when the number of workers in a cluster is too high or the volume of data is huge. So one should avoid the Shuffling in their pipelines. Also at the same time, one should also understand the concept clearly that could help them to design Fault-tolerant, Robust, and Reliable Data Pipelines.

Comments

Post a Comment

Popular posts from this blog

How to Backfill the Data in Airflow

In Apache Airflow, backfilling is the process of running a DAG or a subset of its tasks for a specific date range in the past. This can be useful if you need to fill in missing data, or if you want to re-run a DAG for a specific period of time to test or debug it. Here are the steps to backfill a DAG in Airflow: Navigate to the Airflow web UI and select the DAG that you want to backfill. In the DAG detail view, click on the "Graph View" tab. Click on the "Backfill" button in the top right corner of the page. In the "Backfill Job" form that appears, specify the date range that you want to backfill. You can use the "From" and "To" fields to set the start and end dates, or you can use the "Last X" field to backfill a certain number of days. Optional: If you want to backfill only a subset of the tasks in the DAG, you can use the "Task Instances" field to specify a comma-separated list of task IDs. Click on the "Star...
Post a Comment
Read more

What is BigQuery?

BigQuery is a fully-managed, cloud-native data warehouse from Google Cloud that allows organizations to store, query, and analyze large and complex datasets in real-time. It's a popular choice for companies that need to perform fast and accurate analysis of petabyte-scale datasets. One of the key advantages of BigQuery is its speed. It uses a columnar storage format and a Massively Parallel Processing (MPP) architecture, which allows it to process queries much faster than traditional row-based warehouses. It also has a highly optimized query engine that can handle complex queries and aggregations quickly. BigQuery is also fully integrated with other Google Cloud products, making it easy to build end-to-end data pipelines using tools like Google Cloud Storage, Google Cloud Data Fusion, and Google Cloud Dataproc. It can also be used to power dashboards and reports in tools like Google Data Studio. In addition to its speed and integration capabilities, BigQuery has a number of advance...
Post a Comment
Read more

Difference between ETL and ELT Pipelines

ETL (Extract, Transform, Load) and ELT (Extract, Load, Transform) are two common architectures for data pipelines. Both involve extracting data from one or more sources, loading the data into a destination system, and possibly transforming the data in some way. The main difference between the two approaches is the order in which the transform and load steps are performed. In an ETL pipeline, the transform step is typically performed before the data is loaded into the destination system. This means that the data is cleaned, transformed, and structured into a form that is optimized for the destination system before it is loaded. The advantage of this approach is that it can be more efficient, since the data is transformed once and then loaded into the destination system, rather than being transformed multiple times as it is queried. However, ETL pipelines can be inflexible, since the data must be transformed in a specific way before it is loaded, and it can be difficult to modify the pip...
Post a Comment
Read more