Partition in Spark

[NanthakumarYoga]

Hi Community, Need your help on understanding below topics.. 

I have a huge transaction file ( 20GB ) partition by transaction_date column , parquet file. I have evenly distributed data ( no skew ). There are 10 days of data and we have 10 partition folder each contains 1 GB. 

Path = '\FileStore\Nantha\Trx\data\2024-01-01\'    ....  '\FileStore\Nantha\Trx\data\2024-01-10'

Now, I would like to understand here,

1. While reading the file without where condition, ( a simple read ). How spark partition the data and process in parrallel. ( default partition size is 128MB ). Here i am confused on partition size is 128MB. what is this ??

2. when we use shuffle.partition is 200, which means 200 partitions having each 128MB ? how this partition is refered and calculated. Are these internal one.

3. When we are issues cache or persist on dataframe, will this store the whole dataframe in MEMORY / DISK... or will it store as partition internally. 

[Kaniz]

Hi @NanthakumarYoga, Let’s delve into each of your questions about Spark and data partitioning:

1. Data Partitioning and Parallel Processing:

   • When you read a large Parquet file without any specific where condition (a simple read), Spark automatically partitions the data for parallel processing.
   • By default, Spark aims for a partition size of 128MB. This means that the entire dataset is divided into smaller chunks (partitions), each approximately 128MB in size.
   • These partitions are distributed across the available processing nodes (executors) in your Spark cluster.
   • The goal is to ensure that each executor processes a manageable portion of the data concurrently, maximizing parallelism and overall performance.
   • If your data is evenly distributed across the 10 days and 10 partitions, Spark will create 10 partitions, each containing data for one day.

2. Shuffle Partitions and Their Size:

   • When you perform operations that involve data shuffling (such as joins or aggregations), Spark uses shuffle partitions.
   • The shuffle.partition configuration determines the number of shuffle partitions. By default, it’s set to 200.
   • These shuffle partitions are not directly related to the 128MB partition size used during simple reads.
   • Instead, shuffle partitions control how data is redistributed during shuffling operations.
   • For example, if you perform a join, Spark will shuffle data between partitions to align matching keys.
   • The shuffle partitions help manage the data movement during these operations, ensuring efficient parallelism.
   • The actual size of each shuffle partition depends on the data distribution and the total amount of data being shuffled.

3. Cache and Persist:

   • When you cache or persist a DataFrame, Spark stores the intermediate result in memory or on disk.
   • The storage level determines where the data is stored:
     • MEMORY_ONLY: Stores the data in memory as deserialized Java objects.
     • MEMORY_AND_DISK: Stores the data in memory, and if it doesn’t fit, spills to disk.
     • Other storage levels include serialization options (MEMORY_ONLY_SER, MEMORY_AND_DISK_SER) and disk-only storage (DISK_ONLY).
   • When you cache or persist a DataFrame, it doesn’t store the entire DataFrame as a whole. Instead:
     • Each partition of the DataFrame is cached independently.
     • If a partition doesn’t fit in memory, it spills to disk (based on the storage level).
     • Cached partitions are reused across subsequent actions on the same DataFrame.
   • Caching helps avoid recomputing expensive transformations and improves performance for iterative or interactive workloads.

In summary:

• Spark partitions data for parallel processing during reads.
• Shuffle partitions are used during data shuffling operations.
• Caching/persisting stores individual partitions in memory or on disk, not the entire DataFrame.

Feel free to explore these concepts further, and let me know if you have any more questions! 😊