How Fondant handles partitions#
When working with Fondant, each component deals with datasets. Fondant leverages Dask internally to handle datasets larger than the available memory. To achieve this, the data is divided into smaller chunks called "partitions" that can be processed in parallel. Ensuring a sufficient number of partitions enables parallel processing, where multiple workers process different partitions simultaneously, and smaller partitions ensure they fit into memory.
Check this link for more info on Dask partitions.
Fondant repartitions the loaded dataframe if the number of partitions is fewer than the available workers on the data processing instance. By repartitioning, the maximum number of workers can be efficiently utilized, leading to faster and parallel processing.
By default, Fondant automatically handles the partitioning, but you can disable this and create your own custom partitioning logic if you have specific requirements.
Here's an example of disabling the automatic partitioning:
The code snippet above disables automatic partitions for both the loaded and written dataframes, allowing you to define your own partitioning logic inside the components.
Moreover, you have the flexibility to set your own custom partitioning parameters to override the default settings:
In the example above, each partition of the loaded dataframe will contain approximately one hundred rows, and the size of the output partitions will be around 10MB. This capability is useful in scenarios where processing one row significantly increases the number of rows in the dataset (resulting in dataset explosion) or causes a substantial increase in row size (e.g., fetching images from URLs).
By setting a lower value for input partition rows, you can mitigate issues where the processed data grows larger than the available memory before being written to disk.