DEV Community 👩‍💻👨‍💻

Discussion on: The 5-minute guide to using bucketing in Pyspark

Collapse
luminousmen profile image
luminousmen Author

At their core, they are one and the same thing - optimizing data retrieval based on splitting up data by content. The difference is the number of aka separations that will eventually occur. As you said in order to join data, Spark needs the data that is to be joined (i.e., the data based on each key) to live on the same partition. In essence, we want to move the data (shuffle) as little as possible between the nodes of the cluster so that join can happen as parallel as possible (in each partition).

For example, if you're doing any annual financial reports, it makes sense to partition the data by month - you'll have 12 partitions in total. And if there are 12>= executors in the cluster, then doing join by month we can make join absolutely parallel if the dataset with which we join is also partied in the same way.

The same thing happens with buckets, but with data that don't have some logical way to partition (e.g., logs, user transactions, etc.). We can essentially do the same thing as with the partitions but specifying exact number of buckets (aka partitions).

I hope that makes sense. I'm thinking of writing an article on this while it's in my backlog. I hope I get to the point of writing it eventually :)