So far I've covered the writing part of the shuffle files. You've learned about 3 different shuffle writers, but what happens with their generated files? Who and how reads them? Is the reading an in-memory operation? I will try to answer this and some other questions in this blog post.
Some time ago I participated in an interesting meetup about the MERGE operation in Delta Lake (link in the Further reading section). Jacek Laskowski presented the operation internals and asked an interesting question about the difference between commands and execs. Since I didn't know the answer right away, I decided to explore the commands concepts in this blog post.
With the Adaptive Query Execution module, you can have a feeling that Apache Spark will optimize the job for you. In part, yes, because it'll be able to optimize the job based on the runtime parameters you don't necessarily know. However, you also can master the execution, and ones of these mastery tools are hints.
Previously we discovered what happens when you coalesce a dataset. To recall, it doesn't involve shuffle operation. It's then the opposite of a repartition operation which is a first class shuffle citizen.
Every Apache Spark release brings not only completely new components but also new native functions. The 3.1.1 is not an exception and it also comes with some new built-in functions!
Even though the change I will describe in this blog post is still in progress, it's worth attention, especially that I missed the DataSource V2 evolution in my previous blog posts.
Predicate pushdown is a data processing technique taking user-defined filters and executing them while reading the data. Apache Spark already supported it for Apache Parquet and RDBMS. Starting from Apache Spark 3.1.1, you can also use them for Apache Avro, JSON and CSV formats!
I have waited for writing this blog post since the Data+AI Summit 2020, where Cheng Su presented the ongoing effort to improve shuffle and stream-to-stream joins in Apache Spark 3.1. And in this blog post, I will start by sharing what changed for the joins in the new release of the framework!
That's probably one of the most common questions you may have heard in preliminary job interviews. What's the difference between coalesce and repartition? Many answers exist, but instead of repeating them, I will try to dig a bit deeper in this blog post and see how the coalesce works.
Pivot operation presented 2 weeks ago transforms some cells into columns. The reverse one is called stack and it's time to see how it works!
If you came to data engineering after having a BI career, you certainly know what the pivot is. It was not my case and was quite amazed by this operation that transforms values from rows into columns. If you want to understand how it's possible, this article will present some internals of pivoting data in Apache Spark.
Very often you will find Apache Spark performance tips related to the hardware (memory, GC) or the configuration parameters (shuffle partitions number, broadcast join threshold). But they're not the single ones you can implement. Moreover, IMO, you should start by the ones presented in this article and optimize your pipeline code before going into more complicated hardware and configuration tuning.
Apache Spark has this great capacity to optimize joins of bucketed tables but does it work on partitions as well? No, and to understand why, I invite you to read the following sections of this blog post ?
Have you ever wondered what is the relationship between drop and select operations in Apache Spark SQL? If not, I will shed some light on them in this short blog post.
There are stories like this, the stories that remain in the backlog for a very long time, and finally, they get implemented. That's exactly what happened with the Dynamic Partition Pruning feature added, after almost 4 years in the backlog, to Apache Spark 3.
The Local shuffle reader presented in one of the previous posts might have introduced some doubt in the way how the broadcast join is working. If it's the case, this blog post should shed some light on it. If not, it can give you more in-depth details than the ones introducing this type of join a few years ago.
If you noticed that some filter expressions weren't pushed down to your Apache Parquet files, the situation should change in Apache Spark 3.0. The new release supports this feature called nested data predicate pushdown.
All the operations from the title are natively available in relational databases but doing them with distributed data processing systems is not obvious. Starting from 3.0, Apache Spark gives a possibility to implement them in the data sources.
It's the last part of the series about the Adaptive Query Execution in Apache Spark SQL. So far you learned about the physical plan optimizations. But they're not alone and you will see that in this blog post.
So far you learned about skew optimization and coalesce shuffle partition optimizations made by the Adaptive Query Execution engine. But they're not the single ones and the next one you will discover is also related to the shuffle.
