To close the topic of the new arbitrary stateful processing API in Apache Spark Structured Streaming let's focus on its...batch counterpart!
Last week we discovered the new way to write arbitrary stateful transformations in Apache Spark 4 with the transformWithState API. Today it's time to delve into the implementation details and try to understand the internal logic a bit better.
Arbitrary stateful processing has been evolving a lot in Apache Spark. The initial version with updateStateByKey evolved to mapWithState in Apache Spark 2. When Structured Streaming was released, the framework got mapGroupsWithState and flatMapGroupsWithState. Now, Apache Spark 4 introduces a completely new way to interact with the arbitrary stateful processing logic, the Arbitrary state API v2!
Using cloud managed services is often a love and hate story. On one hand, they abstract a lot of tedious administrative work to let you focus on the essentials. From another, they often have quotas and limits that you, as a data engineer, have to take into account in your daily work. These limits become even more serious when they operate in a latency-sensitive context, as the one of stream processing.
A few months ago I wrote a blog post about event skew and how dangerous it is for a stateful streaming job. Since it was a high-level explanation, I didn't cover Apache Spark Structured Streaming deeply at that moment. Now the watermark topic is back to my learning backlog and it's a good opportunity to return to the event skew topic and see the dangers it brings for Structured Streaming stateful jobs.
Code organization and assertions flow are both important but even them, they can't guarantee your colleagues' adherence to the unit tests. There are other user-facing attributes to consider as well.
Testing batch jobs is not the same as testing streaming ones. Although the transformation (the WHAT from the previous article) is similar in both cases, more complete validation tests on the job logic are not. After all, streaming jobs often iteratively build the final outcome while the batch ones generate it in a single pass.
With this blog I'm starting a follow-up series for my Data+AI Summit 2024 talk. I missed this family of blog posts a lot as the previous DAIS with me as speaker was 4 years ago! As previously, this time too I'll be writing several blog posts that should help you remember the talk and also cover some of the topics left aside because of the time constraints.
I wrote a blog post about OutputModes 6 (yes!) years ago and after reading it a few times, I realized it was not good enough to be a quick refresher. For that reason you can read about OutputModes for the second time here. Hopefully, this one will be a good try!
Streaming jobs are supposed to run continuously but it applies to the data processing logic. After all, sometimes you may need to release a new job package with upgraded dependencies or improved business logic. What happens then?
Apache Spark leverages the observer design pattern for the framework-to-code communication. One of the consumers' implementations is StreamingQueryListener.
That's the question. The lack of the processing time trigger means more a reactive micro-batch triggering but it cannot be considered as the single true best practice. Let's see why.
Apache Spark Structured Streaming relies on the micro-batch pattern which evaluates the same query in each execution. That's only a high level vision, though. Under-the-hood, there are many other interesting things that happen.
I bet you know it already. You can limit the max throughput for Apache Spark Structured Streaming jobs for popular data sources such as Apache Kafka, Delta Lake, or raw files. Have you known that you can also control the lower limit, at least for Apache Kafka?
Data enrichment is one of common data engineering tasks. It's relatively easy to implement with static datasets because of the data availability. However, this apparently easy task can become a nightmare if used with inappropriate technologies.
In March I wrote a blog showing how to use accumulators to know the application of each filter statement. Turns out, the solution may not be perfect as mentioned by Aravind in one of the comments. I bet you already have an idea but if not, keep reading. Everything will be clear in the end!
Watermark, or rather multiple watermarks management, has been a thorn in the side of Apache Spark Structured Streaming. It has improved in the previous release (3.4.0) but still had some room for improvement. Well, it did have because the 3.5.0 release brought a serious fix for the multiple watermarks scenario.
It's time to start the series covering Apache Spark 3.5.0 features. As the first topic I'm going to cover Structured Streaming which has got a lot of RocksDB improvements and some major API changes.
I've already written about watermarks in a few places in the blog but despite that, I still find things to refresh. One of them is the watermark used to filter out the late data, which will be the topic of this blog post.
Do not get the title wrong! Having applyInPandasWithState in the PySpark API is huge! However, due to Python duck typing, some operations are more difficult and more risky to express in the code than in the strongly typed Scala API.
