Some time ago I was wondering why an object created once in the driver is recreated every time with new stage on executors - even if this object is sent through a broadcast variable. After some code digging, the response related to Java serialization appeared.
Some of previous posts (Serialization issues - part 1) presented some of solutions for serialization problems. This post is its continuation.
Issues with not serializable objects are maybe the most painful when we start to work with Spark. But hopefully there are several solutions to them.
Spark has 2 deployment modes that can be controlled in fine-grained way thanks to master URL property.
Keeping in mind which parts of Spark code are executed on driver and which ones on workers is important and can help to avoid some of annoying errors, as the ones related to serialization.
As every library, Spark has methods than are used more often than the others. As often used methods we could certainly define map or filter. In the other side of less popular transformations we could place, among others, tree-like methods that will be presented in this post.
In general Spark's actions reflects logic implemented in a lot of equivalent methods in programming languages. As an example we can consider isEmpty() that in Spark checks the existence of only 1 element and similarly in Java's List. But it can often lead to troubles, especially when more than 1 action is invoked.
After writing a post about testing Spark applications, I decided to take a look at Spark project tests and see which patterns they use to verify framework features.
It's difficult to contest the importance of testing in programming. Tests help to avoid regressions (a lot of regressions) and also to better understand developed code. Spark (and other data processing frameworks by the way) is not an exception of this rule. But, obviously, testing applications working in distributed mode is more tricky than in the case of standalone programs.
Every distributed computation is divided in small parts called jobs, stages and tasks. It's useful to know them especially during monitoring because it helps to detect bottlenecks.
Often a misconfiguration is the reason of all kinds of issues - performance, security or functional. Spark isn't an exception for this rule and it's the reason why this article focuses on configuration properties available for driver and executors.
This small post is the complement for previous article describing big lines of shuffle. It focuses more in details on writing part.
Memory management in Spark went through some changes. In the first versions, the allocation had a fix size. Only the 1.6 release changed it to more dynamic behavior. This change will be the main topic of the post.
As already told in one of previous posts about Spark, shuffle is a process which moves data between nodes. It's orchestrated by a specific manager and it will be the topic of this post.
Cache is an appreciable tool when we have a greedy computation generating a lot of data. Spark also uses this feature to better handle the case of RDD which generation is heavy (for example necessities database connection or data retrieval from external web services).
Checkpointing is, alongside caching, a method allowing to make a RDD persist. But there are some subtle differences between cache and checkpoint.
Serialization frameworks are intrinsic part of Big Data systems. Spark is not an exception for this rule and it offers some different possibilities to manage serialization.
Spark was developed to work on big amount of data. If big means millions of items. For every item one or several costly operations are done, it'll lead quick to performance problems. It's one of the reasons why Spark proposes operations executed once per partition.
Spark has an interesting concept of shared variables among all distributed computations. This special kind of objects is called broadcast variables. But it's not the single possibility to share objects in Spark. The second one are accumulators.
Partitioning in distributed data is quite common concept. Spark is not an exception and it also has some operations related to partitions.
If you want to, click on the banner βοΈ