Big Data Hadoop Certification Training
- 164k Enrolled Learners
- Live Class
Contributed by Prithviraj Bose
In this blog we will discuss the windowing concept of Apache Spark’s stateful transformations.
What is stateful transformation?
Spark streaming uses a micro batch architecture where the incoming data is grouped into micro batches called Discretized Streams (DStreams) which also serves as the basic programming abstraction. The DStreams internally have Resilient Distributed Datasets (RDD) and as a result of this standard RDD transformations and actions can be done.
In streaming if we have a use case to track data across batches then we need state-ful DStreams.
For example we may track a user’s interaction in a website during the user session or we may track a particular twitter hashtag across time and see which users across the globe is talking about it.
Types of state-ful transformation.
State-ful DStreams are of two types – window based tracking and full session tracking.
For stateful tracking all incoming data should be transformed to key-value pairs such that the key states can be tracked across batches. This is a precondition.
Further we should also enable checkpointing, a concept which we will discuss in the later blogs.
> Window based tracking
In window based tracking the incoming batches are grouped in time intervals, i.e. group batches every ‘x’ seconds. Further computations on these batches are done using slide intervals.
For example if the window interval = 3 secs and slide interval = 2 secs, then all incoming data will be grouped in batches every 3 seconds and the computations on these batches will happen every 2 seconds. Alternatively we can say, do computations every 2 seconds on the batches that arrived in the last 3 seconds.
In the above diagram we see that the incoming batches are grouped every 3 units of time (window interval) and the computations are done every 2 units of time (slide interval).
Note: Unlike Apache Flink, Apache Spark does not have a concept of tumbling window, all windows are sliding.
A popular API for window based transformations is
There are several overloaded versions of this API, let’s see the one that has the most number of parameters. After this explanation the rest of the overloaded versions of this API should be self explanatory.
Returns: The transformed DStream[(K, V)]
reduceFunc: The associative reduce function.
invReduceFunc: The inverse of the above reduce function. This is required for efficient computing of incoming and outgoing batches. With the help of this function the value of the batches which are outgoing is deducted from the accumulated value of the above reduce function. For example, if we are computing the sum of the incoming values for the respective keys then for the outgoing batches we will subtract the values for the respective keys (provided they are present in the current batch else ignore).
windowDuration: Units of time for grouping the batches, this should be a multiple of the batch interval.
slideDuration: Units of time for computation, this should be a multiple of the batch interval.partitioner: The partitioner to use for storing the resulting DStream. For more information on partitioning read this.
filterFunc: Function to filter out expired key-value pairs, i.e. for example if we don’t get an update for a key for sometime we may wish to remove it.
Here’s a program to count the words coming from a socket stream. We have used the an overloaded version of the above function with a window interval of 4 secs and a slide interval of 2 secs.
In my next blog I will write about full session tracking and checkpointing.
Got a question for us? Please mention it in the comments section and we will get back to you.