With the exponential growth in the amount of data being generated and advanced data-capturing capabilities, enterprises are facing the challenge of making sense out of this mountain of raw data. On the batch processing front, Hadoop has emerged as the go-to framework to deal with big data. Until recently, there has been a void when one looks for frameworks to build real-time stream processing applications. Such applications have become an integral part of a lot of businesses as they enable them to respond swiftly to events and adapt to changing situations. Examples of this are monitoring social media to analyze public response to any new product that you launch and predicting the outcome of an election based on the sentiments of election-related posts.
Organizations are collecting a large volume of data from external sources and want to evaluate/process the data in real time to get market trends, detect fraud, identify user behavior, and so on. The need for real-time processing is increasing day by day and we require a real-time system/platform that should support the following features:
- Scalable: The platform should be horizontally scalable without any down time.
- Fault tolerance: The platform should be able to process the data even after some of the nodes in a cluster go down.
- No data lost: The platform should provide the guaranteed processing of messages.
- High throughput: The system should be able to support millions of records per second and also support any size of messages.
- Easy to operate: The system should have easy installation and operation. Also, the expansion of clusters should be an easy process.
- Multiple languages: The platform should support multiple languages. The end user should be able to write code in different languages. For example, a user can write code in Python, Scala, Java, and so on. Also, we can execute different language code inside the one cluster.
- Cluster isolation: The system should support isolation so that dedicated processes can be assigned to dedicated machines for processing.
Apache Storm has emerged as the platform of choice for industry leaders to develop distributed, real-time, data processing platforms. It provides a set of primitives that can be used to develop applications that can process a very large amount of data in real time in a highly scalable manner.
Storm is to real-time processing what Hadoop is to batch processing. It is open source software, and managed by Apache Software Foundation. It has been deployed to meet real-time processing needs by companies such as Twitter, Yahoo!, and Flipboard. Storm was first developed by Nathan Marz at BackType, a company that provided social search applications. Later, BackType was acquired by Twitter, and it is a critical part of their infrastructure. Storm can be used for the following use cases:
- Stream processing: Storm is used to process a stream of data and update a variety of databases in real time. This processing occurs in real time and the processing speed needs to match the input data speed.
- Continuous computation: Storm can do continuous computation on data streams and stream the results to clients in real time. This might require processing each message as it comes in or creating small batches over a short time. An example of continuous computation is streaming trending topics on Twitter into browsers.
- Distributed RPC: Storm can parallelize an intense query so that you can compute it in real time.
- Real-time analytics: Storm can analyze and respond to data that comes from different data sources as they happen in real time.
In this chapter, we will cover the following topics:
- What is a Storm?
- Features of Storm
- Architecture and components of a Storm cluster
- Terminologies of Storm
- Programming language
- Operation modes
The following are some of the features of Storm that make it a perfect solution to process streams of data in real time:
- Fast: Storm has been reported to process up to 1 million tuples/records per second per node.
- Horizontally scalable: Being fast is a necessary feature to build a high volume/velocity data processing platform, but a single node will have an upper limit on the number of events that it can process per second. A node represents a single machine in your setup that executes Storm applications. Storm, being a distributed platform, allows you to add more nodes to your Storm cluster and increase the processing capacity of your application. Also, it is linearly scalable, which means that you can double the processing capacity by doubling the nodes.
- Fault tolerant: Units of work are executed by worker processes in a Storm cluster. When a worker dies, Storm will restart that worker, and if the node on which the worker is running dies, Storm will restart that worker on some other node in the cluster. This feature will be covered in more detail in Chapter 3, Storm Parallelism and Data Partitioning.
- Guaranteed data processing: Storm provides strong guarantees that each message entering a Storm process will be processed at least once. In the event of failures, Storm will replay the lost tuples/records. Also, it can be configured so that each message will be processed only once.
- Easy to operate: Storm is simple to deploy and manage. Once the cluster is deployed, it requires little maintenance.
- Programming language agnostic: Even though the Storm platform runs on Java virtual machine (JVM), the applications that run over it can be written in any programming language that can read and write to standard input and output streams.
A Storm cluster follows a master-slave model where the master and slave processes are coordinated through ZooKeeper. The following are the components of a Storm cluster.
The Nimbus node is the master in a Storm cluster. It is responsible for distributing the application code across various worker nodes, assigning tasks to different machines, monitoring tasks for any failures, and restarting them as and when required.
Nimbus is stateless and stores all of its data in ZooKeeper. There is a single Nimbus node in a Storm cluster. If the active node goes down, then the passive node will become an Active node. It is designed to be fail-fast, so when the active Nimbus dies, the passive node will become an active node, or the down node can be restarted without having any effect on the tasks already running on the worker nodes. This is unlike Hadoop, where if the JobTracker dies, all the running jobs are left in an inconsistent state and need to be executed again. The Storm workers can work smoothly even if all the Nimbus nodes go down but the user can't submit any new jobs into the cluster or the cluster will not be able to reassign the failed workers to another node.
Supervisor nodes are the worker nodes in a Storm cluster. Each supervisor node runs a supervisor daemon that is responsible for creating, starting, and stopping worker processes to execute the tasks assigned to that node. Like Nimbus, a supervisor daemon is also fail-fast and stores all of its states in ZooKeeper so that it can be restarted without any state loss. A single supervisor daemon normally handles multiple worker processes running on that machine.
In any distributed application, various processes need to coordinate with each other and share some configuration information. ZooKeeper is an application that provides all these services in a reliable manner. As a distributed application, Storm also uses a ZooKeeper cluster to coordinate various processes. All of the states associated with the cluster and the various tasks submitted to Storm are stored in ZooKeeper. Nimbus and supervisor nodes do not communicate directly with each other, but through ZooKeeper. As all data is stored in ZooKeeper, both Nimbus and the supervisor daemons can be killed abruptly without adversely affecting the cluster.
The following is an architecture diagram of a Storm cluster:
The basic unit of data that can be processed by a Storm application is called a tuple. Each tuple consists of a predefined list of fields. The value of each field can be a byte, char, integer, long, float, double, Boolean, or byte array. Storm also provides an API to define your own datatypes, which can be serialized as fields in a tuple.
A tuple is dynamically typed, that is, you just need to define the names of the fields in a tuple and not their datatype. The choice of dynamic typing helps to simplify the API and makes it easy to use. Also, since a processing unit in Storm can process multiple types of tuples, it's not practical to declare field types.
Each of the fields in a tuple can be accessed by its name,
getValueByField(String), or its positional index,
getValue(int), in the tuple. Tuples also provide convenient methods such as
getIntegerByField(String) that save you from typecasting the objects. For example, if you have a Fraction (numerator, denominator) tuple, representing fractional numbers, then you can get the value of the numerator by either using
You can see the full set of operations supported by
org.apache.storm.tuple.Tuple in the Java doc that is located at https://storm.apache.org/releases/1.0.2/javadocs/org/apache/storm/tuple/Tuple.html.
In Storm terminology, a topology is an abstraction that defines the graph of the computation. You create a Storm topology and deploy it on a Storm cluster to process data. A topology can be represented by a direct acyclic graph, where each node does some kind of processing and forwards it to the next node(s) in the flow. The following diagram is a sample Storm topology:
The following are the components of a Storm topology:
- Tuple: A single message/record that flows between the different instances of a topology is called a tuple.
- Stream: The key abstraction in Storm is that of a stream. A stream is an unbounded sequence of tuples that can be processed in parallel by Storm. Each stream can be processed by a single or multiple types of bolts (the processing units in Storm, which are defined later in this section). Thus, Storm can also be viewed as a platform to transform streams. In the preceding diagram, streams are represented by arrows. Each stream in a Storm application is given an ID and the bolts can produce and consume tuples from these streams on the basis of their ID. Each stream also has an associated schema for the tuples that will flow through it.
- Spout: A spout is the source of tuples in a Storm topology. It is responsible for reading or listening to data from an external source, for example, by reading from a log file or listening for new messages in a queue and publishing them--emitting in Storm terminology into streams. A spout can emit multiple streams, each of a different schema. For example, it can read records of 10 fields from a log file and emit them as different streams of seven-fields tuples and four-fields tuples each.
org.apache.storm.spout.ISpout interface is the interface used to define spouts. If you are writing your topology in Java, then you should use
org.apache.storm.topology.IRichSpout as it declares methods to use with the
TopologyBuilder API. Whenever a spout emits a tuple, Storm tracks all the tuples generated while processing this tuple, and when the execution of all the tuples in the graph of this source tuple is complete, it will send an acknowledgement back to the spout. This tracking happens only if a message ID was provided when emitting the tuple. If null was used as the message ID, this tracking will not happen.
A tuple processing timeout can also be defined for a topology, and if a tuple is not processed within the specified timeout, a fail message will be sent back to the spout. Again, this will happen only if you define a message ID. A small performance gain can be extracted out of Storm at the risk of some data loss by disabling the message acknowledgements, which can be done by skipping the message ID while emitting tuples.
The important methods of spout are:
nextTuple(): This method is called by Storm to get the next tuple from the input source. Inside this method, you will have the logic of reading data from external sources and emitting them to an instance of
org.apache.storm.spout.ISpoutOutputCollector. The schema for streams can be declared by using the
If a spout wants to emit data to more than one stream, it can declare multiple streams using the
declareStream method and specify a stream ID while emitting the tuple. If there are no more tuples to emit at the moment, this method will not be blocked. Also, if this method does not emit a tuple, then Storm will wait for 1 millisecond before calling it again. This waiting time can be configured using the
ack(Object msgId): This method is invoked by Storm when the tuple with the given message ID is completely processed by the topology. At this point, the user should mark the message as processed and do the required cleaning up, such as removing the message from the message queue so that it does not get processed again.
fail(Object msgId): This method is invoked by Storm when it identifies that the tuple with the given message ID has not been processed successfully or has timed out of the configured interval. In such scenarios, the user should do the required processing so that the messages can be emitted again by the
nextTuplemethod. A common way to do this is to put the message back in the incoming message queue.
open(): This method is called only once--when the spout is initialized. If it is required to connect to an external source for the input data, define the logic to connect to the external source in the open method, and then keep fetching the data from this external source in the
nextTuplemethod to emit it further.
Another point to note while writing your spout is that none of the methods should be blocking, as Storm calls all the methods in the same thread. Every spout has an internal buffer to keep track of the status of the tuples emitted so far. The spout will keep the tuples in this buffer until they are either acknowledged or failed, calling the
fail method, respectively. Storm will call the
nextTuple method only when this buffer is not full.
- Bolt: A bolt is the processing powerhouse of a Storm topology and is responsible for transforming a stream. Ideally, each bolt in the topology should be doing a simple transformation of the tuples, and many such bolts can coordinate with each other to exhibit a complex transformation.
org.apache.storm.task.IBolt interface is preferably used to define bolts, and if a topology is written in Java, you should use the
org.apache.storm.topology.IRichBolt interface. A bolt can subscribe to multiple streams of other components--either spouts or other bolts--in the topology and similarly can emit output to multiple streams. Output streams can be declared using the
declareStream method of
The important methods of a bolt are:
execute(Tuple input): This method is executed for each tuple that comes through the subscribed input streams. In this method, you can do whatever processing is required for the tuple and then produce the output either in the form of emitting more tuples to the declared output streams, or other things such as persisting the results in a database.
You are not required to process the tuple as soon as this method is called, and the tuples can be held until required. For example, while joining two streams, when a tuple arrives you can hold it until its counterpart also comes, and then you can emit the joined tuple.
The metadata associated with the tuple can be retrieved by the various methods defined in the
Tuple interface. If a message ID is associated with a tuple, the execute method must publish an
fail event using
OutputCollector for the bolt, or else Storm will not know whether the tuple was processed successfully. The
org.apache.storm.topology.IBasicBolt interface is a convenient interface that sends an acknowledgement automatically after the completion of the execute method. If a fail event is to be sent, this method should throw
prepare(Map stormConf, TopologyContext context, OutputCollector collector): A bolt can be executed by multiple workers in a Storm topology. The instance of a bolt is created on the client machine and then serialized and submitted to Nimbus. When Nimbus creates the worker instances for the topology, it sends this serialized bolt to the workers. The work will desterilize the bolt and call the
preparemethod. In this method, you should make sure the bolt is properly configured to execute tuples. Any state that you want to maintain can be stored as instance variables for the bolt that can be serialized/deserialized later.
Operation modes indicate how the topology is deployed in Storm. Storm supports two types of operation modes to execute the Storm topology:
- Local mode: In local mode, Storm topologies run on the local machine in a single JVM. This mode simulates a Storm cluster in a single JVM and is used for the testing and debugging of a topology.
- Remote mode: In remote mode, we will use the Storm client to submit the topology to the master along with all the necessary code required to execute the topology. Nimbus will then take care of distributing your code.
In the next chapter, we are going to cover both local and remote mode in more detail, along with a sample example.
Storm was designed from the ground up to be usable with any programming language. At the core of Storm is a thrift definition for defining and submitting topologies. Since thrift can be used in any language, topologies can be defined and submitted in any language.
Similarly, spouts and bolts can be defined in any language. Non-JVM spouts and bolts communicate with Storm over a JSON-based protocol over
Storm-starter has an example topology, https://github.com/apache/storm/tree/master/examples/storm-starter/multilang/resources, which implements one of the bolts in Python.
In this chapter, we introduced you to the basics of Storm and the various components that make up a Storm cluster. We saw a definition of different deployment/operation modes in which a Storm cluster can operate.
In the next chapter, we will set up a single and three-node Storm cluster and see how we can deploy the topology on a Storm cluster. We will also see different types of stream groupings supported by Storm and the guaranteed message semantic provided by Storm.