Tuesday, November 25, 2008

Hadoop Map/Reduce Implementation

In my previous post, I talk about the methodology of transforming a sequential algorithm into parallel. After that, we can implement the parallel algorithm, one of the popular framework we can use is the Apache Opensource Hadoop Map/Reduce framework.

Functional Programming

Multithreading is one of the popular way of doing parallel programming, but major complexity of multi-thread programming is to co-ordinate the access of each thread to the shared data. We need things like semaphores, locks, and also use them with great care, otherwise dead locks will result.

If we can eliminate the shared state completely, then the complexity of co-ordination will disappear. This is the fundamental concept of functional programming. Data is explicitly passed between functions as parameters or return values which can only be changed by the active function at that moment. Imagine functions are connected to each other via a directed acyclic graph. Since there is no hidden dependency (via shared state), functions in the DAG can run anywhere in parallel as long as one is not an ancestor of the other. In other words, analyze the parallelism is much easier when there is no hidden dependency from shared state.

Map/Reduce functions

Map/reduce is a special form of such a DAG which is applicable in a wide range of use cases. It is organized as a “map” function which transform a piece of data into some number of key/value pairs. Each of these elements will then be sorted by their key and reach to the same node, where a “reduce” function is use to merge the values (of the same key) into a single result.


map(input_record) {
…
emit(k1, v1)
…
emit(k2, v2)
…
}

reduce (key, values) {
aggregate = initialize()
while (values.has_next) {
    aggregate = merge(values.next)
}
collect(key, aggregate)
}

The Map/Reduce DAG is organized in this way.



A parallel algorithm is usually structure as multiple rounds of Map/Reduce




Distributed File Systems

The distributed file system is designed to handle large files (multi-GB) with sequential read/write operation. Each file is broken into chunks, and stored across multiple data nodes as local OS files.



There is a master “NameNode” to keep track of overall file directory structure and the placement of chunks. This NameNode is the central control point and may re-distributed replicas as needed.

To read a file, the client API will calculate the chunk index based on the offset of the file pointer and make a request to the NameNode. The NameNode will reply which DataNodes has a copy of that chunk. From this points, the client contacts the DataNode directly without going through the NameNode.

To write a file, client API will first contact the NameNode who will designate one of the replica as the primary (by granting it a lease). The response of the NameNode contains who is the primary and who are the secondary replicas. Then the client push its changes to all DataNodes in any order, but this change is stored in a buffer of each DataNode. After changes are buffered at all DataNodes, the client send a “commit” request to the primary, which determines an order to update and then push this order to all other secondaries. After all secondaries complete the commit, the primary will response to the client about the success.

All changes of chunk distribution and metadata changes will be written to an operation log file at the NameNode. This log file maintain an order list of operation which is important for the NameNode to recover its view after a crash. The NameNode also maintain its persistent state by regularly check-pointing to a file.

In case of the NameNode crash, all lease granting operation will fail and so any write operation is effectively fail also. Read operation should continuously to work as long as the clinet program has a handle to the DataNode. To recover from NameNode crash, a new NameNode can take over after restoring the state from the last checkpoint file and replay the operation log.

When a DataNode crashes, it will be detected by the NameNode after missing its hearbeat for a while. The NameNode removes the crashed DataNode from the cluster and spread its chunks to other surviving DataNodes. This way, the replication factor of each chunk will be maintained across the cluster.

Later when the DataNode recover and rejoin the cluster, it reports all its chunks to the NameNode at boot time. Each chunk has a version number which will advanced at each update. Therefore, the NameNode can easily figure out if any of the chunks of a DataNode becomes stale. Those stale chunks will be garbage collected at a later time.


Job Execution

Hadoop MapRed is based on a “pull” model where multiple “TaskTrackers” poll the “JobTracker” for tasks (either map task or reduce task).

The job execution starts when the client program uploading three files: “job.xml” (the job config including map, combine, reduce function and input/output data path, etc.), “job.split” (specifies how many splits and range based on dividing files into ~16 – 64 MB size), “job.jar” (the actual Mapper and Reducer implementation classes) to the HDFS location (specified by the “mapred.system.dir” property in the “hadoop-default.conf” file). Then the client program notifies the JobTracker about the Job submission. The JobTracker returns a Job id to the client program and starts allocating map tasks to the idle TaskTrackers when they poll for tasks.




Each TaskTracker has a defined number of "task slots" based on the capacity of the machine. There are heartbeat protocol allows the JobTracker to know how many free slots from each TaskTracker. The JobTracker will determine appropriate jobs for the TaskTrackers based on how busy thay are, their network proximity to the data sources (preferring same node, then same rack, then same network switch). The assigned TaskTrackers will fork a MapTask (separate JVM process) to execute the map phase processing. The MapTask extracts the input data from the splits by using the “RecordReader” and “InputFormat” and it invokes the user provided “map” function which emits a number of key/value pair in the memory buffer.

When the buffer is full, the output collector will spill the memory buffer into disk. For optimizing the network bandwidth, an optional “combine” function can be invoked to partially reduce values of each key. Afterwards, the “partition” function is invoked on each key to calculate its reducer node index. The memory buffer is eventually flushed into 2 files, the first index file contains an offset pointer of each partition. The second data file contains all records sorted by partition and then by key.

When the map task has finished executing all input records, it start the commit process, it first flush the in-memory buffer (even it is not full) to the index + data file pair. Then a merge sort for all index + data file pairs will be performed to create a single index + data file pair.

The index + data file pair will then be splitted into are R local directories, one for each partition. After all the MapTask completes (all splits are done), the TaskTracker will notify the JobTracker which keeps track of the overall progress of job. JobTracker also provide a web interface for viewing the job status.

When the JobTracker notices that some map tasks are completed, it will start allocating reduce tasks to subsequent polling TaskTrackers (there are R TaskTrackers will be allocated for reduce task). These allocated TaskTrackers remotely download the region files (according to the assigned reducer index) from the completed map phase nodes and concatenate (merge sort) them into a single file. Whenever more map tasks are completed afterwards, JobTracker will notify these allocated TaskTrackers to download more region files (merge with previous file). In this manner, downloading region files are interleaved with the map task progress. The reduce phase is not started at this moment yet.

Eventually all the map tasks are completed. The JobTracker then notifies all the allocated TaskTrackers to proceed to the reduce phase. Each allocated TaskTracker will fork a ReduceTask (separate JVM) to read the downloaded file (which is already sorted by key) and invoke the “reduce” function, which collects the key/aggregatedValue into the final output file (one per reducer node). Note that each reduce task (and map task as well) is single-threaded. And this thread will invoke the reduce(key, values) function in assending (or descending) order of the keys assigned to this reduce task. This provides an interesting property that all entries written by the reduce() function is sorted in increasing order. The output of each reducer is written to a temp output file in HDFS. When the reducer finishes processing all keys, the temp output file will be renamed atomically to its final output filename.

The Map/Reduce framework is resilient to crashes of any components. TaskTracker nodes periodically report their status to the JobTracker which keeps track of the overall job progress. If the JobTracker hasn’t heard from any TaskTracker nodes for a long time, it assumes the TaskTracker node has been crashed and will reassign its tasks appropriately to other TaskTracker nodes. Since the map phase result is stored in the local disk, which will not be available when the TaskTracker node crashes. In case a map-phase TaskTracker node crashes, the crashed MapTasks (regardless of whether it is complete or not) will be reassigned to a different TaskTracker node, which will rerun all the assigned splits. However, the reduce phase result is stored in HDFS, which is available even the TaskTracker node crashes. Therefore, in case a reduce-phase TaskTracker node crashes, only the incomplete ReduceTasks need to be reassigned to a different TaskTracker node, where the incompleted reduce tasks will be re-run.

The job submission process is asynchronous. Client program can poll for the job status at any time by supplying the job id.

Sunday, November 23, 2008

Amazon Web Services

"Cloud computing" is one of the hot topics these days. Especially with the economy downturn, enterprises are looking every opportunity to reduce their spending. Cloud computing provides a number of attractions ...
  • Since resource capacity can be provisioned just in a few clicks, enterprises no longer need to budget for their peak load. They just need budget for the average load (rather than the peak load) and rent for extra resources from the cloud during peak traffic period. This is huge savings for them.
  • By outsourcing the operation environment, enterprise pass the problem of load balancing, fail recovery to the cloud provider. They no longer need to maintain their inhouse IT expertise which is required for running their web sites. This is also huge savings.
There are primarily two architectural models of how cloud providers deliver their services. The IaaS, and PaaS model.

IaaS – Infrastructure as a Service

The IaaS model take a very bottom-up approach and focus in just providing a “virtual machine”. They give their users all the freedom to pick any technology that suits their organization skills sets, budget and particular deployment scenarios. Since existing technology stack can be used in the cloud without any change, existing applications can be migrated to the cloud almost transparently. This is very attractive to enterprise who wants to enjoy the cost benefits that cloud computing provides without requiring them to make a lot of changes to their existing applications. On the other hand, it is very easy to migrate back from the cloud to internal IT. There is less concern of being lock-in.

However, the freedom also comes with a cost. Since there is a lot of infrastructure decision (such as load balancing, fail recovery) still need to be figured out, the user’s organization need to retain pretty much same amount of IT expertise to make these decisions and operate it.


PaaS – Platform as a Service

PaaS model takes a different approach in that they are not just covering the lowest machine layer but also provides rich technology in the upper technology stacks based the vendor's experience in running large scale websites. Enterprises can keep their focus in developing business logic and not worry about how the infrastructure should be setup.

The downside of PaaS is that enterprises has to write their application in specific language and API which is predefined by the cloud provider. That means existing application have to be first rewritten before they can be run inside PaaS. Also, to a high degree, the written application will be locked in to the cloud provider’s particular technology stack.

Predicting adoption

We predict PaaS model to be more attractive to small companies and startups who are primarily green field development and want to quickly deliver their features without worrying too much about infrastructure setup. On the other hand, larger enterprise who have more legacy system, or companies who just wants to test out cloud computing with minimum effort, or those who concern about vendor’s lock-in, will find IaaS to be much more attractive.

Amazon Web Services

Amazon is the current leader of Cloud provider based on the IaaS model. At the heart of its technology stack, they have the virtual machine layer called EC2. Amazon also provides a set of surrounding services for data storage, metadata storage, message queues, payment processing, content cache … etc.

EC2 – Elastic Computing

Amazon has procured a large number of commoditized Intel boxes running virtualization software Xen. On top of Xen, Linux or Windows can be run as the guest OS . The guest operating system can have many variations with different set of software packages installed.

Each configuration is bundled as a custom machine image (called AMI). Amazon host a catalog of AMI for the users to choose from. Some AMI is free while other requires a usage charge. User can also customize their own setup by starting from a standard AMI, make their special configuration changes and then create a specific AMI that is customized for their specific needs. The AMIs are stored in Amazon’s storage subsystem S3.

Amazon also classifies their machines in terms of their processor power (no of cores, memory and disk size) and charged their usage at a different rate. These machines can be run in different network topology specified by the users. There is an “availability zone” concept which is basically a logical data center. “Availability zone” has no interdependency and is therefore very unlikely to fail at the same time. To achieve high availability, users should consider putting their EC2 instances in different availability zones.

“Security Group” is the virtual firewall of Amazon EC2 environment. EC2 instances can be grouped under “security group” which specifies which port is open to which incoming range of IP addresses. So EC2 instances that running applications at various level of security requirements can be put into appropriated security groups and managed using ACL (access control list). Somewhat very similar to what network administrator configure their firewalls.

User can start the virtual machine (called an EC2 instance) by specifying the AMI, the machine size, the security group, and its authentication key via command line or an HTTP/XML message. So it is very easy to startup the virtual machine and start running the user’s application. When the application completes, the user can also shutdown the EC2 instance via command line or HTTP/XML message. The user is only charged for the actual time when the EC2 instance is running.

One of the issue of extremely dynamic machine configuration (such as EC2) is that a lot of configuration setting is transient and does not survive across reboot. For example, the node name and IP address may have been changed, all the data stored in local files is lost. Latency and network bandwidth between machines may also have changed. Fortunately, Amazon provides a number of ways to mitigate these issues.
  • By paying some charge, user can reserve a stable IP address, called “elastic IP”, which can be attached to EC2 instance after they bootup. External facing machine is typically done this way.
  • To deal with data persistence, Amazon also provides a logical network disk, called “elastic block storage” to store the data. By paying some charges, EBS is reserved for the user and it survives across EC2 reboots. User can attach the EBS to EC2 instances after the reboot.

S3 – Simple Storage Service

Amazon S3 provides a HTTP/XML services to save and retrieve content. It provides a file system-like metaphor where “objects” are group under “buckets”. Based on a REST design, each object and bucket has its own URL.

With HTTP verbs (PUT, GET, DELETE, POST), user can create a bucket, list all the objects within the bucket, create object within a bucket, retrieve an object, remove an object, remove a bucket … etc.

Under S3, each object has a unique URI which serves as its key. There is no query mechanism in S3 and User has to lookup the object by its key. Each object is stored as an opaque byte array with maximum 5GB size. S3 also provides an interesting partial object retrieval mechanism by specifying the ranges of bytes in the URL.

However, partial put is not current support but it can be simulated by breaking the large object into multiple small objects and then do the assembly at the app level. Breaking down the object also help to speed up the upload and download by doing the data transfer in parallel.

Within Amazon S3, each S3 objects are replicated across 2 (or more) data center and also cache at the edge for fast retrieval.

Amazon S3 is based on an “eventual consistent” model which means it is possible that an application won’t see the change it just made. Therefore, some degree of tolerance of inconsistent view is required by the application. Application should avoid the situation of having two concurrent modifications to the same object. And application should wait for some time between updates, and also should expect all the data it reads is potentially stale for few seconds.

There is also no versioning concept in S3, but it is not hard to build one on top of S3.


EBS – Elastic Block Storage

Based on RAID disks, EBS provides a persistent block storage device for data persistence where user can attach it to a running EC2 instance within the same availability zone. EBS is typically used as a file system that is mounted to EC2 instance, or as raw devices for database.

Although EBS is a network devices to the EC2 instance, benchmark from Amazon shows that it has higher performance than local disk access. Unlike S3 which is based on eventual consistent model, EBS provides strict consistency where latest updates are immediately available.


SimpleDB – queriable data storage

Unlike S3 where data has to be looked up by key, SimpleDB provides a semi-structured data store with querying capability. Each object can be stored as a number of attributes where the user can search the object by the attribute name.

Similar to the concepts of “buckets “ and “objects” in S3, SimpleDB is organized as a set of “items” grouped by “domains”. However, each item can have a number of “attributes” (up to 256). Each attribute can store one or multiple values and the value must be a string (or a string array in case of multi-valued attribute). Each attribute can store up to 1K bytes, so it is not appropriate to store binary content.

SimpleDB is typically used as a metadata store in conjuction with S3 where the actual data is being stored. SimpleDB is also schema-less. Each item can define its own set of attributes and is free to add more or remove some attributes at runtime.

SimpleDB provides a query capability which is quite different from SQL. The “where” clause can only match an attribute value with a constant but not with other attributes. On the other hand, the query result only return the name of the matched items but not the attributes, which means subsequent lookup by item name is needed. Also, there is no equivalent of “order by” and the returned query result is unsorted.

Since all attribute are store as strings (even number, dates … etc). All comparison operation is done based on lexical order. Therefore, special encoding is needed for data type such as date, number to string to make sure comparison operation is done correctly.

SimpleDB is also based on an eventual consistency model like S3.


SQS – Simple Queue Service

Amazon provides a queue services for application to communicate in an asynchronous way with each other. Message (up to 256KB size) can be sent to queues. Each queue is replicated across multiple data centers.

Enterprises use HTTP protocol to send messages to a queue. “At least once” semantics is provided, which means, when the sender get back a 200 OK response, SQS guarantees that the message will be received by at least one receiver.

Receiving messages from a queue is done by polling rather than event driven calling interface. Since messages are replicated across queues asynchronously, it is possible that receivers only get some (but not all) messages sent to the queue. But the receiver keep polling the queue, he will eventually get all messages sent to the queue. On the other hand, message can be delivered out of order or delivered more than once. So the message processing logic needs to be idempotent as well as independent of message arrival order.

Once message is taken by a receiver, the message is invisible to other receivers for a period of time but it is not gone yet. The original receiver is supposed to process the message and make an explicit call to remove the message permanently from the queue. If such “removal” request is not made within the timeout period, the message will be visible in the queue again and will be picked up by subsequent receivers.

CloudWatch -- Monitoring Services

CloudWatch provides an API to extract system level metrics for each VM (e.g. CPU, network I/O and disk I/O) as well as for each load balancer services (e.g. response time, request rate). The collected metrics is modeled as a multi-dimensional data cube and therefore can be queried and aggregated (e.g. min/max/avg/sum/count) in different dimensions, such as by time, or by machine groups (by ami, by machine class, by particular machine instance id, by auto-scaling group).

This metrics is also used to drive the auto-scaling services (described below). Note that the metrics are predefined by Amazon and custom metrics (application level metrics) is not supported at this moment.

Load Balancing Services

Load balancer provides a way to group identical VMs into a pool. Amazon provides a way to create a software load balancer in a region and then attach EC2 instances (of the same region) to the it. The EC2 instances under a particular load balancer can be in different availability zone but they have to be in the same region.

Auto-Scaling Services

Auto-scaling allows the user to group a number of EC2 instances (typically behind the same load balancer) and specify a set of triggers to grow and shrink the group. Trigger defines the condition which is matching the collected metrics from the CloudWatch and match that against some threshold values. When match, the associated action can be to grow or shrink the group.

Auto-scaling allows resource capacity (number of EC2 instances) automatically adjusted to the actual workload. This way user can automatically spawn more VMs as the workload increases and shutdown the VM as the load decreases.

Elastic Map/Reduce

Amazon provides an easy way to run Hadoop Map/Reduce in the EC2 environment. They provide a web UI interface to start/stop a Hadoop Cluster and submit jobs to it. For a detail of how Hadoop works, see here.

Under elastic MR, both input and output data are stored into S3 rather than HDFS. This means data need to be loaded to S3 before the Hadoop processing can be started. Elastic also provides a job flow definition so user can concatenate multiple Map/Reduce job together. Elastic MR supports the program to be written in Java (jar) or any programming language (Hadoop streaming) as well as PIG and Hive.

Relational DB Services

RDS is basically running MySQL in the EC2.


Virtual Private Cloud

VPC is a VPN solution such that the user can extend its data center to include EC2 instances running in the Amazon cloud. Notice that this is an "elastic data center" because its size can grow and shrink when the user starts / stops EC2 instances.

User can create a VPC object which represents an isolated virtual network in the Amazon cloud environment and user can create multiple virtual subnets under a VPC. When starting the EC2 instance, the subnet id need to be specified so that the EC2 instance will be put into the subnet under the corresponding VPC.

EC2 instances under the VPC is completely isolated from the rest of Amazon's infrastructure at the network packet routing level (of course it is software-implemented isolation). Then a pair of gateway objects (VPN Gateway on the Amazon side and Customer gateway on the data center side) need to be created. Finally a connection object is created that binds these 2 gateway objects together and then attached to the VPC object.

After these steps, the two gateway will do the appropriate routing between your data center and the Amazon VPC with VPN technologies used underneath to protect the network traffic.


Things to watch out

While Amazon AWS provides a very transparent model for enterprise to migrate their existing IT infrastructure, there are a number of limitations that needs to pay attention to …
  • Multicast communication is not supported between EC2 instances. This means application has to communicate using TCP point-to-point protocol. Some cluster replication framework based on IP multicast simply doesn’t work in EC2 environment.
  • EBS currently can be attached to a single EC2 instance. This means some application (e.g. Oracle cluster) which based on having multiple machines accessing a shared disk simply won’t work in EC2 environment.
  • Except EC2, using any of the other API that Amazon provides is lock-in to Amazon’s technology stack. This issue may be somewhat mitigated as there are open source clone (e.g. Eucalyptus) to the Amazon AWS services

Sunday, November 16, 2008

Design for parallelism

There has been a lot of interests around parallel computing recently. One of the main reasons is that we all know the Moore's law (which promise to double the CPU power on a single chip every 18 months) has reached its limit. We cannot no expect the speed of a single CPU to go much further. Instead of attempting to advance the clock rate of a CPU, many of the chip manufacturer has shifted their development focus to multi-core machines.

On the other hand, highly scalable system based on large pool of inexpensive commodity hardware has demonstrated significant success. Google has published the Map/Reduce model which is their underlying computing infrastructure and there are open source clone like Apache Hadoop. All these provides a very rich framework for implementing massively parallel system.

However, most software algorithms that we are using today are sequential in nature. We need to refactor them in order to fit into the parallel computing architecture

How do we do that ?

There are two different approaches to restructure a sequential algorithm into parallel, “functional decomposition” is typically used to deal with complex logic flow; and “map reduce” is used to deal with algorithm with large volume of input data with simple logic flow.


Functional Decomposition

This model attempts to break down the sequential algorithm into multiple “work units” from a functionality perspective and see if different work units can be executed in parallel. The whole analysis and design will typically go through the following steps.

Decomposition

The purpose of this step is to identify the function boundary of each work unit, which is the basic unit of execution that occurs in a specific machine sequentially
  • Analyze the processing steps from a functionality boundary perspective. Break down the whole processing into a sequence of work units where each work unit represents a focused function.
  • At this stage, we typically breakdown to the finest level of granularity so that we have more flexibility in the design stage to maximize the degree of parallelism.
Dependency analysis

After we break down the whole process into the finest grain of work units, we analyze the sequential dependency between different work units.

Lets say workUnitB is following workUnitA in the sequential version of algorithm, and R(B) and W(B) represents the read set and write set of work unit B. Then workUnitB is directly dependent on workUnitA if any of the following conditions is true
  • W(B) and W(A) overlaps
  • R(B) and W(A) overlaps
  • W(B) and R(A) overlaps
If we represent each work unit as a node and each “directly dependent” relationship as an arc, we will end up having a DAG (directed acyclic graph). The DAG gives us a good picture about what is the maximum parallelism that we can obtain. The critical path of the DAG provides the lower bound of the total execution time.



Analyzing communication overhead
However, as data need to be fed from an upstream work unit to its downstream work units, communication is not free as it consumes bandwidth and latency. In fact, parallelism introduces communication and coordination overhead. This purpose of this step is to understand the associated communication cost when data flow between work units.

Depends on the chosen framework technology, the communication mechanism can be one of the following …
  • TCP Point to point: Persistent TCP connections are maintained between different machines and will be used to pass data between its residing work units.
  • Multicast pub/sub: Downstream work units subscribe their interests to upstream work units and use a multicast mechanism to deliver data. The implementation of multicast can be based on IP multicast or epidemic message spreading over an overlay network.
  • Queue: Upstream work unit put their result into a queue, which is polled by its downstream work units. FIFO semantics is provided.
  • DFS: Upstream work unit put their results into a distributed file system, which is consumed by downstream work units. Unlike a queue, the communicating work units need to synchronize their access to the DFS themselves.

Aggregating work units

The purpose of this step is to regroup the work unit into coarser granularity to reduce communication overhead.

For example, if workUnitA is feeding large amount of data into workUnitB, then both work units should be put into the same machine to reduce the network bandwidth consumption. When there are multiple work units residing in the same machine, then they can be further aggregated into a larger unit. This aggregation can reduce the number of nodes in the dependency graph and hence make the scheduling more straightforward.


Another DAG is produced at the end of this step where each node represents the work aggregate.


Schedule execution

The work aggregates eventually need to be executed in some machines in the network. It is the responsibility of the scheduler to ship the job to available processors, and synchronize their execution.

A node (in the DAG) is ready for execution when all the preceding nodes are completed. There is also a pool of idle processors. A simple-mind scheduler will schedule a ready-to-execute node to a randomly picked processor from the idle pool. After the processor finishes executing a node, it will report back to the scheduler which will update the DAG and the idle processor pool. The cycle repeats.

A more sophisticated scheduler will consider more factors such as the network bandwidth between processors, estimated execution time of each node … etc. in order to provide an optimal scheduling where network bandwidth consumption is minimized.


Map Reduce

For data intensive application, large amount of data need to be processed within a single work unit although the DAG itself is simple. In this model, just running different work unit in parallel is not sufficient, the execution within a work unit also need to be parallelized and run across multiple machines.

The design methodology is different here. Instead of focusing in the flow between work units, we need to focus the input data pattern of a single work unit. Map/Reduce model is a common choice to handle this scenario. The analysis and design will typically go through the following steps.
  1. Identify the repetition of input data, determine the basic unit of input record. ie: input
  2. Identify the selection criteria of each input record. ie: select() function
  3. For each input record, determine how many entries to be emitted and how the emit entries should be grouped and process together. ie: handle_map(), key(), value() function
  4. Determine the aggregation logic of grouped entries. ie: handle_reduce() function
  5. Identify the selection criteria of each aggregated result. ie: having() function
If we use the Map/Reduce framework such as Hadoop, we can structure the map() and reduce() function as follows:


Conclusion

By following a systematic methodology to transform a sequential application into parallel one, we can take advantage of the parallelism to make the application more scalable.