Author Archives: Leo Cheung

Scala IoT Systems With Akka Actors II

In a previous blog post, I assembled a Scala application simplified from an IoT prototype using Akka Actors and MQTT to illustrate how an IoT system fits into the selected tech stack. The stripped-down application uses a single actor to simulate requests from a bunch of IoT devices.

In this post, I would like to share an expanded version of the previous application that uses loosely-coupled lightweight actors to simulate individual IoT devices, each of which maintains its own internal state and handles bidirectional communications via non-blocking message passing. Using a distributed workers system adapted from a Lightbend template along with a persistence journal, the end product is an IoT system equipped with a scalable fault-tolerant data processing system.

Main components

Below is a diagram and a summary of the revised Scala application which consists of 3 main components:

IoT with MQTT and Akka Actor Systems v.2

1. IoT

  • An IotManager actor which:
    • instantiates a specified number of devices upon start-up
    • subscribes to a MQTT pub-sub topic for the work requests
    • sends received work requests via ClusterClient to the master cluster
    • notifies Device actors upon receiving failure messages from Master actor
    • forwards work results to the corresponding devices upon receiving them from ResultProcessor
  • Device actors each of which:
    • simulates a thermostat, lamp, or security alarm with random initial state and setting
    • maintains and updates internal state and setting upon receiving work results from IotManager
    • generates work requests and publishes them to the MQTT pub-sub topic
    • re-publishes requests upon receiving failure messages from IotManager
  • A MQTT pub-sub broker and a MQTT client for communicating with the broker
  • A configuration helper object, MqttConfig, consisting of:
    • MQTT pub-sub topic
    • URL for the MQTT broker
    • serialization methods to convert objects to byte arrays, and vice versa

2. Master Cluster

  • A fault-tolerant decentralized cluster which:
    • manages a singleton actor instance among the cluster nodes (with a specified role)
    • delegates ClusterClientReceptionist on every node to answer external connection requests
    • provides fail-over of the singleton actor to the next-oldest node in the cluster
  • A Master singleton actor which:
    • registers Workers and distributes work to available Workers
    • acknowledges work request reception with IotManager
    • publishes work results from Workers to ‘work-results’ topic via Akka distributed pub-sub
    • maintains work states using persistence journal
  • A ResultProcessor actor in the master cluster which:
    • gets instantiated upon starting up the IoT system (more on this below)
    • consumes work results by subscribing to the ‘work-results’ topic
    • sends work results received from Master to IotManager

3. Workers

  • An actor system of Workers each of which:
    • communicates via ClusterClient with the master cluster
    • registers with, pulls work from the Master actor
    • reports work status with the Master actor
    • instantiates a WorkProcessor actor to perform the actual work
  • WorkProcessor actors each of which:
    • processes the work requests from its parent Worker
    • generates work results and send back to Worker

Master-worker system with a ‘pull’ model

While significant changes have been made to the IoT actor system, much of the setup for the Master/Worker actor systems and MQTT pub-sub messaging remains largely unchanged from the previous version:

  • As separate independent actor systems, both the IoT and Worker systems communicate with the Master cluster via ClusterClient.
  • Using a ‘pull’ model which generally performs better at scale, the Worker actors register with the Master cluster and pull work when available.
  • Paho-Akka is used as the MQTT pub-sub messaging client.
  • A helper object, MqttConfig, encapsulates a MQTT pub-sub topic and broker information along with serialization methods to handle MQTT messaging using a test Mosquitto broker.

What’s new?

Now, let’s look at the major changes in the revised application:

First of all, Lightbend’s Activator has been retired and Sbt is being used instead.

On persisting actors state, a Redis data store is used as the persistence journal. In the previous version the shared LevelDB journal is coupled with the first seed node which becomes a single point of failure. With the Redis persistence journal decoupled from a specific cluster node, fault tolerance steps up a notch.

As mentioned earlier in the post, one of the key changes to the previous application is the using of actors representing individual IoT devices each with its own state and capability of communicating with entities designated for interfacing with external actor systems. Actors, lightweight and loosely-coupled by design, serve as an excellent vehicle for modeling individual IoT devices. In addition, non-blocking message passing among actors provides an efficient and economical means for communication and logic control of the device state.

The IotManager actor is responsible for creating and managing a specified number of Device actors. Upon startup, the IoT manager instantiates individual Device actors of random device type (thermostat, lamp or security alarm). These devices are maintained in an internal registry regularly updated by the IoT manager.

Each of the Device actors starts up with a random state and setting. For instance, a thermostat device may start with an ON state and a temperature setting of 68F whereas a lamp device might have an initial state of OFF and brightness setting of 2. Once instantiated, a Device actor will maintain its internal operational state and setting from then on and will report and update the state and setting per request.

Work and WorkResult

In this application, a Work object represents a request sent by a specific Device actor and carries the Device’s Id and its current state and setting data. A WorkResult object, on the other hand, represents a returned request for the Device actor to update its state and setting stored within the object.

Responsible for processing the WorkResult generated by the Worker actors, the ResultProcessor actor simulates the processing of work result – in this case it simply sends via the actorSelection method the work result back to the original Device actor through IotManager. Interacting with only the Master cluster system as a cluster client, the Worker actors have no knowledge of the ResultProcessor actor. ResultProcessor receives the work result through subscribing to the Akka distributed pub-sub topic which the Master is the publisher.

While a participant of the Master cluster actor system, the ResultProcessor actor gets instantiated when the IoT actor system starts up. The decoupling of ResultProcessor instantiation from the Master cluster ensures that no excessive ResultProcessor instances get started when multiple Master cluster nodes start up.

Test running the application

Complete source code of the application is available at GitHub.

To run the application on a single JVM, just git-clone the repo, run the following command at a command line terminal and observe the console output:

The optional NumOfDevices parameter defaults to 20.

To run the application on separate JVMs, git-clone the repo to a local disk, open up separate command line terminals and launch the different components on separate terminals:

Sample console log

Below is filtered console log output from the console tracing the evolving state and setting of a thermostat device:

The following annotated console log showcases fault-tolerance of the master cluster – how it fails over to the 2nd node upon detecting that the 1st node crashes:

Scaling for production

While the application has an underlying architecture that emphasizes on scalability, it would require further effort in the following areas to make it production ready:

  • IotManager uses the ‘ask’ method for message receipt confirmation via a Future return by the Master. If business logic allows, using the fire-and-forget ‘tell’ method will be significantly more efficient especially at scale.
  • The MQTT broker used in the application is a test broker provided by Mosquitto. A production version of the broker should be installed preferably local to the the IoT system. MQTT brokers from other vendors like HiveMQ, RabbitMQ are also available.
  • As displayed in the console log when running the application, Akka’s default Java serializer isn’t best known for its efficiency. Other serializers such as Kryo, Protocol Buffers should be considered.
  • The Redis data store for actor state persistence should be configured for production environment

Further code changes to be considered

A couple of changes to the current application might be worth considering:

Device types are currently represented as strings, and code logic for device type-specific states and settings is repeated during instantiation of devices and processing of work requests. Such logic could be encapsulated within classes defined for individual device types. The payload would probably be larger as a consequence, but it might be worth for better code maintainability especially if there are many device types.

Another change to be considered is that Work and WorkResult could be generalized into a single class. Conversely, they could be further differentiated in accordance with specific business needs. A slightly more extensive change would be to retire ResultProcessor altogether and let Worker actors process WorkResult as well.

State mutation in Akka Actors

In this application, a few actors maintain mutable internal states using private variables (private var):

  • Master
  • IotManager
  • Device

As an actor by-design will never be accessed by multiple threads, it’s generally safe enough to use ‘private var’ to store changed states. But if one prefers updating states in the way like a state machine, Akka Actors provides a method to hot-swap an actor’s internal state.

Hot-swapping like state machine

Below is a sample snippet that illustrates how hot-swapping mimics a state machine without having to use any mutable variable for maintaining the actor state:

Simplified for illustration, the above snippet depicts a Worker actor that pulls work from the Master cluster. The context.become method allows the actor to switch its internal state at run-time like a state machine. As shown in the simplified code, it takes an ‘Actor.Receive’ (which is a partial function) that implements a new message handler. Under the hood, Akka manages the hot-swapping via a stack. As a side note, according to the relevant source code, the stack for hot-swapping actor behavior is, ironically, a mutable ‘private var’ of List[Actor.Receive].

Recursive transformation of immutable parameter

Another functional approach to mutating actor state is via recursive transformation of an immutable parameter. As an example, we can avoid using a mutable ‘private var registry’ as shown in the following ActorManager actor and use ‘context.become’ to recursively transform a registry as an immutable parameter passed to be updateState method:

Text Mining With Akka Streams

Reactive Systems, whose core characteristics are declared in the Reactive Manifesto, have started to emerge in recent years as message-driven systems that emphasize scalability, responsiveness and resilience. It’s pretty clear from the requirements that a system can’t be simply made Reactive. Rather, it should be built from the architectural level to be Reactive.

Akka’s actor systems, which rely on asynchronous message-passing among lightweight loosely-coupled actors, serve a great run-time platform for building Reactive Systems on the JVM (Java Virtual Machine). I posted a few blogs along with sample code about Akka actors in the past. This time I’m going to talk about something different but closely related.

Reactive Streams

While bearing a similar name, Reactive Streams is a separate initiative that mandates its implementations to be capable of processing stream data asynchronously and at the same time automatically regulating the stream flows in a non-blocking fashion.

Akka Streams, built on top of Akka actor systems, is an implementation of Reactive Streams. Equipped with the back-pressure functionality, it eliminates the need of manually buffering stream flows or custom-building stream buffering mechanism to avoid buffer overflow problems.

Extracting n-grams from text

In text mining, n-grams are useful data in the area of NLP (natural language processing). In this blog post, I’ll illustrate extracting n-grams from a stream of text messages using Akka Streams with Scala as the programming language.

First thing first, let’s create an object with methods for generating random text content:

Source code: TextMessage.scala

Some minimal effort has been made to generate random clauses of likely pronounceable fake words along with punctuations. To make it a little more flexible, lengths of individual words and clauses would be supplied as parameters.

Next, create another object with text processing methods responsible for extracting n-grams from input text, with n being an input parameter. Using Scala’s sliding(size, step) iterator method with size n and step default to 1, a new iterator of sliding window view is generated to produce the wanted n-grams.

Source code: TextProcessor.scala

Now that the text processing tools are in place, we can focus on building the main streaming application in which Akka Streams plays the key role.

First, make sure we have the necessary library dependencies included in build.sbt:

Source code: build.sbt

As Akka Streams is relatively new development work, more recent Akka versions (2.4.9 or higher) should be used.

Let’s start with a simple stream for this text mining application:

Source code: NgramStream_v01.scala

As shown in the source code, constructing a simple stream like this is just defining and chaining together the text-generating source, the text-processing flow and the text-display sink as follows:

Graph DSL

Akka Streams provides a Graph DSL (domain-specific language) that helps build the topology of stream flows using predefined fan-in/fan-out functions.

What Graph DSL does is somewhat similar to how Apache Storm‘s TopologyBuilder pieces together its spouts (i.e. stream sources), bolts (i.e. stream processors) and stream grouping/partitioning functions, as illustrated in a previous blog about HBase streaming.

Back-pressure

Now, let’s branch off the stream using Graph DSL to illustrate how the integral back-pressure feature is at play.

Source code: NgramStream_v02.scala

Streaming to a file should be significantly slower than streaming to the console. To make the difference more noticeable, a delay is deliberately added to streaming each line of text in the file sink.

Running the application and you will notice that the console display is slowed down. It’s the result of the upstream data flow being regulated to accommodate the relatively slow file I/O outlet even though the other console outlet is able to consume relatively faster – all that being conducted in a non-blocking fashion.

Graph DSL create() methods

To build a streaming topology using Graph DSL, you’ll need to use one of the create() methods defined within trait GraphApply, which is extended by object GraphDSL. Here are the signatures of the create() methods:

Note that the sbt-boilerplate template language is needed to interpret the create() method being used in the application that takes multiple stream components as input parameters.

Materialized values

In Akka Streams, materializing a constructed stream is the step of actually running the stream with the necessary resources. To run the stream, the implicitly passed factory method ActorMaterializer() is required to allocate the resources for stream execution. That includes starting up the underlying Akka actors to process the stream.

Every processing stage of the stream can produce a materialized value. By default, using the via(flow) and to(sink) functions, the materialized value of the left-most stage will be preserved. As in the following example, for graph1, the materialized value of the source is preserved:

To allow one to selectively capture the materialized values of the specific stream components, Akka Streams provides functions viaMat(flow) and toMat(sink) along with a combiner function, Keep. As shown in the above example, for graph2, the materialized value of the flow is preserved, whereas for graph3, materialized values for both the flow and sink are preserved.

Back to our fileSink function as listed below, toMat(fileIOSink)(Keep.right) instructs Akka Streams to keep the materialized value of the fileIOSink as a Future value of type IOResult:

Using Graph DSL, as seen earlier in the signature of the create() method, one can select what materialized value is to be preserved by specifying the associated stream components accordingly as the curried parameters:

In our case, we want the materialized value of fileSink, thus the curried parameters should look like this:

Defining the stream graph

Akka Streams provides a number of functions for fan-out (e.g. Broadcast, Balance) and fan-in (e.g. Merge, Concat). In our example, we want a simple topology with a single text source and the same n-gram generator flow branching off to two sinks in parallel:

Adding a message counter

Let’s further expand our n-gram extraction application to include displaying a count. A simple count-flow is created to map each message string into numeric 1, and a count-sink to sum up all these 1′s streamed to the sink. Adding them as the third flow and sink to the existing stream topology yields something similar to the following:

Source code: NgramStream_v03.scala

Full source code of the application is at GitHub.

Final thoughts

Having used Apache Storm, I see it a rather different beast compared with Akka Streams. A full comparison between the two would obviously be an extensive exercise by itself, but it suffices to say that both are great platforms for streaming applications.

Perhaps one of the biggest differences between the two is that Storm provides granular message delivery options (at most / at least / exactly once, guaranteed message delivery) whereas Akka Streams by design questions the premise of reliable messaging on distributed systems. For instance, if guaranteed message delivery is a requirement, Akka Streams would probably not be the best choice.

Back-pressure has recently been added to Storm’s v.1.0.x built-in feature list, so there is indeed some flavor of reactiveness in it. Aside from message delivery options, choosing between the two technologies might be a decision basing more on other factors such as engineering staff’s expertise, concurrency model preference, etc.

Outside of the turf of typical streaming systems, Akka Streams also plays a key role as the underlying platform for an emerging service stack. Viewed as the next-generation of Spray.io, Akka HTTP is built on top of Akka Streams. Designed for building HTTP-based integration layers, Akka HTTP provides versatile streaming-oriented HTTP routing and request/response transformation mechanism. Under the hood, Akka Streams’ back-pressure functionality regulates data streaming between the server and the remote client, consequently conserving memory utilization on the server.

PostgreSQL Table Partitioning

With the ever growing demand for data science work in recent years, PostgreSQL has gained superb popularity especially in areas where extensive geospatial/GIS (geographic information system) functionality is needed. In a previous startup venture, MySQL was initially adopted and I went through the trouble of migrating to PostgreSQL mainly because of the sophisticated geospatial features PostGIS offers.

PostgreSQL offers a lot of goodies, although it does have a few things that I wish were done differently. Most notable to me is that while its SELECT statement supports SQL-92 Standard’s JOIN syntax, its UPDATE statement would not. For instance, the following UPDATE statement would not work in PostgreSQL:

Partial indexing

Nevertheless, for general performance and scalability, PostgreSQL remains one of the top candidates with proven track record in the world of open source RDBMS. In scaling up a PostgreSQL database, there is a wide variety of approaches. Suitable indexing are probably some of the first strategies to be looked into. Aside from planning out proper column orders in indexes that are optimal for the frequently used queries, there are also a couple of indexing features that help scaling.

Partial indexing allows an index to be built over a subset of a table based on a conditional expression. For instance:

In the case of a table with large amount of rows, this feature could make an otherwise gigantic index much smaller, thus more efficient for queries against the selectively indexed data.

Scaling up with table partitioning

However, when a table grows to certain volume, say, beyond a couple of hundreds of million rows, and if periodically archiving off data from the table isn’t an option, it would still be a problem even with applicable indexing strategy. In many cases, it might be necessary to do something directly with the table structure and table partitioning is often a good solution.

There are a few approaches to partition a PostgreSQL table. Among them, partitioning by means of table inheritance is perhaps the most popular approach. A master table will be created as a template that defines the table structure. This master table will be empty whereas a number of child tables inherited from this master table will actually host the data.

The partitioning is based on a partition key which can be a column or a combination of columns. In some common use cases, the partition keys are often date-time related. For instance, a partition key could be defined in a table to partition all sales orders by months with constraint like the following:

order_date >= ’2016-12-01 00:00:00′ AND order_date < ’2017-01-01 00:00:00′

Other common cases include partitioning geographically, etc.

A table partitioning example

When I was with a real estate startup building an application that involves over 100 millions nationwide properties, each with multiple attributes of interest, table partitioning was employed to address the demanding data volume. Below is a simplified example of how the property sale transaction table was partitioned to maintain a billion rows of data.

First, create the master table which will serve as the template for the table structure.

Next, create child tables inheriting from the master table for the individual states. For simplicity, I only set up 24 states for performance evaluation.

Nothing magical so far, until a suitable trigger for propagating insert is put in place. The trigger essentially redirects insert requests against the master table to the corresponding child tables.

Let’s test inserting data into the partitioned tables via the trigger:

A Python program for data import

Now that the master table and its child tables are functionally in place, we’re going to populate them with large-scale data for testing. First, write a simple program using Python (or any other programming/scripting language) as follows to generate simulated data in a tab-delimited file for data import:

Run the Python program to generate up to 1 billion rows of property sale data. Given the rather huge output, make sure the generated file is on a storage device with plenty of space. Since it’s going to take some time to finish the task, it would better be run in the background, perhaps along with mail notification, like the following:

Next, load data from the generated infile into the partitioned tables using psql. In case there are indexes created for the partitioned tables, it would generally be much more efficient to first drop them and recreate them after loading the data, like in the following:

Query with Constraint Exclusion

Prior to querying the tables, make sure the query optimization parameter, constraint_exclusion, is enabled.

With constraint exclusion enabled, the query planner will be smart enough to examine query constraints to exclude scanning of those partitioned tables that don’t match the constraints. Unfortunately, though, if the constraints involve matching against non-constants like the NOW() function, the query planner won’t have enough information to filter out unwanted partitions hence won’t be able to take advantage of the optimization.

Final notes

With a suitable partitioning scheme applied to a big table, query performance can be improved by an order of magnitude. As illustrated in the above case, the entire partitioning scheme centers around the key column used for partitioning, hence it’s critical to properly plan out which key column (or combination of columns) to partition. Number of partitions should also be carefully thought out, as too few partitions might not help whereas too many partitions would create too much overhead.

Relational Database Redemption

Relational databases, such as PostgreSQL and Oracle, can be traced back to the 80′s when they became a dominant type of data management systems. Their prominence was further secured by the ANSI standardization of the domain specific language called SQL (Structured Query Language). Since then, RDBMS (relational database management system) has been the de facto component most data-centric applications would be architecturally centered around.

What happened to relational Databases?

It’s a little troubling, though, over the past 10-15 years, I’ve witnessed relational databases being sidelined from the core functionality requirement review or architectural design in many software engineering projects that involve data-centric applications. In particular, other kinds of databases would often be favored for no good reasons. And when relational databases are part of the core technology stack, thorough data model design would often be skipped and using of SQL would often be avoided in cases where it would be highly efficient.

So, why have relational databases been treated with such noticeably less preference or seriousness? I believe a couple of causes have led to the phenomenon.

Object-oriented data persistence architecture

First, there was a shift in application architecture in late 90′s when object-oriented programming began to increasingly dominate in the computing world. In particular, the backend data persistence component of object-oriented applications began to take over the heavy lifting of the database CRUD (create/read/update/delete) operations which used to reside within the database tier via SQL or procedural language PL/SQL.

Java EJB (enterprise Java bean), which was aimed to emulate data persistence and query functionality among other things, took the object-oriented programming world by storm. ORM (object-relational mapping) then further helped keep software engineers completely inside the Object world. Realizing that the initial EJB specifications were over-engineered, it later evolved into JPA (Java Persistence API) which also incorporates ORM functionality. All that doesn’t eliminate the need of relational databases, but engineering design focus has since been pulled away from the database tier and SQL has been treated as if it was irrelevant.

NoSQL databases

Then, in late 00′s came column-oriented NoSQL databases like HBase and Cassandra, which were designed to primarily handle large-scale datasets. Designed to run on scalable distributed computing platforms, these databases are great for handling Big Data at the scale that conventional relational databases would have a hard time to perform well.

Meanwhile, document-based NoSQL databases like MongoDB also emerged and have increasingly been adopted as part of the core technology stack by software engineers. These NoSQL databases have all of a sudden stole the spotlight in the database world. Relational databases were further perceptually “demoted” and SQL wouldn’t look right without a negation prefix.

Object-oriented data persistence versus SQL, PL/SQL

Just to be clear, I’m not against having the data persistence layer of the application handle the business logic of data manipulations and queries within the Object world. In fact, I think it makes perfect sense to keep data access business logic within the application tier using the same object-oriented programming paradigm, shielding software engineers from having to directly deal with things in the disparate SQL world.

Another huge benefit of using the object-oriented data persistence is that it takes advantage of any scaling mechanism provided by the application servers (especially for those on distributed computing platforms), rather than, say, relying everything on database-resident PL/SQL procedures that don’t scale well.

What I’m against, though, is that proper design and usage best practices are skipped when a relational database is used, hallucinating that the ORM would just magically handle all the data manipulations/queries of a blob of poorly structured data. In addition, while ORMs can automatically generate SQLs for a relatively simple data model, they aren’t good at coming up with optimal efficient SQLs for many sophisticated models in the real world.

NoSQL databases versus Relational databases

Another clarification point I thought I should raise is that – I love both SQL-based relational and NoSQL databases, and have adopted them as core parts of different systems in the past. I believe they have their own sweet spots as well as drawbacks, and should be adopted in accordance with the specific need in data persistence and consumption.

I’ve seen some engineering organizations flocking to the NoSQL world for valid reasons, and others just for looking cool. I’ve also seen in a couple of occasions that companies decided to roll back from a NoSQL platform to using relational databases to better address their database transaction need after realizing that their increasing data volume demand can actually be handled fine with a properly designed relational database system.

In general, if your database need leans towards data warehousing and the projected data volume is huge, NoSQL is probably a great choice; otherwise, sticking to using relational databases might be the best deal. It all boils down to specific business requirement, and these days it’s also common that both database types are simultaneously adopted to complement each other. As to what’s considered huge, I would say it warrants a NoSQL database solution when one or more tables need to house 100′s of millions or more rows of data.

Why do relational databases still matter?

The answer to whether relational databases still matter is a decisive yes:

  1. Real-world need of relational data models — A good portion of structured and inter-related data in the real world is still best represented by relational data models. While column-oriented databases excel in handling very large datasets, they aren’t designed for modeling relational data entities.
     
  2. Transactional CRUD operations — Partly due to NoSQL database’s fundamental design, data often need to be stored in denormalized form for performance, and that makes transactional operations difficult. On the contrary, relational database is a much more suitable model for transactional CRUD operations that many types of applications require. That, coupled with the standard SQL language for transactional CRUD makes the role of relational databases not easily replaceable.
     
  3. Bulk data manipulations — Besides proven a versatile powerful tool in handling transactional CRUD, SQL also excels in manipulating data in bulk without compromise in atomicity. While PL/SQL isn’t suitable for all kinds of data manipulation tasks, when used with caution it provides procedural functionality in bulk data processing or complex ETL (extract-transform-load).
     
  4. Improved server hardware — Improvement in server processing power and low cost of memory and storage in recent years have helped make relational databases cope with the increasing demand of high data volume. On top of that, prominent database systems are equipped with robust data sharding and clustering features that also decidedly help in scalability. Relational databases with 10′s or even 100′s of million rows of data in a table aren’t uncommon these days.
     

Missing skills from today’s software architects

In recent years, I’ve encountered quite a few senior software engineers/architects with advanced programming skills but poor relational data modeling/SQL knowledge. With their computing backgound I believe many of these engineers could pick up the essential knowledge without too much effort. (That being said, I should add that while commanding the relational database fundamentals is rather trivial, becoming a database guru does require some decent effort.) It’s primarily the lack of drive to sharpen their skills in the specific domain that has led to the said phenomenon.

The task of database design still largely falls on the shoulders of the software architect. Most database administrators can configure database systems and fine-tune queries at the operational level to ensure the databases are optimally run, but few possess business requirement knowledge or, in many cases, skills for database design. Suitable database design and data modeling requires intimate knowledge and understanding of business logic of the entire application that is normally in the software architect’s arena.

Even in the NoSQL world of column-oriented databases, I’ve noticed that database design skills are also largely missing. Part of NoSQL database’s signature is that data columns don’t have to be well-defined upfront and can be added later as needed. Because of that, many software architects tend to think that they have the liberty to bypass proper schema design upfront. The truth is that NoSQL databases do need proper schema design as well. For instance, in HBase, due to the by-design limitation of indexing, one needs to carefully lay out upfront what the row key is comprised of and what column families will be maintained.

Old and monolithic?

Aside from causes related to the disruptive technologies described above, some misconceptions that associate relational databases with obsolete technology or monolithic design have also helped contribute to the unwarranted negative attitude towards RDBMS.

Old != Obsolete — Relational database technology is old. Fundamentally it hasn’t changed since decades ago, whereas new computing and data persistence technology buzzwords keep popping up left and right non-stopped. Given so many emerging technologies that one wants to learn all at once, old RDBMS often gets placed at the bottom of the queue. In any case, if a technology is old but continues to excel within its domain, it isn’t obsolete.

RDBMS != Monolith — Contemporary software architects have been advocating against monolithic design. In recent years, more and more applications have been designed and built as microservices with isolated autonomous services and data locality. That’s all great stuff in the ever-evolving software engineering landscape, but when people automatically categorize an application with a high-volume relational database a monolithic system, that’s a flawed assumption.

Bottom line, as long as much of the data in the real world is still best represented in relational data models, RDBMS will have its place in the computing world.