Transactions Come and Go, but Events are Forever

As record-keeping first became computerized, the high price of compute, memory, and storage meant that it was too expensive to keep every event. To reduce costs, the database world endeavored to make it possible to only keep the current state, and thus gave birth to the “transaction”. Databases and applications that focused entirely on the current state proliferated and became what we know of today as On-Line Transaction Processing (OLTP). After a few decades of OLTP, methods to automate reporting and analysis of transactions were defined as On-Line Analytics Processing (OLAP).  Now, after a few decades of exploration of both OLTP and OLAP, we are seeing a new split emerge, not on the basis of “transactional processing versus analytical processing” but on the basis of “Transactions versus Events”.

Early use of event processing, by Midjourney

This new split of transactions versus events is giving rise to a new set of infrastructure: HTAP as a one-stop shop for “Transactions” and “Event Stores” as a one-stop shop for Events.  But, we are now jumping ahead of ourselves.  Before we can reach that conclusion, we need to first go back and discuss the difference between events and transactions.

Events make Transactions, but Transactions do not make Events

We, as people, are all a product of our past experiences.  We have seen different things, experienced different things and all of those serve as inputs into who we are as human beings.  This very directly mimics the difference between Transactions and Events, all of our experiences as humans are Events that happened in our lives, but who we are right now, here, today, that is the Transactional representation of us.  Even this description is probably too meta, so let’s actually boil this down to a real-world example: T-shirts!

First, we’ll explain Events by purchasing a t-shirt.  When we make that purchase, we are essentially creating an “Event.”  Events happen at a point in time, for a very specific amount of money, and these purchases never change: if you purchased something (such as a new shirt) yesterday, even if you return it today, it doesn’t change the fact that you purchased it yesterday.  This specific notion of “never changing” is a central theme for Events, they are occurrences that will never change.

Purchasing a shirt in a store, with data, by midjourney

When we purchase that shirt, however, the inventory of shirts in the store changes.  If there were 10 shirts, there are now 9, this is a Transaction.  The inventory of shirts in the store is defined by its change and its most important property is that it is up-to-date.  If the store manager is looking at the inventory and it is not up-to-date, there is a risk of purchasing too much stuff or running out of stock.  Just as “never changing” is a central theme for Events, “changeable” is a central theme for a Transaction.

So, what is the relationship between your T-shirt purchases and the inventory management at the store?

Well, let’s start by discussing the world of that store manager who is managing their inventory.  If they have a consistent view of how many of each shirt that they have, then it is relatively easy to decide if you have enough inventory right this moment.  But, if all you know is your current inventory, you do not have enough information to decide how much inventory you will need in the future.  Likely, you pick a level of inventory, say 10 of each shirt, and just try to keep that much in stock.  But, do you tend to sell more Medium size shirts? Large shirts? XXXL shirts?  Are there days of the week that you tend to sell more shirts and days when you sell less?  These questions are hard to answer with only the current stock levels available: the best you can do is notice that you are running low on inventory and then stocking back up.

When you want to get a bit more intelligent about what the right level of inventory is and whether you need to be restocked by tomorrow or if it can wait til next week, you need more information.  You might start keeping track of how much inventory you sold each day, and then you can go back and look at the history of days to get a better idea of how much and how quickly you need to restock.  Even better than waiting for the end of the day to figure out how much inventory changed, you could also keep a record of each and every interaction, counting each sale as it happens.  Whether you do this every evening or on each sale, you are recording Events: the fact that someone purchased that inventory never changes.  

But, what about if we return that shirt?  Doesn’t that reverse the purchase of the item?  This is where the separation between a Transaction and an Event becomes even more important.  When that T-shirt gets returned, the inventory increases again and it looks like nobody ever bought the shirt in the first place, so the “Transactional View” would end up having you believe that nobody purchased anything.  However, that doesn’t change the fact that it actually was purchased, the Event view would actually register 2 separate Events: one for the purchase and one for the return.

In this way, as each new purchase or return occurs, a new Event happens.  With each of these events, the inventory level increases or decreases: a Transaction change.  Events and Transactions exist everywhere and follow this same relationship.  The Transaction is a view of a specific state that sometimes changes when a new Event happens..

But, Don’t Databases already have Transactions?

A Transaction is a state that changes, but databases and especially OLTP databases have been defined by their support for “database transactions” for many decades.  Truly, it’s right in the name: OLTP = On-Line Transaction Processing, but the section above doesn’t really sound like a “transaction” from a Databases 101 class, so, what gives?

These are very deeply related concepts and that’s why we are choosing to reuse the label “Transaction”, but there is a slight difference in nuance.  Specifically, database transactions exist to ensure the correctness of changes applied to the Transaction.  Going back to our purchasing example, let’s say that both I and a family member are shopping online.  We both try to return the same item at the same time.  The store should accept only one return and reject the other, but which one?  The simplest answer to this is “the first one should be accepted and the latter one rejected.”  But, if they truly happened at the exact same instant, how do we know which one is the first one and which one is the second?  This is the problem that a database transaction solves: it gives us a way to ensure that one return is applied before the other, giving us the ability to accept one and reject the other.

Databases put great importance on making each database transaction ACID-compliant, guaranteeing Atomicity, Consistency, Isolation, and Durability. Without this, we couldn’t correctly apply the Events (our product returns) to the Transaction (the inventory level) in the database.

Even with database transactions ensuring that only one return gets processed, we still have a chance for a bad experience.  If my return was the latter one and it got rejected, I will want to know why it was rejected.  Specifically, I need to be able to see some record that the item I believe I’m returning was already processed and returned by my family member, so that I can double check with them that we both tried to do the same thing.

Taking this back to the generic “Transaction” versus “Event” terminology, a database transaction exists to enforce the orderly application of Event(s) to the state of a Transaction.  It is a very necessary and powerful tool in the OLTP tool chest, but in an OLTP system, the database transaction fundamentally exists to allow us to minimize storage by applying Events to the Transaction and only store the most up-to-date, final state of the object.

At least, that was a fundamentally important element to dealing with data in the past.  As technology has evolved, storage has become cheap while compute is ever-more accessible and usually available on demand. We’re evolved from a world where Events were too expensive to store to the modern world where the insights from Events are too valuable to lose.

What if we used Infrastructure designed to keep all of those Events?

All of this might sound like a new concept, like we are preaching some new doctrine, some new religion that was just discovered.  But, in reality, we are not.  Event-native infrastructure has been brewing for many years. Perhaps it started with streaming systems, where every message put into a Kafka topic is fundamentally equivalent to an Event.  For analytics, data is commonly copied from OLTP systems to OLAP systems using an Extract-Load-Transform process. When this ELT process is completed, the resulting data is very commonly a ‘reverse conversion’ from a Transaction into Events (for example, the process might subtract today’s inventory from yesterday’s to figure out how many of each item was purchased).

Either way, the growth of the Internet has driven a fundamental shift in infrastructure, creating a need and development towards infrastructure that can deal with Events directly and natively.  The first large-scale uses of event processing were driven by a very common use case: Web Analytics.  Every interaction with a web site or a mobile app is an Event. Understanding the flow and how people go through logins, signups, and other activities is a collection of events, but not really a Transaction. 

Events are streaming, by Midjourney

While events can be stored in databases designed for Transactions, the requirements to implement ACID-compliant database operations are usually at odds with the requirements to address the scale, performance, and concurrency needs of working with Events.  Given that Events are forever constant, adding them to a database only requires an append and does not require changing any other state.  This is a fundamental assumption that must be leveraged to truly scale out to the 10s, 100s, even 1000s of millions of Events per second that exist in our physical world.  Infrastructure not built for this world often tries to overcome the cost of supporting ACID-compliant changes to state through adopting a microbatching strategy, converting a stream of Events to a series of database transactions.

Event databases house Events as the immutable pieces of treasure that they are. Event databases are built to store the Events and make them accessible so that it is always possible to understand the series of Events that caused something to happen. Since the Events are often arriving as a real-time stream, these databases are often called Real-Time Analytics Databases.

Diving deeper into our shopping example, we can take a case study from the realm of e-commerce.  Amazon.com and other pioneers of e-commerce realized early in their journey that some functions, such as inventory tracking and financial settlement, need ACID-compliant transactions, but far more valuable information is in the flow of events. Understanding how each customer comes to the e-storefront, navigates through offerings, responds to recommendations, builds shopping carts (or, sometimes, abandons those carts), and changes patterns at different times of day, days of the week, and seasons of the year provides the knowledge to build lasting relationships. The events are transported as a data stream to an Event Store, where they can be evaluated and used to generate insights.

This need exists in all industries and beyond. Search engines use Events to validate the relevance of search results.  Medical devices like a glucose meter can provide a stream of Events that can be leveraged to orchestrate an insulin pump to administer insulin while maintaining an auditable trail of the state of a patient and Event data to power predictive analytics and future research.  Smart devices in a home provide streams of Events from their sensors which can be combined and aggregated to understand the overall efficiency of a building, or what time of day a household consumes their electricity.

Our Event-Driven Future

OLTP and OLAP are mature, and together they form the basis of transaction processing and analytics. Relational databases, non-relational databases, and analytics databases, are each designed to store, retrieve, and otherwise work with Transactions. We have found ways of scaling these databases, culminating in the advent and expansion of Hybrid Transactional/Analytic Processing (HTAP) databases which promise to remove the boundaries between OLAP and OLTP, for Transactions, still focusing on computing and analyzing a specific state at a specific point in time.

Future Data with Event Streams by midjourney

There is another world that is emerging from this that HTAP doesn’t address, it is the Event-focused world, where you want to understand the actions taken and provenance of a specific state more than just the current state.  While, on the surface, it might seem like Transaction-oriented systems should be able to do what an Event-oriented system can do, this is the same argument that existed for many years that an OLTP system can do the same things that an OLAP system can do.  

For low levels of scale, the sheer power of compute these days can overcome the inefficiencies of using a Transaction-oriented system to work with Events.  But, there will always exist use cases and points of scale where an Event-oriented system will be required to provide the best experience at the best overall cost. 

The fundamental assumptions of how you want to interact with Events versus Transactions are different.  When you think about it, all actions taken by humans and machines are Events.  So, the needs of Event-oriented infrastructure are not likely to disappear any time soon.

(If you’d like an easy way to try out an event database, get a free trial of Imply Polaris, the database-as-a-service built from Apache Druid.)