Your CIO is all fired up about moving your legacy inventory management app to the Cloud. Lower capital costs! Dynamic provisioning! Outsourced infrastructure! So you get out your shoehorn, provision some storage and virtual machine instances, and forklift the whole mess into the stratosphere. (OK, there’s more to it than that, but bear with me.)
Everything seems to work at first. But then the real test comes: the Holiday season, when you do most of your online business. You breathe a sigh of relief as your Cloud provider seamlessly scales up to meet the spikes in demand. But then your boss calls, irate. Turns out customers are swamping the call center with complaints of failed transactions.
You frantically dive into the log files and diagnostic reports to see what the problem is. Apparently, the database has not been keeping an accurate count of your inventory—which is pretty much what an inventory management system is all about. You check the SQL, and you can’t find the problem. Now you’re really beginning to sweat.
You dig deeper, and you find the database is frequently in an inconsistent state. When the app processes orders, it decrements the product count. When the count for a product drops to zero, it’s supposed to show customers that you’ve run out. But sometimes, the count is off. Not always, and not for every product. And the problem only seems to occur in the afternoons, when you normally experience your heaviest transaction volume.
The Problem: Consistency in the Cloud
The problem is that while it may appear that your database is running in a single storage partition, in reality the Cloud provider is provisioning multiple physical partitions as needed to provide elastic capacity. But when you look at the fine print in your contract with the Cloud provider, you realize they offer eventual consistency, not immediate consistency. In other words, your data may be inconsistent for short periods of time, especially when your app is experiencing peak load. It may only be a matter of seconds for the issue to resolve, but in the meantime, customers are placing orders for products that aren’t available. You’re charging their credit cards and all they get for their money is an error page.
From the perspective of the Cloud provider, however, nothing is broken. Eventual consistency is inherent to the nature of Cloud computing, a principle we call the CAP Theorem: no distributed computing system can guarantee (immediate) consistency, availability, and partition tolerance at the same time. You can get any two of these, but not all three at once.
Of these three characteristics, partition tolerance is the least familiar. In essence, a distributed system is partition tolerant when it will continue working even in the case of a partial network failure. In other words, bits and pieces of the system can fail or otherwise stop communicating with the other bits and pieces, and the overall system will continue to function.
With on-premise distributed computing, we’re not particularly interested in partition tolerance: transactional environments run in a single partition. If we want ACID transactionality (atomic, consistent, isolated, and durable transactions), then we should stick with a partition intolerant approach like a two-phase commit infrastructure. In essence, ACID implies that a transaction runs in a single partition.
But in the Cloud, we require partition tolerance, because the Cloud provider is willing to allow that each physical instance cannot necessarily communicate with every other physical instance at all times, and each physical instance may go down unpredictably. And if your underlying physical instances aren’t communicating or working properly, then you have either an availability or a consistency issue. But since the Cloud is architected for high availability, consistency will necessarily suffer.
The Solution: Rethink Your Priorities
The kneejerk reaction might be that since consistency is nonnegotiable, we need to force the Cloud providers to give up partition tolerance. But in reality, that’s entirely the wrong way to think about the problem. Instead, we must rethink our priorities.
As any data specialist will tell you, there are always performance vs. flexibility tradeoffs in the world of data. Every generation of technology suffers from this tradeoff, and the Cloud is no different. What is different about the Cloud is that we want virtualization-based elasticity—which requires partition tolerance.
If we want ACID transactionality then we should stick with an on-premise partition intolerant approach. But in the Cloud, ACID is the wrong priority. We need a different way of thinking about consistency and reliability. Instead of ACID, we need BASE (catchy, eh?)
BASE stands for Basic Availability (supports partial failures without leading to a total system failure), Soft-state (any change in state must be maintained through periodic refreshment), and Eventual consistency (the data will be consistent after a set amount of time passes since an update). BASE has been around for several years and actually predates the notion of Cloud computing; in fact, it underlies the telco world’s notion of “best effort” reliability that applies to the mobile phone infrastructure. But today, understanding the principles of BASE is essential to understanding how to architect applications for the Cloud.
Thinking in a BASE Way
Let’s put the BASE principles in simple terms.
Basic availability: stuff happens. We’re using commodity hardware in the Cloud. We’re expecting and planning for failure. But hey, we’ve got it covered.
Soft state: the squeaky wheel gets the grease. If you don’t keep telling me where you are or what you’re doing, I’ll assume you’re not there anymore or you’re done doing whatever it is you were doing. So if any part of the infrastructure crashes and reboots, it can bootstrap itself without any worries about it being in the wrong state.
Eventual consistency: It’s OK to use stale data some of the time. It’ll all come clean eventually. Accountants have followed this principle since Babylonian times. It’s called “closing the books.”
So, how would you address your inventory app following BASE best effort principles? First, assume that any product quantity is approximate. If the quantity isn’t near zero you don’t have much of a problem. If it is near zero, set the proper expectation with the customer. Don’t charge their credit card in a synchronous fashion. Instead, let them know that their purchase has probably completed successfully. Once the dust settles, let them know if they got the item or not.
Of course, this inventory example is an oversimplification, and every situation is different. The bottom line is that you can’t expect the same kind of transactionality in the Cloud as you could in a partition intolerant on-premise environment. If you erroneously assume that you can move your app to the Cloud without reworking how it handles transactionality, then you are in for an unpleasant surprise. On the other hand, rearchitecting your app for the Cloud will improve it overall.
The ZapThink Take
Intermittently stale data? Unpredictable counts? States that expire? Your computer science profs must be rolling around in their graves. That’s no way to write a computer program! Data are data, counts are counts, and states are states! How could anything work properly if we get all loosey-goosey about such basics?
Welcome to the twenty-first century, folks. Bank account balances, search engine results, instant messaging buddy lists—if you think about it, all of these everyday elements of our wired lives follow BASE principles in one way or another.
And now we have Cloud computing, where we’re bundling together several different modern distributed computing trends into one neat package. But if we mistake the Cloud for being nothing more than a collection of existing trends then we’re likely to fall into the “horseless carriage” trap, where we fail to recognize what’s special about the Cloud.
The Cloud is much more than a virtual server in the sky. You can’t simply migrate an existing app into the Cloud and expect it to work properly, let alone take advantage of the power of the Cloud. Instead, application migration and application modernization necessarily go hand in hand, and architecting your app for the Cloud is more important than ever.