Building Robust Systems with ACID and Constraints

In 1983, Andreas Reuter and Theo Härder coined the acronym ACID as shorthand for atomicity, consistency, isolation, and durability. They were building on earlier work by Jim Gray who’d proposed atomicity, consistency, and durability, but had initially left out the I. ACID is one of those inventions of the 80s that’s not only just still in use in the form of major database systems like Postgres, Oracle, and MSSQL, but which has never been displaced by a better idea.

In the last decade we’ve seen the emergence of a number of new flavors of data store that come with untraditional features like streaming changesets, JavaScript APIs, or nestable JSON documents. Most of them assume that the need for horizontal partitioning is a given in this day and age and therefore ACID is put to the altar (this doesn’t necessarily have to be the case, see below). Every decision comes with trade offs, but trading away these powerful guarantees for the novelties du jour or an unexamined assumption that horizontal scaling will very soon be a critically required feature is as raw of a deal as you’ll ever see in the technical world.

But why the mismatch in values? I think it’s because many of us have taught ourselves programming on frameworks like Rails, or been educated in environments where ACID databases were a part of the package, and we’ve taken them for granted. They’ve always been there, and we’ve never necessarily considered closely exactly what they can do for us and why they’re important. In many cases this also leads to their most powerful features being underutilized.

I want to convince you that ACID databases are one of the most important tools in existence for ensuring maintainability and data correctness in big production systems. Lets start by digging into each of their namesake guarantees.

The “A” in ACID. It dictates that within a given database transaction, the changes to be committed will be all or nothing. If the transaction fails partway through, the initial database state is left unchanged.

Software is buggy by nature and introducing problems that unintentionally fail some operations is inevitable. Any sufficiently large program is eventually going to want to have an operation that writes two or more objects consecutively, and by wrapping that operation in a transaction, we get to ensure that even in the these worst case scenarios state is left undamaged. Every subsequent operation will start with safe initial state.

It’s never desirable to fail transactions that we hoped to commit, but atomicity cancels the expensive fallout.

Some requests. Each wraps its database operations using an atomic transaction so that they either all commit, or none of them do.
Some requests. Each wraps its database operations using an atomic transaction so that they either all commit, or none of them do.

Many products will claim “document-level atomicity” (e.g. MongoDB, RethinkDB, CouchBase, …) which means that writing any one row is atomic, but nothing beyond that. What happens in a world like this where any failed operation leaves invalid state behind?

The default will be that a subsequent retry won’t be able to reconcile the broken state, and that the data will need to be repaired before it’s usable again.

Here’s an example of a simple GitHub-like service. When a user opens a pull request, we have a number of objects that we have to save in succession before finishing the request: a pull request modeling the created resource, a webhook to fire off to any listeners on the repository, a reviewer record mapping to whomever we’ve assigned review, and an event to store in the audit log.

Demonstration of how without an atomicity guarantee, a failed request results in an invalid state of data.
Demonstration of how without an atomicity guarantee, a failed request results in an invalid state of data.

A request that fails after the first two saves fails to create a valid set of objects, but with transactional atomicity can’t revert the changes it did make. The result? An invalid pull request. A subsequent request that tries to look it up might error as the code tries to load state that was only partially created.

You might hope that projects in this position would have automated protections in place to try and roll back bad partial transactions. While this may exist somewhere, it’s much more likely that the overarching strategy is an optimistic sense of hope that these kinds of problems won’t happen very often. Code paths begin to mutate to load data defensively so that they handle an exponential number of combinations of bad state that have accumulated in the data store over time.

Bad incidents will necessitate heavy manual intervention by operators, or even a specially crafted “fixer script” to clean up state and get everything back to normal. After a certain size, this sort of thing will be happening frequently, and your engineers will start to spend less time as engineers, and more time as data janitors.

A grid of pillars at the Jewish Museum in Berlin. Real world consistency at its best.
A grid of pillars at the Jewish Museum in Berlin. Real world consistency at its best.

The “C” in ACID. It dictates that every transaction will bring a database from one valid state to another valid state; there’s no potential for anything in between.

It might be a little hard to imagine what this can do for a real world app in practice, but consider one the very common case where a user signs up for a service with their email address We don’t want to two accounts with the same email in the system, so when creating the account we’d use an initial check:

  1. Look for any existing users in the system. If there is one, reject the request.

  2. Create a new record for

Regardless of data store, this will generally work just fine until you have a system with enough traffic to start revealing edge cases. If we have two requests trying to register that are running almost concurrently, then the above check can fail us because both could have validated step one successfully before moving on to create a duplicated record.

Without guaranteed consistency, there's nothing to stop the database from transitioning to an invalid state.
Without guaranteed consistency, there's nothing to stop the database from transitioning to an invalid state.

You can solve this problem on an ACID database in multiple ways:

  1. You could use a strong isolation level like SERIALIZABLE on your transactions, which would guarantee that only one creation would be allowed to commit.

  2. You could put a uniqueness check on the table itself (or on an index) which would prevent a duplicate record from being inserted.

Without ACID, its up to your application code to solve the problem. You could implement some a locking system of sorts to guarantee that only one registration for any given email address can be in flight at once. Realistically, many providers on non-ACID databases will probably elect to just not solve the problem. Maybe later, after it causes painful fall out in production.

The “I” in ACID. It ensures that two simultaneously executing transactions that are operating on the same information don’t conflict with each other. Each one has access to a pristine view of the data (depending on isolation level) even if the other has modified it, and results are reconciled when the transactions are ready to commit. Modern RDMSes have sophisticated multiversion concurrency control systems that make this possible in ways that are correct and efficient.

Isolation Level Dirty Read Nonrepeatable Read Phantom Read Serialization Anomaly
Read uncommitted Allowed Possible Possible Possible
Read committed Not possible Possible Possible Possible
Repeatable read Not possible Not possible Allowed Possible
Serializable Not possible Not possible Not possible Not possible
Transaction isolation levels and the contention phenomena that they allow. See the Postgres docs if you want to learn more.

Concurrent resource access is a problem that every real world web application is going to have to deal with. So without isolation, how do you deal with the problem?

The most common technique is to implement your own pessimistic locking system that constrains access to some set of resources to a single operation, and forces others to block until it’s finished. So for example, if our core model is a set of user accounts that own other resources, we’d lock the whole account when a modification request comes in, and only unlock it again after we’ve finished our work.

Demonstration of pessimistic locking showing 3 requests to the same resource. Each blocks the next in line.
Demonstration of pessimistic locking showing 3 requests to the same resource. Each blocks the next in line.

This approach is all downsides:

  1. It’s slow. Operations waiting on a lock may have to wait for very extended periods for resources to become available. The more concurrent access, the worse it is (which probably means that your large users will suffer the most).

  2. It’s inefficient. Not every blocking operation actually needs to wait on every other operation. Because the models you lock on tend to be broad to reduce the system’s complexity, many operations will block when they didn’t necessarily have to.

  3. It’s a lot of work. A basic locking system isn’t too hard to implement, but if you want to improve its speed or efficiency then you quickly need to move to something more elaborate which gets complicated fast. With an ACID database, you’ll get a fast, efficient, and correct locking system built-in for free.

  4. It’s probably not right. Locks and software are hard. Implementing your own system is going to yield problems; it’s just a question of what magnitude.

The “D” in ACID. It dictates that committed transactions stay committed, even in the event of a crash or power loss. It’s so important that even data stores that don’t support the rest of ACI* tend to get it right. I wrote a separate article about MongoDB’s lengthy road to achieving durability for example.

An often cited features document data stores is that they allow you to bootstrap quickly and get to a prototype because they don’t bog you down with schema design. Rich Hickey has a great talk where he makes a distinction between “simple” and “easy”, with simplicity meaning the opposite of complex, and ease meaning “to be at hand” or “to be approachable” in that it may provide short term gratification, even if it’s to the detriment of long term maintainability. Schemaless databases are not simple; they’re easy.

First of all, the claim around faster prototyping isn’t actually true – an experienced developer reasonably competent with their RDMS of choice and armed with an ORM and migration framework can keep up with their document store-oriented counterpart (and almost certainly outpace them), but even if it were true, it’s optimizing for exactly the wrong thing.

While building a prototype quickly might be important for the first two weeks of a system’s lifespan, the next ten years will be about keeping it running correctly by minimizing bugs and data consistency problems that will lead to user and operator pain and attrition. Life is artificially difficult when your User records aren’t even guaranteed to come with an id or email field, and even steadfast schemaless enthusiasts will acquiesce to allow some form of constraints.

By the time an organization hits hundreds of models and thousands of fields, they’ll certainly be using some kind of object modeling framework in a desperate attempt to get a few assurances around data shape into place. Unfortunately by that point, things are probably already inconsistent enough that it’ll make migrations difficult in perpetuity, and application code twisted and complicated as it’s built to safely handle hundreds of accumulated edge cases.

For services that run in production, the better defined the schema and the more self-consistent the data, the easier life is going to be. Valuing miniscule short-term gains over long-term sustainability is a pathological way of doing anything; when building production-grade software, it’s a sin.

A common criticism of ACID databases is that they don’t scale, and by extension horizontally scalable (and usually non-ACID) data stores are the only valid choice.

First of all, despite unbounded optimism for growth, the vast majority will be well-served by a single vertically scalable node; probably forever. By offloading infrequently needed “junk” to scalable alternate stores and archiving old data, it’s reasonable to expect to vertically scale a service for a very long time, even if it has somewhere on the order of millions of users. Show me any databases that’s on the scale of TBs or larger, and I’ll show you the 100s of GBs that are in there when they don’t need to be.

After reaching scale on the order of Google’s, there’s an argument to be made for giving up aspects of ACID in return for certain kinds of partitioning tolerance and guaranteed availability, but advances in newer database technologies that support some ACID along with scaling mean that you don’t have to go straight to building on top of a glorified key/value store anymore. For example, Citus gives you per-shard ACID guarantees. Google Spanner provides distributed locking read-write transactions for when you need them.

For best results, build your app on solid foundations.
For best results, build your app on solid foundations.

There’s a common theme to everything listed above:

  • You can get away without atomicity, but you end up hacking around it with cleanup scripts and lots of expensive engineer-hours.

  • You can get away without consistency, but only through the use of elaborate application-level schemes.

  • You can get away without isolation, but only by building your own probably slow, probably inefficient, and probably buggy locking scheme.

  • You can get away without constraints and schemas, but only by internalizing a nihilistic understanding that your production data isn’t cohesive.

By choosing a non-ACID data store, you end up reimplementing everything that it does for you in the user space of your application, except worse.

Your database can and should act as a foundational substrate that offers your application profound leverage for fast and correct operation. Not only does it provide these excellent features, but it provides them in a way that’s been battle-tested and empirically vetted by millions of hours of running some of the heaviest workloads in the world.

My usual advice along these lines is that there’s no reason not to start your projects with an RDMS providing ACID and good features around constraints. In almost every case the right answer is probably to just use Postgres.

Did I make a mistake? Please consider sending a pull request.