Among the many options available for running managed Postgres, Amazon Aurora Serverless initially stood out as unique when it was announced. It promised to introduce the concepts of scaling to zero, autoscaling, and usage-based pricing to Postgres.

A lot has changed since then, including AWS's decision to deprecate scale-to-zero in Aurora. Today, developers have other options for running serverless Postgres, such as Neon. In this comparison, we'll examine the key differences between Aurora and Neon, focusing on their serverless capabilities and pricing models.

Navigating the Amazon Aurora universe

Let’s start by clarifying terminology. When developers refer to “Amazon Aurora”, they might be referring to three different products:

• Amazon Aurora provisioned is the “traditional” version of Amazon Aurora, where you provision database instances with a fixed capacity. You have to specify the instance size upfront, and you are billed based on the allocated resources regardless of usage.
• Amazon Aurora Serverless v1 came next as the first serverless version of Amazon Aurora. The two core functionalities it introduced were scale to zero and autoscaling: Aurora Serverless v1 instances automatically start up, shut down, and scale capacity up or down based on your application's needs. It's positioned as a more optimal choice for applications with intermittent, unpredictable, or variable workloads.
• Amazon Aurora Serverless v2 aimed to address the limitations of v1. It claimed to offer more fine-grained scaling, improved performance, and the same high availability and durability as the provisioned instances. But these improvements came at a high price: *losing scale to zero. *

In Aurora Serverless v1, if there were no database connections or activity for a period of time, the database could automatically pause and reduce capacity to zero, effectively eliminating costs during idle periods. This capability was essential to the claim of serverless as a way to reduce costs for applications with “infrequent usage patterns” running in Aurora.

In contrast, Aurora Serverless v2 maintains a minimum capacity of 0.5 ACUs (Aurora Capacity Units) even when there is no database activity. As we’ll explore later in the post, this means that there are always some costs incurred, regardless of usage. This approach was taken to ensure instant provisioning and to eliminate the latency associated with cold starts, but it came with a trade-off in costs for users.

Now, meet Neon

Neon's architecture is inspired by Amazon Aurora and its separation of compute and storage. But Neon takes this design one step further by adopting a custom-built storage engine that keeps a history of Postgres transactions. This enables Neon not only to offer a truly serverless experience with scale to zero, but also to focus on improving development workflows by offering features like database branching.

Neon automatically scales compute instances to zero when inactive for a specified period (5 minutes by default). Similar to Aurora Serverless, it includes autoscaling to dynamically adjust compute resources based on the current load within user-defined limits. Differently to Aurora, Neon comes with a free tier.

Features: Neon vs Aurora Serverless v2

Let’s dig deeper into how Neon compares to Amazon Aurora Serverless in terms of features. This table gives you the high-level view:

*If you’re wondering what the heck is an ACU, see the next section.

Pricing model: Aurora Serverless

When it’s time to evaluate pricing for Aurora Serverless, you’ll very quickly be confronted with what seems like an easy question to answer: what is an ACU?

Let’s backtrack. In Aurora Serverless, ACUs (Aurora Capacity Units) are the units of measure used to define the capacity of database instances. When running an instance as a user, you’ll define a minimum and maximum ACU limit. Aurora will scale up and down automatically between these minimum and maximum limits, in 0.5 ACU increments.

The minimum number of ACUs varies between Aurora Serverless v1 and v2:

• In Aurora Serverless v1, you can define 0 as the minimum limit (v1 scales to zero).
• In Aurora Serverless v2, the minimum possible limit is 0.5 ACU. We’ll break down what this implies cost-wise in the next section.

Now, to size an Aurora Serverless instance, the first thing you would like to know is how much CPU and memory is contained in each ACU. You will be billed according to how many ACUs you have used in a month, so this is highly relevant: for example, you may suspect that 1 CPU would be enough to handle your peak load, and therefore you would like to set up your maximum ACU limit at a corresponding capacity.

Unfortunately, AWS is opaque in disclosing this information, making pricing in Aurora very uncertain. According to the Aurora docs, an ACU is “a combination of approximately 2 gibibytes (GiB) of memory, corresponding CPU, and networking”.

We’re as confused as you are about what this means. Some folks online have experimented with this and concluded that probably 1 ACU = 0.25 vCPU, 2 GiB memory. But we can’t know for sure.

ACU mysteries aside, your monthly Aurora Serverless bill will be calculated as the sum of a few elements, included below. If you avoid the I/O charges by using I/O optimized storage (highly recommended), compute and database storage will most likely be the main elements contributing to your monthly costs.

Since Aurora Serverless v2 doesn’t have a free tier and has minimum requirements for both compute and storage, we can estimate the minimum costs for the smallest database possible running 24/7:

• Minimum monthly cost: $65.65 USD.

This is the absolute monthly minimum for a database running in us-east-1, assuming we’re using I/O optimized storage to avoid extra I/O charges.

This calculation assumes that you’re using 0.5 ACU (the minimum) at all times. However, an important practical consideration is that in reality, you’ll be forced to pick an ACU range, and the lowest maximum ACU possible is 1 ACU. So, a better expectation is that, in the previous example, costs would oscillate between $60 and $120 USD approximately.

Pricing model: Neon

Compute in Neon with autoscaling enabled works similarly to Aurora Serverless, but without the opacity. In Neon, 1 CU = 1 vCPU and 4 GiB of memory. You’ll be able to set up minimum and maximum autoscaling limits (with the minimum being able to scale to zero if you wish), and your compute consumption will be billed in CU-hours at the end of the month.

In terms of billing components, Neon offers three different pricing plans. Your monthly bill will account for:

• The monthly fee corresponding to each plan ($0, $19, or $69)
• Any additional compute or storage usage over what is included within each plan
• Charges per additional projects (the logical equivalent of an instance in Neon)

Comparing compute costs: Neon vs Aurora Serverless v2

Estimating compute costs is often the hardest piece with serverless databases. To bring some clarity to this, let’s work through some example workloads that teams might see for serverless applications.

For this exercise, we’ll use the following equivalence:

• 1 ACU in Aurora = 0.5 CU in Neon = 0.5 vCPU, 2 GiB memory.

Taking into consideration that AWS discloses that 1 ACU equals 2 GiB of memory, and that 1 CU in Neon equals 4 GiB of memory, this equivalence seems like a fair assumption—but note that this is approximate. We have reasons to believe that ACUs are even smaller than that CPU-wise, so your** real workload may require higher ACU limits** (and therefore higher costs) than estimated here.

Low compute, testing workloads

Imagine a small team working on a new feature. They need multiple dev, staging, and testing environments, but each environment has minimal traffic and data storage needs. These databases are often idle for extended periods and only need to be active during specific testing windows.

Using database branches, we could do this on the Neon free tier. But, if this workload requires multiple projects, we cand use the Launch tier instead. Let’s say we’re using 1 vCPU (1 CU) for each of our three projects (dev, staging, and testing), but overall, they are idle 80% of the time. So, this becomes:

• 1 CU * 3 Projects * 730 * 0.2 = 438 CU-hours

This is over the 300 compute hours included in the Launch tier, so we’ll also have to pay $0.16 per extra compute hour:

• $19 + (138 compute hours * $0.16) = $41.08

So, just over $40 monthly for this testing workload with Neon. For Aurora Serverless v2, it is:

• 2 ACU * 3 instances * 730 = 4,380 ACU-hours

This costs us many more compute hours because we have no scale to zero. Now, using the standard configuration pricing of $0.16 per ACU-hour in I/O optimized instances:

• 4,380 ACU-hours * $0.16 = $700.8

The Aurora compute cost would be 17x more than the Neon cost for this sceario, mainly due to Neon’s ability to scale to zero.

Medium compute, analytics workloads

Here, the team might need to batch-run analytics queries and generate reports to gain insights into user behavior and application performance.

Let’s do Neon first. Let’s assume we’re still in the Launch tier, that we’ll use 2 vCPUs (CUs), and 1 project. Again, these analytics runs aren’t constant—we’ll assume that we’re using them 50% of the time and that they’re idle the rest. With Neon, this looks like:

• 2 CUs * 1 Project * 730 * 0.5 = 730 CU-hours

If we stayed on the Launch tier, we’d have to pay 430 extra compute hours, so the extra cost would be:

• $19 + (430 CU-hours * $0.16) = $448.8

*We can get this cost much lower if we upgraded to the Scale tier, which includes 750 CU-hours within the $69 monthly fee. *

With Aurora Serverless v2, we get:

• 4 ACUs * 1 Instance * 730 = 2,920 ACU-hours

Again, if we assume a I/O-optimized storage so we’re not abused by I/O costs, the monthly price would be:

• 2,920 ACU-hours* $0.16 = $467.2

High compute, application workloads

A high compute, application workload will be a production environment with significant traffic and low latency.

Here, we’ll use a variable workload, with 8 vCPUs used during working hours (180 hours / month) and 2 /vCPUs during off-peak hours (550 hours / month). We’ll assume 5 instances/projects. In this scenario, there is no idle time.

With the Scale tier on Neon, this works out as:

• (8 CUs *180) + (2 CUs * 550) * 5 Projects = 6,940 CU-hours

We have 750 CU-hours included in the Scale tier, so the cost for this would be:

• $69 + (6,190 * $0.16) = $1,059.4

In Aurora Serverless:

• (16 ACUs * 180) + (4 ACUs * 550) * 5 Instances = 25,400 ACU-hours

And the monthly price:

• 25,400 ACU-hours * $0.16 = $4,064

Conclusion

Both Amazon Aurora and Neon offer serverless options to run managed Postgres instances. While Aurora provides robust scalability and a rich set of features, Neon stands out with some advantages, mainly the capacity to scale to zero and a simpler and more transparent pricing structure with a free tier. This makes it a more attractive choice for startups and mid-sized businesses.