DynamoDB is Not a Database

Daniel Vassallo
Jan 6 · 5 min read

Amazon describes DynamoDB as a database, but it’s best seen as a highly-durable data structure in the cloud. A partitioned B-tree data structure, to be precise.

DynamoDB is much more similar to a Redis than it is to a MySQL. But, unlike Redis, it is immediately consistent and highly-durable, centered around that single data structure. If you put something into DynamoDB, you’ll be able to read it back immediately and, for all practical purposes, you can assume that what you have put will never get lost.

It is true that DynamoDB can replace a relational database. Sometimes, this simple data structure happens to be enough. But you can only get away with DynamoDB as your database if you can afford to have all your data in a primitive B-tree.

Compared to a relational database, DynamoDB requires you to do most of the data querying yourself within your application. You can either read a single value out of DynamoDB, or you can get a contiguous range of data. But if you want to aggregate, filter, or sort, you have to do that yourself, after you receive the requested data range.

Having to do most query processing on the application side isn’t just inconvenient. It also comes with performance implications. Relational databases run their queries close to the data, so if you’re trying to calculate the sum total value of orders per customer, then that rollup gets done while reading the data, and only the final summary (one row per customer) gets sent over the network. However, if you were to do this with DynamoDB, you’d have to get all the customer orders (one row per order), which involves a lot more data over the network, and then you have to do the rollup in your application, which is far away from the data. This characteristic will be one of the most important aspects of determining whether DynamoDB is a viable choice for your needs.

Another factor to consider is cost. Storing 1 TB in DynamoDB costs $256/month. For comparison, storing 1 TB in S3 costs $23.55/month. Data can also be compressed much more efficiently in S3, which could make this difference even bigger. However, storage cost is rarely a large factor when deciding whether DynamoDB is a viable option. Instead, it’s generally request pricing that matters most.

By default, you should probably start with DynamoDB’s on-demand pricing and only consider provisioned capacity as a cost optimization. On-demand costs $1.25 per million writes, and $0.25 per million reads. Now, since DynamoDB is such a simple data structure, it’s often not that hard to estimate how many requests you will need. You will likely be able to inspect your application and map every logical operation to a number of DynamoDB requests. For example, you might find that serving a web page will require four DynamoDB read requests. Therefore, if you expect to serve a million pages per day, your DynamoDB requests for that action would cost $1/day.

Then, if the performance characteristics of DynamoDB are compatible with your application and the on-demand request pricing is in the ballpark of acceptability, you can consider switching to provisioned capacity. On paper, that same workload that cost $1/day to serve 1 million pages would only cost $0.14/day with provisioned capacity, which seems like a very spectacular cost reduction. However, this calculation assumes both that requests are evenly distributed over the course of the day and that there is absolutely zero capacity headroom. (You would get throttled if there were a million and one requests in a day.) Obviously, both of these assumptions are impractical. In reality, you’re going to have to provision abundant headroom in order to deal with the peak request rate, as well as to handle any general uncertainty in demand. With provisioned capacity, you will have the burden to monitor your utilization and proactively provision the necessary capacity.

In general, you will almost always want to start with on-demand pricing (no capacity management burden). Then, if your usage grows significantly, you will almost always want to consider moving to provisioned capacity (significant cost savings). However, if you believe that on-demand pricing is too expensive, then DynamoDB will very likely be too expensive, even with provisioned capacity. In that case, you may want to consider a relational database, which will have very different cost characteristics than DynamoDB.

It is important to note that, with on-demand pricing, the capacity you get is not perfectly on-demand. Behind the scenes, DynamoDB adjusts a limit on the number of reads and writes per second, and these limits change based on your usage. However, this is an opaque process and, if you want to ensure that you reserve capacity for big fluctuations in usage, you may want to consider using provisioned capacity for peace of mind.

A final word about DynamoDB indexes. They come in two flavors: local and global. Local indexes came first in early 2013, and global indexes were added just a few months later. The only advantage of local indexes is that they’re immediately consistent, but they do come with a very insidious downside. Once you create a local index on a table, the property that allows the table to keep growing indefinitely goes away. Local indexes come with the constraint that all the records that share the same partition key need to fit in 10 GB, and once that allocation gets exhausted, all writes with that partition key will start failing. Unless you know for sure that you won’t ever exceed this limit, you should avoid local indexes.

On the other hand, global indexes don’t constrain your table size in any way, but reading from them is eventually consistent (although the delay is almost always unnoticeable). Global indexes also have one insidious downside, but for most scenarios it is much less worrisome than that of local indexes. DynamoDB has an internal queue-like system between the main table and the global index, and this queue has a fixed (but opaque) size. Therefore, if the provisioned throughput of a global index happens to be insufficient to keep up with updates on the main table, then that queue can get full. When that happens, disaster strikes: all write operations on the main table start failing. The most problematic part of this behavior is that there’s no way to monitor the state of this internal queue. So, the only way to prevent it is to monitor the throttled request count on all your global indexes, and then to react quickly to any throttling by provisioning additional capacity on the affected indexes. Nevertheless, this situation tends to only happen with highly active tables, and short bursts of throttling rarely cause this problem.

Global indexes are still very useful, but keep in mind the fact that they’re eventually consistent and that they can indirectly affect the main table in a very consequential manner if they happen to be underprovisioned.


This is an excerpt from The Good Parts of AWS. If you liked this perspective on DynamoDB, you may also like what we have to say on other AWS services.

Daniel Vassallo

Written by

Being independent. https://twitter.com/dvassallo.

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade