Skip to main content

MongoDB vs DynamoDB vs Tigris - a NoSQL Database Comparison

ยท 11 min read
Matthew Revell

Choosing how your app stores and manages data is one of those foundational decisions that will impact just about every other step you take. But with so much variety, how do you make the right call?

Tigris is an open source NoSQL database and search platform that you can run in any cloud, on your own hardware, or have it managed for you via Tigris Cloud. Here we'll look at how Tigris compares both to MongoDB and DynamoDB, as well as providing an overview of some of the features and tradeoffs you'll need to consider.

Banner for MongoDB vs DynamoDB vs Tigris - a NoSQL Database Comparison

Just want to see the feature comparison table? No problem, view the NoSQL database comparison table >

Types of NoSQL databaseโ€‹

There are three fundamental things you need to consider when looking at different data stores:

  • NoSQL Data model
  • Scaling
  • Consistency

Each of those impacts the role a database can play in your application architecture. Tigris, MongoDB, and DynamoDB all take a different approach, and understanding where they differ will help you decide which makes most sense for your needs.

Data modelsโ€‹

Let's start with the data model. In a relational DBMS, such as PostgreSQL or Oracle, data falls into the familiar rows and columns of a table, with the option to create relationships across those tables.

With NoSQL databases, on the other hand, there's much more variety in how they store data. The most common non-relational data models are:

  • Document: Both MongoDB and Tigris store data as JSON documents, rather than in tables.
  • Key value: Probably the simplest form of database, a key-value store such as DynamoDB has a one-to-one relationship between the key and an item of data. Typically, the database system has no insight into what is stored: it could be raw text, a binary blob, numeric, JSON, or anything else.
  • Column store: Similar to a relational database, column stores like Cassandra or Google's Bigtable organizes data into families of columns and rows, but rather than adding data as a new row, it is added as a new column.
  • Time series database: Optimized for time-stamped data, such as a series of readings from industrial IoT sensors.
  • Multi-model: A data platform, such as Tigris, or a database like FoundationDB, might have one underlying data model (such as document or key-value) but provide multiple ways of querying that data so as to break out of the confines of one particular data model.


At its simplest, databases can scale up or scale out. Scaling up means adding more RAM, disk space, CPU, and so on to a single machine or VM. That approach is easy, but it has a couple of downsides. The most obvious is that bare metal servers and virtual machines can get only so big. The other is that relying on one machine introduces a single point of failure.

Scaling out, on the other hand, expands capacity by adding more database instances and spreading both the data stored and the query load across them. That usually means storing more than one copy of the same data in different places or storing the data in one place, but it might be that an index that refers to that data lives on another machine. However it plays out, the moment your database spreads across more than one machine, you have to choose between two data consistency scenarios.

Data consistencyโ€‹

In a scale out database, data consistency comes in one of two forms:

  • Strongly consistent: The database ensures that every copy of and reference to an item of data is the same however and wherever it is accessed. As a result, the data might sometimes be unavailable in order to avoid inconsistency.
  • Highly available: All data is always available. Even if another user is updating one instance of a document, for example, you can still access the older version before their write completes. The tradeoff is that different users might see different versions of the same data.

MongoDB, DynamoDB, and Tigris all take somewhat different approaches to data model, scaling, and data consistency. And that impacts how they deliver on features such as ACID transactions, full-text search, and globally consistent secondary indexes. Let's look in detail at some of those key features that differ between MongoDB, DynamoDB, and Tigris.

Interactive ACID transactionsโ€‹

Writing to a database is usually made up of a chain of operations. Let's say you're updating an existing document. The process might look a little like this:

  1. Fetch the existing document.
  2. Find and fetch any indexes that reference that document.
  3. Write the new version of the document.
  4. Share that new version to any replicas.
  5. Update existing references in indexes, add any new ones, and delete those that no longer apply.

Let's say that just after writing the new document, someone puts a backhoe through the data center's power cable, and the back-up generator fails to kick-in on time. Or, more likely, a hard drive fails unexpectedly and takes the server down. Once the server was back online, some NoSQL databases would plough on despite two major problems:

  • There would be two different versions of the same document available to use.
  • The indexes would be inaccurate.

ACID transactions prevent those types of inconsistency. In an ACID transaction, either every step of a change takes place, or none takes place. In our scenario above, on regaining power, the server would either reverse the steps that had taken place or complete those that hadn't.

Strongly consistent global secondary indexesโ€‹

Secondary indexes make it easier and faster to find data within the body of a document. Let's say we have a database of people. We can be sure that each person's email address is unique, so we use that as the key for the documents.

However, if we also want to find people from a certain city, then the database would have to trawl through each document to find those that match. The resultant read could take a substantial amount of time. A more efficient way is to maintain an index of which people are in each city. That way, if we want to find all the people from Berlin, for example, the database can perform a single look-up in the cities index to return all the relevant documents.

The accuracy of those indexes depends on when they get updated. Some databases update indexes in the background, separately from writing the document itself. That way, it's never really clear if the index is accurate or stale. Strongly consistent global secondary indexes, on the other hand, get updated when the document is updated. That way, they're always up to date.

Database branchingโ€‹

Imagine working on a codebase without branching. Any change you committed would immediately be in the mainline. It would stifle most engineering teams and make innovations such as CI/CD nearly impossible.

But that's how most databases work. With some effort, you can create a replica, but it's neither automated nor designed as a standard path. Database branching makes creating new database copies as simple as making a git branch. That way, you can create and destroy database branches as part of an automated development workflow without impacting the main branch.

Secondary indexes are ideal when you know upfront how you'll need to retrieve data. But they're less useful in a few scenarios:

  • A query is rare enough that it's not worth the overhead of maintaining a secondary index. For example, in our user profiles, the system might only occasionally need to retrieve people with particular dietary requirements.
  • Where you can't anticipate the data you'll need. Let's say each user profile also includes a free text resume field. Without a schema for the resume, it could hold any data, so searches could be similarly unpredictable.
  • Where potential results are spread across multiple fields. Retrieving the profiles of every person who has a connection with London, for example, could require diving into multiple free text fields.

Full-text search is more efficient than a brute-force query of every document, because it maintains indexes of its own that point to the locations of specific words and phrases.

Automatic database shardingโ€‹

The moment a database outgrows a single server, the question arises of how to spread the data across each node. There are several common ways to shard, or partition, a database, but as an application developer, you almost certainly don't want to have to get into the detail of how that happens.

Automatic database sharding finds the most efficient way to spread data across multiple servers. Depending on the database in question, it will maintain multiple copies of each record to speed up data access and improve redundancy.

Cloud-native architectureโ€‹

Many of the database systems in use today date back to a world of monoliths running mostly on bare metal servers. That made provisioning and scaling a thing of detailed planning.

A datastore built for cloud-native architecture, on the other hand, separates out each aspect -- such as query, indexing, and persistence -- into complementary services that scale independently of each other. That makes it easier to adjust capacity either up or down as demand changes.

Self-managed Data hostingโ€‹

Where you host a database can decide how much control you have over it. A purely cloud-based service might make it harder to move your data should your needs change. Even if you start out with a cloud-based service, the option to self-host gives you flexibility and avoids lock-in.

Comparing Tigris with MongoDB and DynamoDBโ€‹

When making a choice between Tigris, MongoDB, and DynamoDB you need to take each of the above criteria into account. Here's how each of them compares.

Interactive ACID Transactions
Find out more about the Tigris ACID transactions
Strongly Consistent Global Secondary Indexes
See this MongoDB forum thread
Database Branching
Find out more about the Tigris Database branching
Integrated Full Text Search
Find out more about the Tigris full-text search
Automatic Database Sharding
Cloud-native Architecture
Find out more about the Tigris cloud-native architecture
Self-managed Data hosting

Tigris: the MongoDB and DynamoDB alternativeโ€‹

We built Tigris to give developers the consistency of ACID transactions with the flexibility of NoSQL at a fraction of the cost of alternatives. Compared to MongoDB and DynamoDB, Tigris gives you:

  • Consistent data for mission-critical applications: With Tigris, you can be sure that the data you retrieve is always up to date. That's not the case for all NoSQL databases.
  • Developer-first workflow: With database branching, you can treat your database just like your codebase, whether that's for testing, feature branches, or CI/CD.
  • Effortless scaling: Tigris's cloud-native architecture makes scaling second nature, meaning you can focus on your data while Tigris takes care of the mechanics.
  • Self-host or cloud host: Tigris is fully open source and cloud agnostic, so you can host it wherever you choose. Or leave the heavy lifting to us by hosting in Tigris Cloud. There's no vendor lock-in and no heavy DevOps burden.
  • MongoDB compatibility: Already have an application built for MongoDB? No problem, with our MongoDB compatibility layer, you can port to Tigris with minimal refactoring.
  • Lower total cost of ownership: Tigris costs less to run, even at petabyte scales.

Sign up for a free Tigris Cloud account to see the benefits of our serverlesss NoSQL Database and Search platform.

Have questions? Get in touch.

Stay connected

Subscribe to the Tigris Newsletter: