dbt Deployment Options

dbt is wildly popular and has become a fundamental part of many data stacks. While it’s easy to spin up a project and get things running on a local machine, taking the next step and deploying dbt to production isn’t quite as simple.  

In this article we will discuss options for deploying dbt to production, comparing some high, medium, and low effort options so that you can find which works best for your business and team. You might be deploying dbt using one of these patterns already; if you are, hopefully this guide will help highlight some improvements you can make to your existing deployment process.  

We're going to assume you know how to run dbt on your own computer (aka your local dbt setup).  We’re also going to assume that you either want to or need to run dbt in a “production” environment – a place where other tools and systems make use of the data models that dbt creates in your warehouse.

Enhancing understanding of dbt deployment

The deployment process for dbt jobs extends beyond basic scheduling and involves a multifaceted approach. This includes establishing various dbt environments with distinct roles and requirements, ensuring the reliability and scalability of these environments, integrating dbt with other tools in the (EL)T stack, and implementing effective scheduling strategies for dbt tasks. By focusing on these aspects, a comprehensive and robust dbt deployment strategy can be developed. This strategy will not only address current data processing needs but also adapt to future challenges and changes in your data landscape, ensuring long-term success and reliability.

dbt environments

In deploying dbt you have the creation and management of certain dbt environments. The development environment is the initial testing ground for creating and refining dbt models. It allows for experimentation without impacting production data. Following this, the testing environment, including stages like UAT and regression testing, rigorously evaluates the models for accuracy and performance. Finally, the production environment is where these models are executed on actual data, demanding high stability and performance.

Data reliability and scalability

Reliability and scalability of data models are also important. Ensuring that the data models produce accurate and consistent results is essential for maintaining trust in your data. As your data grows, the dbt deployment should be capable of scaling, handling increased volumes, and maintaining performance.

End-to-End data pipeline integration

Integration with other data tools and systems is another key aspect. A seamless integration of dbt with EL tools, data visualization platforms, and data warehouses ensures efficient data flow and processing, making dbt a harmonious component of your broader data stack.

dbt job scheduling

Effective dbt scheduling goes beyond mere time-based scheduling. It involves context-aware execution, such as triggering jobs based on data availability or other external events. Managing dependencies within your data models is critical to ensure that transformations occur in the correct sequence. Additionally, adapting to varying data loads is necessary to scale resources effectively and maintain the efficiency of dbt job executions.

The main flavors for dbt deployment are:  

• Cron Job
• Cloud Runner Service (like dbt Cloud)
• Integrated Platform (like Datacoves)
• Fully Custom (like Airflow, Astronomer, etc)

They each have their place, and the trade-offs between setup costs and long-term maintainability is important to consider when you’re choosing one versus another.  

We can compare these dbt deployment options across the following criteria:

• Ease of Use / Implementation
• Required Technical Ability
• Configurability
• Customization
• Best for End-to End Deployment

Cron job

Cron jobs are scripts that run at a set schedule. They can be defined in any language. For instance, we can use a simple bash script to run dbt. It’s just like running the CLI commands, but instead of you running them by hand, a computer process would do it for you.  

Here’s a simple cron script:

In order to run on schedule, you’ll need to add this file to your system’s crontab.

As you can tell, this is a very basic dbt run script; we are doing the bare minimum to run the project. There is no consideration for tagged models, test, alerting, or more advanced checks.  

Even though Cron jobs are the most basic way to deploy dbt there is still a learning curve. It requires some technical skills to set up this deployment. Additionally, because of its simplicity, it is pretty limited. If you are thinking of using crons for multi-step deployments, you might want to look elsewhere.  

While it's relatively easy to set up a cron job to run on your laptop this defeats the purpose of using a cron altogether. Crons will only run when the daemon is running, so unless you plan on never turning off your laptop, you’ll want to set up the cron on an EC2 instance (or another server). Now you have infrastructure to support and added complexity to keep in mind when making changes. Running a cron on an EC2 instance is certainly doable, but likely not the best use of resources. Just because it can be done does not mean it should be done. At this point, you’re better off using a different deployment method.

The biggest downside, however, is that your cron script must handle any edge cases or errors gracefully.  If it doesn’t, you might wind up with silent failures – a data engineer’s worst enemy.  

Who should use Cron for dbt deployment?

Cron jobs might serve you well if you have some running servers you can use, have a strong handle on the types of problems your dbt runs and cron executions might run into, and you can get away with a simple deployment with limited dbt steps. It is also a solid choice if you are running a small side-project where missed deployments are probably not a big deal.  

Use crons for anything more complex, and you might be setting yourself up for future headaches.  

Ease of Use / Implementation – You need to know what you’re doing

Required Technical Ability – Medium/ High

Configurability – High, but with the added complexity of managing more complex code

Customization – High, but with a lot of overhead. Best to keep things very simple  

Best for End-to-End Deployment - Low.

Cloud Service Runners (like dbt Cloud)

Cloud Service Runners like dbt Cloud are probably the most obvious way to deploy your dbt project without writing code for those deployments, but they are not perfect.  

dbt Cloud is a product from dbt Labs, the creators of dbt. The platform has some out-of-the-box integrations, such as Github Actions and Webhooks, but anything more will have to be managed by your team. While there is an IDE (Integrated Developer Experience) that allows the user to write new dbt models, you are adding a layer of complexity by orchestrating your deployments in another tool. If you are only orchestrating dbt runs, dbt Cloud is a reasonable choice – it's designed for just that.  

However, when you want to orchestrate more than just your dbt runs – for instance, kickoff multiple Extract / Load (EL) processes or trigger other jobs after dbt completes – you will need to look elsewhere.  

dbt Cloud will host your project documentation and provide access to its APIs. But that is the lion’s share of the offering. Unless you spring for the Enterprise Tier, you will not be able to do custom deployments or trigger dbt runs based on incoming data with ease.  

Deploying your dbt project with dbt Cloud is straightforward, though. And that is its best feature. All deployment commands use native dbt command line syntax, and you can create various "Jobs" through their UI to run specific models at different cadences.

Who should use dbt Cloud for dbt deployment?

If you are a data team with data pipelines that are not too complex and you are looking to handle dbt deployments without the need for standing up infrastructure or stringing together advanced deployment logic, then dbt Cloud will work for you. If you are interested in more complex triggers to kickoff your dbt runs - for instance, triggering a run immediately after your data is loaded – there are other options which natively support patterns like that. The most important factor is the complexity of the pieces you need to coordinate, not necessarily the size of your team or organization.  

Overall, it is a great choice if you’re okay working within its limitations and support a simple workflow. As soon as you reach any scale, however, the cost may be too high.  

Ease of Use / Implementation – Very easy

Required Technical Ability – Low

Configurability – Low / Medium

Customization – Low

Best for End-to-End Deployment - Low

Integrated platform (like Datacoves)

The Modern Data Stack is a composite of tools. Unfortunately, many of those tools are disconnected because they specialize in handling one of the steps in the ELT process. Only after working with them do you realize that there are implicit dependencies between these tools. Tools like Datacoves bridge the gaps between the tools in the Modern Data Stack and enable some more flexible dbt deployment patterns. Additionally, they cover the End-to-End solution, from Extraction to Visualization, meaning it can handle steps before and after Transformation.

Coordinating dbt runs with EL processes

If you are loading your data into Snowflake with Fivetran or Airbyte, your dbt runs need to be coordinated with those EL processes. Often, this is done by manually setting the ETL schedule and then defining your dbt run schedule to coincide with your ETL completion. It is not a hard dependency, though. If you’re processing a spike in data or running a historical re-sync, your ETL pipeline might take significantly longer than usual. Your normal dbt run won’t play nicely with this extended ETL process, and you’ll wind up using Snowflake credits for nothing.

This is a common issue for companies moving from early stage / MVP data warehouses into more advanced patterns. There are ways to connect your EL processes and dbt deployments with code, but Datacoves makes it much easier. Datacoves will trigger the right dbt job immediately after the load is complete. No need to engineer a solution yourself. The value of the Modern Data Stack is being able to mix and match tools that are fit for purpose.

Seamless integration and orchestration

Meeting established data freshness and quality SLAs is challenging enough, but with Datacoves, you’re able to skip building custom solutions for these problems. Every piece of your data stack is integrated and working together. If you are orchestrating with Airflow, then you’re likely running a Docker container which may or may not have added dependencies. That’s one common challenge teams managing their own instances of Airflow will meet, but with Datacoves, container / image management and synchronization between EL and dbt executions are all handled on the platform. The setup and maintenance of the scalable Kubernetes infrastructure necessary to run Airflow is handled entirely by the Datacoves platform, which gives you flexibility but with a lower learning curve. And, it goes without saying that this works across multiple environments like development, UAT, and production.

Streamlining the Datacoves experience

With the End-to-End Pipeline in mind, one of the convenient features is that Datacoves provides a singular place to access all the tools within your normal analytics workflow - extraction + load, transformation, orchestration, and security controls are in a single place. The implicit dependencies are now codified; it is clear how a change to your dbt model will flow through to the various pieces downstream.  

Who should use Datacoves for dbt deployment?

Datacoves is for teams who want to introduce a mature analytics workflow without the weight of adopting and integrating a new suite of tools on their own. This might mean you are a small team at a young company, or an established analytics team at an enterprise looking to simplify and reduce platform complexity and costs.

There are some prerequisites, though. To make use of Datacoves, you do need to write some code, but you’ll likely already be used to writing configuration files and dbt models that Datacoves expects. You won't be starting from scratch because best practices, accelerators, and expertise are already provided.

Ease of Use / Implementation – You can utilize YAML to generate DAGS for a simpler approach, but you also have the option to use Python DAGS for added flexibility and complexity in your pipelines.

Required Technical Ability – Medium

Configurability – High

Customization – High. Datacoves is modular, allowing you to embed the tools you already use

Best for End-to-End Deployment - High. Datacoves takes into account all of the factors of dbt Deployment

Fully custom (like Airflow, Dagster, etc)

What do you use to deploy your dbt project when you have a large, complex set of models and dependencies? An orchestrator like Airflow is a popular choice, with many companies opting to use managed deployments through services such as Astronomer.  

For many companies – especially in the enterprise – this is familiar territory. Adoption of these orchestrators is widespread. The tools are stable, but they are not without some downsides.  

These orchestrators require a lot of setup and maintenance. If you’re not using a managed service, you’ll need to deploy the orchestrator yourself, and handle the upkeep of the infrastructure running your orchestrator, not to mention manage the code your orchestrator is executing. It’s no small feat, and a large part of the reason that many large engineering groups have dedicated data engineering and infrastructure teams.  

Running your dbt deployment through Airflow or any other orchestrator is the most flexible option you can find, though. The increase in flexibility means more overhead in terms of setting up the systems you need to run and maintain this architecture. You might need to get DevOps involved, you’ll need to move your dbt project into a Docker image, you’ll want an airtight CI/CD process, and ultimately have well defined SLAs. This typically requires Docker images, container management, and some DevOps work. There can be a steep learning curve, especially if you’re unfamiliar with what’s needed to take an Airflow instance to a stable production release.

There are 3 ways to run Airflow, specifically – deploying on your own, using a managed service, or using an integrated platform like Datacoves. When using a managed service or an integrated platform like Datacoves, you need to consider a few factors:

1. Airflow use cases

2. Ownership and contributing teams

3. Integrations with rest of your stack

‍

Airflow use cases

Airflow is a multi-purpose tool. It’s not just for dbt deployments. Many organizations run complex data engineering pipelines with Airflow, and by design, it is flexible. If your use of Airflow extends well beyond dbt deployments or ELT jobs oriented around your data warehouse, you may be better suited for a dedicated managed service.

Ownership and contributing teams

Similarly, if your organization has numerous teams dedicated to designing, building and maintaining  your data infrastructure, you may want to use a dedicated Airflow solution. However, not every organization is able to stand up platform engineering teams or DevOps squads dedicated to the data infrastructure. Regardless of the size of your team, you will need to make sure that your data infrastructure needs do not outmatch your team’s ability to support and maintain that infrastructure.

Integrations with the rest of your stack

Every part of the Modern Data Stack relies on other tools performing their jobs; data pipelines, transformations, data models, BI tools - they are all connected. Using Airflow for your dbt deployment adds another link in the dependency chain. Coordinating dbt deployments via Airflow can always be done through writing additional code, but this is an additional overhead you will need to design, implement, and maintain. With this approach, you begin to require strong software engineering and design principles. Your data models are only as useful as your data is fresh; meeting your required SLAs will require significant cross-tool integration and customization.  

Who should use a fully custom setup for dbt deployment?

If you are a small team looking to deploy dbt, there are likely better options. If you are a growing team, there are certainly simpler options with less infrastructure overhead. For Data teams with complex data workflows that combine multiple tools and transformation technologies such as Python, Scala, and dbt, however, Airflow and other orchestrators can be a good choice.  

Ease of Use / Implementation – Can be quite challenging starting from scratch

Required Technical Ability – High

Configurability – High

Customization – High, but build time and maintenance costs can be prohibitive ‍

Best for End-to-End Deployment - High, but requires a lot of resources to set up and maintain

dbt Deployment Options Overview

Final thoughts

The way you should deploy your dbt project depends on a handful of factors – how much time you’re willing to invest up front, your level of technical expertise, as well as how much configuration and customization you need.  

Small teams might have high technical acumen but not enough capacity to manage a deployment on their own. Enterprise teams might have enough resources but maintain disparate, interdependent projects for analytics. Thankfully, there are several options to move your project beyond your local and into a production environment with ease. And while specific tools like Airflow have their own pros and cons, it’s becoming increasingly important to evaluate your data stack vendor solution holistically. Ultimately, there are many ways to deploy dbt to production, and the decision comes down to spending time building a robust deployment pipeline or spending more time focusing on analytics.