Dylan originally published this comparison in March 2019 on his personal blog. Since then, Looker has added a lot of functionality to PDTs. For an updated comparison, check out our post PDT to dbt Revisited.
Over the past few years, Looker has become a market-leading BI tool. Its data governance and self-service capabilities have now become paramount to thousands of BI stacks. For many, the Third Wave of BI (TM Frank Bien) is well and truly here.
Persistent derived tables (PDTs) are a big part of Looker's success. They allow businesses to quickly transform their data with simple SQL select statements, abstracting lots of the data engineering that might otherwise be required. They allow you to easily materialize those models as tables in your data warehouse, speeding up the query times for your end users.
They are addictive. Build a few PDTs, define an explore. Push to production. Build a few more. Push. Re-define the calculation behind a dimension. Push. While derived tables are great for a number of use-cases (many of which I come on to later), there can come a point where their use can become problematic.
Models can start getting slow, without the ability to optimise fully. Models can get repetitive, with large portions of code written out over and over. Models can break, without the ability to test their output.
dbt is to PDTs what good sourdough with cultured butter is to white sandwich bread with margarine: broadly the same thing, but one somewhat more palatable than the other.
As Tristan Handy, founder of dbt-building Fishtown Analytics, explains:
dbt is the T in ELT (extract, load, transform). It doesn’t extract or load data, but it’s extremely good at transforming data that’s already loaded into your warehouse. Every model is exactly one SELECT query, and this query defines the resulting data set.
Sounds pretty similar, right? It is. dbt simply allows you to abide by certain principals that PDTs don't. dbt allows you some flexibity that PDTs don't.
In this article, I'm going to cover some of the issues with doing all of the transformative modelling in Looker. I'll describe the pitfalls that may befall you and I'll lay out an argument for why dbt may be a better option for some of your modelling. In the next part of this series, I'll outline the practical steps of actually moving your PDTs into dbt.
A quick note
Before I go any further, I want to remark that I wholeheartedly believe that Looker is a fantastic BI tool. It has been transformative for my work at a number of different businesses. It has powered most of the reporting I have done in the past three years. It is great. I fully intend to work with Looker for years to come. I wholly endorse it.
It excels at data governance. It excels at self-service. It excels at integrating with other tools. It's just not the best tool for certain types of data modelling.
It may also be right for your business's modelling right now. It's a great way to get started with data modelling if you've never done it before. It removes steps of complication that having another tool in the mix adds.
So, with that said, here are the instances where I would recommend a move to dbt and the instances where Looker is going to better serve your needs.
Performance problems and incremental models
Incredibly often, the issue that brings people from PDTs to dbt is the performance of their models. Most of the time, this stems from one of two problems.
The first is an inability to build PDTs incrementally.
Imagine you've recently set up Heap, Segment, Snowplow or some other solution for event tracking. You build a PDT that turns these events into page views. Initially, you've only got a few days of data, then a few weeks, then a few months. Every time your PDT needs to rebuild, it rebuilds the whole data set. Your data from day one hasn't changed, but it has to be included in the data that gets refreshed. Gradually, this causes the model to rebuild more and more slowly, frustrating your end users who just want to know how many times people viewed the product page for your new special flying mattress yesterday.
With dbt, you can build models incrementally, only building or rebuilding the rows that have been created or changed since you last ran the model. This can substantially speed up the build of your models. It will also greatly reduce the load on your database when the models build, allowing other queries to be served more quickly.
The second problem is a tendency to never build intermediary PDTs to speed up model builds.
There is a tendency in Looker to only build views (and therefore PDTs) that are necessary as a component of an explore. Often though, it would be beneficial for the speed of your rebuilds to break one long model into a handful of tables and then reference those in a final query. base_pdt -> intermediary_pdt -> final_pdt can often be quicker than having all the logic in one final larger query. It's not that this can't be done in Looker - it's just that it feels wrong to generate all those views in Looker if they aren't being surfaced to your end users.
The workflow described above is common place in dbt. Ultimately, it creates a separation between the preparation of data and the LookML that instructs Looker how to expose your final modelled output.
Testing your models
Another feature that dbt brings to the table is model testing, similar to unit tests present in most software development languages.
Looker's data governance capabilities are fantastic. They ensure that everyone is looking at the same data - the same metrics. However, it doesn't ensure that the data feeding those metrics is correct. If someone makes a mistake in the SQL, everyone is still looking at the same metric, that metric just isn't accurate anymore.
Imagine the following situation. You use Salesforce as your CRM and you have a LookML view called sf_account:
Everything is working correctly. You're completely confident in the data generated in your reports.
Then, you decide you want to add the stage name from the related Salesforce opportunities to the model. Absent-mindedly, you forget that Accounts -> Opportunities is a one-to-many relationship. You make the following change and all your data from the sf_account view is suddenly fanned out, distorting metrics in all your reports.
The ID column was supposed to be your unique primary key and now it isn't. There is currently no automated mechanism in Looker to test the output of that query. No way of knowing that a record for ID 176 now has five records, one for each opportunity generated for that account.
In dbt, you can define assumptions on the model, giving you the ability to identify instances like these in your development workflow.
When you run dbt test, you'd realise your mistake and fix it before your Head of Sales starts shouting at you because all the numbers changed. Though testing of that sort hasn't always been commonplace in analytics, it's incredibly important so that everyone can have confidence in the reporting they consume, particularly when that reporting is self-serve.
Making data usage-agnostic
Claire from your data science team wants to build a fancy model, predicting how many flying mattresses you'll sell next week. To do so, she wants to use the customer_status dimension defined in your customers view in Looker. She doesn't want to have re-define it in her work. She understandably thinks you should all be using the same definition.
She jumps into her database explorer to find the name of the PDT in the looker_scratch schema, the location all the PDTs are built. She finds it. The table is called lr$zzzxu3bfa1lkc2x5ne75_customers. She goes off and builds his model. It has an AUC score of 0.96. She's incredibly pleased. She shares the analysis.
Your COO opens it up, tries to run it. To his (and Claire's) dismay, the following error occurs: ERROR: relation "looker_scratch.lr$zzzxu3bfa1lkc2x5ne75_customers" does not exist. The table doesn't exist anymore.
Every time you make a change to the SQL in a PDT, Looker will rename the underlying table in your database. As is the case here, Looker is often not the only platform or location where data analysis will be produced in your business. Often you'll have a data science function. Often you'll have people who need to write ad-hoc sql queries for specific pieces of analysis. As Claire initially asserted, all those functions should be building off the same models and the same definitions, building off the same single source of truth.
The way Looker builds its PDTs isn't conducive to that principal. The table names will continue to change (and aren't memorable for analysts to use in ad-hoc querying).
dbt will build your customers model in a table called customers in the schema of your choice (target in dbt parlance). That will never change unless you tell it to. It can be consumed by every user and every platform in perpetuity. Claire can use it for her analysis. She could even put that model into production because she can be confident that your analytics team controls if that table exists or doesn't.
One way around this with Looker is having people build models on top of the Looker API. This can work well, but still requires the data being exposed together in an explore. With external models, people can join up datasets that aren't otherwise modelled together in Looker.
Moving away from Looker
That leads me to a slightly meta point.
Looker is great. I love Looker. I wholly endorse Looker. I don't see that changing any time soon.
That said, eventually, you or I may choose to move to a different BI platform. It may no longer be the tool best suited for what you are trying to achieve.
At that point, given that the underlying data models need to be used by other users and platforms, you're going to need to retain those models somewhere. Might I suggest dbt?
SQL is a notoriously WET (write every time) language. It's veritably soaking with repetition.
Need to pivot out a column into 5 columns? 5 lines of SQL.
Need to group by 76 columns? (Not sure why.) You need to write out every number between 1 and 76.
Need to calculate the time between two dates, subtracting for weekends in between? The same lovely SQL snippet over and over and over.
Most people's SQL is rife with repetition. Most people's SQL is rife with repetition.
This isn't a Looker problem. But it is a problem that frequently presents itself in PDT modelling. dbt's Jinja-powered 'macros' lets you write far more DRY (don't repeat yourself) code. It allows for more manageable, scalable SQL queries that your analysts will be happy to manage.
Want to pivot that colour column? Initially, you might have written something like this:
Instead, you can write a small macro and generate SQL as follows:
Now, when you need to change how the logic of your 'pivots' work, possibly now present in many files, you only have to change it in the one location you've defined the macro.
When PDTs are best
There aren't many situations where Persistent Derived Tables are necessary over dbt models. However, there are many reasons why non-persisted, vanilla Derived Tables (DTs) are necessary. DTs practically function in a very similar way to PDTs, they just don't get materialised in the database.
DTs are necessary when your query requires user input. Have a window function that you want to calculate dynamically? You'll need a Derived Table.
DTs should be used when you want to extend an underlying model. Have a opportunities table that you need to turn into live_opportunities and lost_opportunities? You should used a Derived Table to filter the original opportunities table. You probably don't want to create each of those individual tables in your data warehouse.
Ultimately though, if you are using dbt as well, your Derived Tables should build from dbt models, not the raw source tables.
And your LookML views will still need to exist. You'll still need to define how measures are created. You'll still need to define how you want to expose your data to your end-user. Looker is vital for those steps.
Arguments against dbt
There are also reasons to not adopt dbt, distinct of Looker's functionality.
You would be adding another tool to your stack. That adds complexity and extra management responsibilities.
You would be adding another step to your development workflow. That requires a bit more work from analysts and engineers.
dbt is a batch-based tool. You wouldn't be able to re-build tables off of custom triggers like you can with PDTs.
All that said, I strongly believe the benefits dbt generates far outweigh the downsides. If you are using PDTs and starting to experience some of the pain-points outlined here, it is the ideal tool to add to your stack.
Moving from one to the other isn't a big leap. It's all still select statements, which in my opinion is the biggest draw to PDTs in the first place. Initially, all of your PDT SQL can be seamlessly picked up and dropped into a dbt project. I've seen migrations happen in days or weeks, depending on how much development you have already done in Looker.
You can also use both in tandem, slowly moving PDTs over to dbt models as you work. There's nothing stopping you from testing dbt out for a few models to ensure you're comfortable with it before making the final jump.