In the previous blog posts (part 1, part 2, part 3, and part 4) in this series, we talked about why we decided to build a marketing data warehouse. This endeavor started by figuring out how to deal with the first part: making the data lake. In the fourth blog post, a more technical one, I’ll give some insights into how we’re leveraging Apache’s Airflow to build the more complicated data pipelines, and I give you some tips on how to get started.
This blog post is part of a series of five? (maybe more, you never know), in which we’ll dive into the details of why we wanted to create a data warehouse, how we created the data lake, how we used the data lake to create a data warehouse. It is written with the help of @RickDronkers and @Hussain / MarketLytics, who we’ve worked with alongside during this (ongoing) project.
Getting Started with Cloud Composer
Cloud Composer is part of Google’s Cloud Platform and brings you most of the upside of using Apache Airflow (open source) and barely any of the downsides (setup, maintenance, etc.). Or to follow their main USP: “A fully managed workflow orchestration service built on Apache Airflow.” While we had worked with Airflow before, we weren’t looking forward to spending time having to worry about its management as we planned to spend most time setting up and maintaining the data pipelines. In the end, then you have to stick to create pipelines (DAGs).
What is it suitable for?
You want to load data from the Google Analytics API, store it locally, translate some values to something new, and have it available in Google BigQuery. However, you would build it; it’s multiple tasks and functions that are depending on itself. You wouldn’t want to load the data into BigQuery when the data wouldn’t have been cleaned (trash in, trash out sounds familiar?). With BigQuery, the next task is only being processed if the previous step was successful.
Tasks are in almost every case; just one thing: get data from BigQuery, upload a file from GCS into BigQuery, download a file from Cloud Storage to local, process data. What makes Airflow very efficient to work with is that the majority of data processing tasks already have pre-built functions. The first three tasks that I listed here are operators (GoogleCloudStorageDownloadOperator, GoogleCloudStorageToBigQueryOperator) that operate as functions.
Versus Google Cloud Functions
If you mainly run very simple ‘pipelines’ that only exist of 1 function that needs to be executed or have only a handful use cases, it is likely overkill to leverage Cloud Composer; the costs might be too high, you still have overhead with DAGs. In that case, you might be better off with Google Cloud Functions as you can write similar scripts that will enable you to also trigger them with Google Cloud Scheduler to run at a specific time.
The costs for Google Cloud Composer are doable, for a basic setup, it’s around 450 dollars (if you run the instances 24 hours * 7 days a week) as you leverage multiple (a minimum of 3) small instances. For more information on the costs, I would point to this pricing example.
See above an example data pipeline, in typical Airflow fashion every task is depending on the previous task. In other words: notify_slack_channel would not run if any of the previous tasks would fail. All tasks are happening in a particular order from left to right. In most cases, data pipelines become more complicated as you can have multiple flows going on at the same time and combining them at the end.
Tips & Tricks
Google Cloud Build, Repositories
The files for Google Cloud Composer are saved in Google Cloud Storage. Which is smart in itself, but at the same time, you want them to live in a Git repository so you can efficiently work on it together. By leveraging this blog post, you’re able to connect the Cloud Storage bucket to a repository and set up a sync between the two. This will help you build a deployment pipeline basically and make sure that only production-ready code from your master branch ends up in GCS.
After working with it for a few months now, I’m still not sure if managing dependencies through Google Cloud Composer is a good or bad thing, as it creates some obstacles if you want to run a deployment and want to add some Python libraries (as your servers could be down 10-30 mins at a time). For other setups, this usually is a bit more smooth and creates less downtime.
Sendgrid for Email Alerts
One of the upsides of Apache Airflow is that it sends alerts upon failure of tasks. Make sure to set up the Sendgrid notifications while you’re setting up Google Cloud Composer. This will be the most straightforward way of receiving email alerts (for free, as in most cases, you shouldn’t get too many failure emails).
Document the crap out of your setup and DAGs. When I took over some of the pipelines that were used at Postmates for XML sitemap generation it was a nightmare, it was hard to read, the code didn’t make a lot of sense, and we had to refactor certain things just because of that. As sometimes pipelines (just like regular code) can be left untouched/unviewed for months (as they literally sometimes only have one job) you want to make sure that you come back and understand what happens inside the tasks.
In the previous blog posts (part 1, part 2, and part 3) in this series, we talked about why we decided to build a marketing data warehouse. This endeavor started by figuring out how to deal with the first part: building the data lake. In the fourth blog post, we’ll chat about how we are visualizing all the data we saved in previous steps by using Google DataStudio.
This blog post is part of a series of four? (maybe more, you never know), in which we’ll dive into the details of why we wanted to create a data warehouse, how we created the data lake, how we used the data lake to create a data warehouse. It is written with the help of @RickDronkers and @Hussain / MarketLytics who we’ve worked with alongside during this (ongoing) project.
How we build dashboards
Try to think ahead about what you need: date ranges, data/date comparisons, filters, what type of visualization. This will help you build a better first version right away as it gives you the ability to have a good version right away. What that mainly looked like for us:
Date ranges: The business is so seasonal that our Year over Year growth is most important for RVshare, and since we often don’t get to see all the context on metrics on a weekly basis, we default to 30 days.
Filters: For some channels (PPC, Social), it’s more relevant to be able to filter down the data on a campaign or social network level. Because in most cases the aggregate level doesn’t tell the whole story right away.
Visualization: We need the top metrics: sessions and revenue in view right away with the comparison YoY right away so we know within seconds what is going on and how that can improve things.
Talking to Stakeholders (Part Deux)
In the first blog post, we talked about connecting with our stakeholders (mainly our channel owners) and gathering their feedback to build the first initial versions of their dashboarding (beginning with the end in mind). We used this approach to put the first charts, tables, and graphs on the dashboards after which we connected back again with the owners to see what data points were missing and in some cases to validate the data that they were seeing on their dashboards. This helped us get additional feedback for fast follows and made for quick iterations on data that we had and could also show. For social media, as an example, it turned out that we wanted to show additional metrics that we hadn’t thought of initially but were in our data lake anyway. These sessions provided a good way for us to build additional pieces into our data warehouse while we were at it. These days some of these dashboards are used weekly to report to other teams in the organization or used within the team itself.
Do you want to add this to Google DataStudio, or do you want to create synced/aggregate tables in BigQuery? For most of our use cases, we have opted for using DataStudio to create JOIN blended data sources. It’s easier – we have the ability to quickly pull some new data together versus having to deal with the data structures and complicated queries. In some use cases, we noticed that we were missing data in our warehouse tables (not the lake) and were able to make adjustments/improvements to them by creating dashboards.
Single Account Owner
Because we work with Rick and Hussain as ‘third parties’, we opted for using 1 shared owner account, transferring owner access is incredibly hard when it’s a Google Apps account so we made sure that the dashboards are owned through an @rvshare.com account, it’s not a big topic but could cause tons of headache in the long term.
Keep It Simple St*pid
Your stakeholders probably have the desire / and the time to look at less than you think. Instead of having them jump through too many charts start simple and then add based on feedback if they want to see more, less is more in this case.
This has added benefit of them feeling engaged and more interested in using it. In our own use case, we leverage our reporting on a weekly basis for a team meeting, which makes it already a more often leveraged us case.
Calculated Fields – Yay or Nay?
As we made most of the tables that we leverage in DataStudio from scratch during our ETL process, we had the opportunity to decide if we wanted to leverage calculated fields in BigQuery or if we do the work in the queries itself. Honestly, the answer wasn’t easy, and as we made modifications in the dashboards, it became clear that having them set up in DataStudio wasn’t always scalable and easy as with data modifications or changes in tables they are removed.
Tables or Queries? In our case, we often used the table information from BigQuery and the specific columns in there to drive the visualization in DataStudio. The alternative for some of them is that we directly query the data in BigQuery, with the BI Engine reservation that we have in there we can speed up intense queries rather easily.
In the previous blog posts (part 1 and part 2) in this series, we talked about why we decided to build a marketing data warehouse. This endeavor started by figuring out how to deal with the first part: building the data lake. We’ll try to go a bit more into detail on how you can do this yourself in this post in which we transformed our marketing data lake into an actual data warehouse.
This blog post is part of a series of four? (we found enough content to add more articles ;-)), in which we’ll dive into the details of why we wanted to create a data warehouse, how we created the data lake, how we used the data lake to create a data warehouse. It is written with the help of @RickDronkers and @hu_me / MarketLytics who we’ve worked with alongside during this (ongoing) project.
The Process of Building a Data Warehouse
In our endeavor of building a data warehouse, we had a couple of big initiatives that we first wanted to get done. We needed some reporting and visualization tables and aligned with that, we needed to make sure that we could have data that was cleaned for other purposes (deduplication, standardization: some typical ETL problems).
In order to streamline the processes, we used three different ways of getting the data streamlined:
Google Cloud Functions
Google Cloud Functions is used for both transforming our data as loading our initial data for a few use cases. Early on we noticed that not every vendor was available through regular data loading platforms, like StitchData. An example of that was Google Search Console, as we didn’t want to have the need to run additional infrastructure for just dealing with Load scripts we leveraged Cloud Functions to run a daily script (with support from Scheduler to make them daily).
After loading the data we also Transform some other tables from our marketing data lake to new production tables using Cloud Functions. All of our scripts are currently written in Python or Node.js, but as Cloud Functions makes it possible to deal with multiple languages it provides us with the flexibility to leverage others over time.
Backfill: As Functions can easily be rewritten and tested within the interface, it also provides us with a good way to backfill data easily as we can easily adjust the dates that a script needs to run.
In some other cases, we can also leverage Google BigQuery’s scheduled queries. In a few instances, we just want to load the data from raw data lake tables into a production table. Mainly because we don’t always need all the columns, we can limit our data drilling and be able to clean the data in the query itself. In that case, scheduled queries can come in pretty handy as they run on a certain schedule, can be easily updated, and already point towards another data set and table.
For more complicated data flows we’re currently using Airflow via Google Cloud Composer. Cloud Composer, as we mentioned in a previous blog post, enables us to not have to worry about maintaining the Airflow infrastructure but gives us all the other upside of it. This gives us the ability to focus on creating and maintaining the DAGs that help drive the actual data structuring flows.
How we mainly use Airflow is to combine, clean and enhance data from multiple sources and then reupload it back into Google BigQuery for visualization in other tools. Singular use cases are more easily captured by one or two tasks, but in Airflow we run flows that usually have multiple tasks that need to be executed in a certain order and not at all if one of them fails. This is what Airflow is meant to do, and that’s how we’re leveraging it too. As an example of our affiliate marketing campaigns, we have a structure set up that only pays out once travel is concluded (a very standard approach in the travel industry). This means that we need to retrieve orders from our partner > verify them with our database > create a new format to upload back to our vendor and run the actual upload. In addition, we want to set up some alerting for the team as well. Resulting in 6 tasks in this case that need to be executed in the right order: the perfect use case for Airflow.
In the previous blog post, I touched on how we wanted to set up raw tables that are transformed once or multiple times. We decided to do this to both make the data more streamlined and also to make them ready for visualization on our channel dashboards. The MarketLytics team did a great job documenting this with a very visual result that you can see here:
As discussed previously, we go through multiple stages with the data that we get into the data lake and transform it to the data warehouse.
Example of data enhancement: One of the most common scenarios that we’ve tried to solve for was connecting existing data from a vendor back to the data that we receive in our web analytics tool: Google Analytics. As an example, if properly tagged we should be able to get the data from a specific newsletter campaign from the UTM parameters and then connect the data to (in our case) Marketo to what we have there on deliverability and open rate (%).
This blog post is part of a series of three, in which we’ll dive into the details of why we wanted to create a data warehouse, how we created the data lake, how we used the data lake to create a data warehouse. It is written with the help of @RickDronkers and @hu_me / MarketLytics, who we’ve worked with alongside during this (ongoing) project.
Six months ago, we were looking at Google Sheets, Google Analytics, and many other tools/vendors daily and had no marketing data live in a data warehouse (nor a lake). We wanted to get better at using (marketing) data but mostly find a better way to connect insights from multiple sources to find insights but also make more data available to the rest of the internal teams (think: PPC, SEO, Email, Product).
Prove business value: Everything we do as Marketing needs to prove some kind of business value. We’re not doing things just to keep ourselves busy. So we wanted to make sure upfront that the business cases that we had in mind would have an upside by creating this setup. Some of these were: detailed insights into social media performance, attribution level data, and a way for us to connect transactions between data sources.
Access to ‘our’ raw data: As often with marketing data, you’re using the interface of the actual marketing vendor to pull your daily numbers. But since so much of the marketing data overlaps with internal data from other channels we wanted to get access to all the raw data that is ‘ours’ (to begin with) and have the ability to combine it. This, for example, would provide us with insights on the lifetime value of customers by channels or the ability to export cohorts of users.
The analysis doesn’t scale in a spreadsheet and isn’t repeatable: Besides creating reporting monkeys, we wanted our teams to not have to spend time pulling data from different sources and manually combining it. It’s too high a risk for potential mistakes, the process doesn’t scale, and valuable time of team members would be lost in the process.
More control of data quality: We want to control our data pipelines and be involved in the ETL (Extract, Transform, Load) process so we know how data is shaped. There shouldn’t be a black box of data being created in the background. With having this process potentially in place, we can control the quality and usefulness of data more (and continue more over time).
More flexibility for visualization and ‘deep analysis’: With having access to our raw data in a warehouse, we knew that we would have to rely less on building analysis right on top of vendor’s APIs or worry about potential issues on how to visualize data (on dashboards).
Looking at the complete picture
Marketing data is often confined within the platforms and tools used to distribute the marketing message. But your prospects don’t confine themselves to just one source or medium to interact with your company (the path length is usually longer than one session before a transaction takes place). To be genuinely able to measure the effectiveness of your marketing activities, extracting the data from these platforms and having the ability to analyze them is essential. This goes back to our first point on Why. It needs to solve cases that provide business value.
Beginning with the end in mind
One of the biggest challenges in working with data is that often you don’t actually know precisely what data you need to solve for your use case. Usually, you tend to go through the cycle of “capture – store – enrich – analyze” a couple of times before arriving at really valuable insights.
By creating your own data warehouse, you create the flexibility to transform and relate the data to each other in ways that aren’t possible when your data is locked within separate platforms.
Build versus Buy?
While it was clear that we wanted to own as much of the data ourselves, an apparent next discussion was how to gain access to the actual data. Would we build the pipelines ourselves? Would we buy tools to do so, and how much of the transformation of data would fall on the team? How about maintenance?
Pros of Buying a solution
Plug and Play: A strong Pro for buying software usually is a decreased time to value because you’re skipping the building phase. So you avoid needing to have a strong learning curve and having to deal with the hard/soft-ware setup. What was a requirement for us is that we didn’t want to deal with a long process of setting up resources.
Scalability: As we grow, our datasets will grow. While paying extra for additional volume isn’t a problem, we wanted to make sure that the infrastructure would hold up and could last while we reach a certain factor of scale.
Cons for Buy & Build
High Costs: Buying tools would likely, over the long term, have more direct costs associated with it. While obviously, the indirect costs for building the solutions ourselves (increased headcount) would also be expected to be high.
Vendor Buy-in: Regardless of either buy or build, it’s a strong Con that we’re basically bought into a vendor’s platform that we wanted to make sure that certain aspects of it are able to be moved over to another service if we deem that to be the right choice at that point in time.
Standard data formatting: With some vendors, it was harder to change the data schema (often for a good reason) or pull additional data formats that we wanted to collect from, for example Facebook (the data was available in the API, just not in the vendor’s export).
Time to first use: Building pipelines for extracting the data from certain platforms could take a long time. But obviously the real business value will, in most cases, only show up when you’re actively working with the data. Losing months while dealing with extracting the data was something (also taking into account our seasonality, which provides a ton of data) that would delay this. So going with platforms that could directly provide us with extracted data was a great pro versus a strong con for building it ourselves.
Maintenance: We didn’t want to maintain extract pipelines from the marketing platforms that we use. While it’s a really important part of our data strategy, we wanted to avoid the need to spend a significant amount of our time to stay up to date and maintaining pipelines. With a buy solution, we would rely on the resources of the vendor to deal with this part of the data flows, which is very much what we would prefer at this point, mainly considering that the resources of our data team are minimal.
In the end, we decided to go with a combination of the two strategies. We decided to let most of the Extract processes be handled by software that we would buy, or that would already be available opensource as this is a task that we wanted to make as cost-effective as possible. Most of the Transform will be done through a set of tools that we would maintain and build ourselves as there are too many custom use cases and needs. With that, it meant that we also needed to ‘Load’ the data back into our data warehouse mainly ourselves. More on the details of that in our upcoming blog posts.
An often asked question in this process is what kind of costs we were expecting to have. While we won’t disclose the costs for the additional help received, we’ll quickly touch on the costs for this whole setup thus far. Parts of this are variable based on the size of your data but so far our costs:
Google Cloud Platform: ~$400-700 monthly.
This includes costs for Google BigQuery, Cloud Composer and other tools. We’re expecting that this will increase at some point but gradually will go up over time.
Note: In another blog post we might dive deeper into Google Cloud Composer and how we leverage that for Airflow (processing data pipelines). Right now, Cloud Composer presents at least half the costs for Google Cloud Platform. If you decide not to use it, your costs will be significantly lower for GCP.
Stitch Data: $ 350-500 monthly based on volume for now. We’re expecting that this will increase at some point but gradually will go up over time.
SuperMetrics: $190 monthly per license per vendor (depending on the number of licenses you might need).
Strategy going forward
With some of these decisions made upfront, we decided to move forward with collecting the data channel by channel, starting with the most important channels. This way, instead of needing to collect all the data for every channel, we were able to directly start building dashboards on top of the data that would be useful for teams. The steps to take to get there were as follows, and we’ll dive into some of them in detail in upcoming blog posts:
Extract all the data using Google BigQuery Transfer Service, Supermetrics and StitchData.
Validate that the data in Google BigQuery is correct by comparing it to the regular data sets, existing dashboards, and the data that we would look at in the vendor’s platform.
Transform the data so data sources can be combined and Transform the data to be directly useful for reporting purposes.
Load this new data into new reporting tables, backfill historical data where possible, and sync it back into Google BigQuery.
Depending on the channel, build dashboarding in Google DataStudio or Tableau.
In the next blog post, we’ll go into depth on how we created a marketing data lake, where all our raw marketing data from our vendors (Google Search Console, Google Ads, Bing Ads, etc.) lives. Expect that to go live within the next two weeks. Until then, feel free to share your thoughts on this blog post on Twitter. Again… This blog post is written with the help of @RickDronkers and @hu_me / MarketLytics, who we’ve worked with alongside during this (ongoing) project.
Updated June 4, 2019: Added a YouTube video which will guide you through the setup process in RStudio and how to run the script yourself.
Averages lie & average click-through rates aren’t very helpful! Here I said it. What I do believe in, is that you can calculate click-through rates (CTR) for your own site in a great way though. With data from Google Search Console. Especially while using R, so let’s dive in on how this works and what you can do with it!
Why ‘Averages’ Lie?
Look at the graph below, a great CTR for position 1. But for example, this research shows that the average CTR for position 1 is: 24% (AWR data, Feb 2019). Which one is correct? Neither of them. As it really depends on the industry, what features show up in the search results that might decrease CTR (think rich snippets like local packs, news results). All of this is making it really hard to make a good analysis of what you could expect if you rank higher for a bunch of keywords in your industry. So while I was working at Postmates on ranking certain category pages better we decided to calculate our own CTR and were intrigued by how far CTRs were off from research (the research isn’t wrong! It’s just generalized across industries). Eventually, with the data in hand, we were able to make better estimates/forecasting of how much traffic we could expect when rankings would increase in that segment. In the rest of this post, I’ll go more in-depth on the specific practice on how we calculated this.
Using Google Search Console Data
Visual of Google Search Console Report (this is not RVshare data)
You’ve seen this report in Google Search Console, providing you with a detailed view of the performance of your keywords and the average position for your keywords. In this graph, we see something positive, a CTR & position that go up slightly over time. But what if you would want to know the average CTR for a certain segment of keywords per position. That’s way harder to do in the interface. Because of that, I used the R script from Mark Edmondson that he wrote about here almost three years ago.
It will help you extract the data from Google Search Console in a raw way so you can use it to digest it and create your own visualizations (like the one we’ll talk about next).
Visualizing CTR Curves in R
CTR Curve visualized CTR per Position (note: this is randomized data for an unknown site)
So let’s dive right into how you can do this yourself, I’ll provide you with the full R script and you will need to download RStudio yourself in step 1.
Run these commands to install the right packages for RStudio:
install.packages(“ggplot2”) if necessary
Line #21 – Change this to the property name from Google Search Console
Line #25 – Not neccesary: If you want the CTR curve for positions over 20, change the number.
Line #40 – Recommended: Exclude the word(s) that are part of your brand name. So you get the right CTR curve for non-branded keywords only
Line #41 – Not necessary: This script is taking a ‘sample’ of 50.000 keywords to calculate your CTR curve of. You can increase this limit to more if needed, if you have less than 50.000 keywords it’s not an issue
Run the script! The output should be a visual as shown earlier in this post
Want to take a deep breath and let me help you go over this again? I’ve made a quick screen share video of what to do in RStudio and how to use the R script.
Hopefully, now, you’ve had a better chance to understand what the actual CTR is for your own site and you can use this to visualize CTR curves for specific parts of your site or pages that have a similar META description. Over time you could use this, for example, for measuring the impact on CTR.
Credits where credits are due! There are many use cases for using CTR data by visualizing it with R, and I’m grateful that a while ago Mark Edmondson opened my eyes about this + credits to Tim Wilson’s documentation on using R and improving visualizations.
Want to read this article in Spanish? Read it here.