Whispering Data
Published in

Whispering Data

How To Measure a Data Engineering Team

Data teams love calculating and tracking everything with metrics. We already have the infrastructure in place to do so… yet often fail to apply the same strategy for our own work. Let’s fix that!

In my old office, there was a shelf that held tchotchkes and stuffed animals collected over the years.

When people would visit me, then serving as CTO of SimilarWeb, they would sometimes fiddle with these items. Before leaving work on most days, I made sure everything was neatly returned to its proper place.

I remember hurrying through this therapeutic ritual one afternoon to make it to an evening performance of The Book of Mormon… when I was interrupted by a regional sales director, Adia, asking if we could speak.

She was struggling to convert new trial users to enterprise plans. And issues with data insights my team provided in the product, she explained, were contributing to the drop in conversion.

Situations like these arise all the time in companies, and can go one of two ways.The first is characterized by time-consuming investigations, anecdotal decision-making, and endless meetings. The second features quick resolutions driven by constructive, pointed conversations.What determines which way a given situation will go?

Metrics as the Differentiator

In my case, the presence of internal metrics tracking the performance of my team anchored the ensuing discussion towards the second path.

I swiveled my monitor towards Adia and pulled up a dashboard showing an indicator of Data Quality over time. The metric was both stable over the last several months — even when isolating for her region — and above the internal target we set at the beginning of the year.

“Either this metric isn’t reflective of reality.” I said pointing to the screen, “Or, there is a disconnect between what you believe is hurting conversion rates and what actually is. Let’s work together to figure out which it is.”

From this objective terra firma, we were able to have a constructive discussion on what was more likely responsible for the decline. Satisfied, we both left the office. I hailed a cab and made it to my seat right as Elder Price came on stage…

Picking Meaningful Metrics

What this (only partially!) fictionalized situation illustrates one of the benefits you’ll get from measuring the output of a data engineering team. While most people agree this is something they should do, many get stuck on what metrics to track, where to track them, and when they should be reviewed.

Data Quality was in fact one of three primary internal metrics we used to measure the data engineering unit. Along with quality, we looked at Uptime and Velocity as additional metrics to gauge ourselves.

Why these three and how did we define them? Read on to find out!

Metric #1: Data Quality

Data quality can be measured by the number of incidents triggered by data issues. An incident can be defined to include internal alerts, failed tests, and issues raised by external consumers, if relevant.

At SimilarWeb, the data engineering team’s output was surfaced to external users since we were, primarily, a data product. Therefore any issues raised to customer support about data issues negatively impacted our data quality metric.

The option to include weighting for severity of incident and resolution time are there, but will involve additional complexity to incorporate.

Higher is not always better

It is tempting to pick a really high number to set as your data quality goal.

“We want the data to be accurate 99% of the time,” or “There should be no incidents on 95% of days.” may sound good on the surface. However, aiming high isn’t always worth it because your team will pay an opportunity cost to maintain it.

This takes the form of time spent investigating data quality problems that prove rare or shouldn’t be prioritized. Another important metric that we’ll discuss later, velocity, will slow down as a result.

Instead of aiming to maximize one metric in a vacuum, it’s important to be mindful of the potential trade-offs between them. Choosing a data quality goal of, say, 80% may be right for you.

The exact number isn’t as important as finding a monthly or quarterly cadence to review and tune it!

Metric #2: Data Uptime

Data uptime is the percent of time a dataset is delivered on time, relative to the expected frequency or SLA requirements.

Let’s look more closely at these two variables:

  • Expected Frequency: How often a dataset is expected to be updated. Most commonly daily, hourly, or real-time.
  • SLA (Service Level Agreement) Requirement: The frequency clause stipulated by an SLA agreement between a data producer and consumer for when data must be updated by. For example, an SLA might state an absolute time like 7:00am or a relative time like every 3 hours.

If multiple data sets are delivered, we should measure uptime per data set. To combine the uptime stats from all our data sets into one metric, we can simply use the weighted average. Just like with application development uptime, we can use the Nine System to define what our data uptime should be.

It is fine to look at just one of the expected frequency or SLA stipulations to determine uptime, but I recommend using both to paint a full picture of your data delivery performance.

Metric #3: Development Velocity

Just like in app development, we can look at velocity — the number of user story points completed per iteration — when measuring data engineering teams. Why mention it here if it’s the same metric for app development?

I’ve heard the argument that data engineers cannot work in an agile environment because data research and exploration don’t fit the framework. My experience managing data orgs for the past decade leads me to a different conclusion.

Research can absolutely be expressed in small, testable user stories and epics. Furthermore, I’ve seen that working in agile methodology and measuring velocity, even when research is part of the work, improves the efficiency of data engineering and data science teams.

Final Thoughts

If you feel like you have no way to quantify your data engineering team, tracking data quality, data uptime, and velocity is a great place to start. Insights from these metrics will allow your team to make informed technical decisions and better communicate the impact of those decisions outside the org.

One of the hardest things to get organizational buy-in for is an initiative aimed to reduce technical debt. Explaining the benefit in terms of delivering more accurate data faster (and having numbers to back it up!) is an effective way to do so.

Ultimately, you can measure any number of KPIs, but my recommendation is to pick three-to-four that are focused and simple to understand; otherwise, you’ll optimize for too many things.

By the way, LakeFS the open-source project I created (along with my co-founder Oz Katz), influences all three metrics by providing simplified workflow management to data environments.

Originally published by Einat Orr, P.h.D. on the lakeFS blog.




Whispering Data is a Medium publication for all the data & productivity secrets you wish you knew years ago!

Recommended from Medium

Game Analytics

Calculating the Probability of Loan Repayment — Using MLOps for Credit Scoring

Variable Evaluation Within Exploratory Data Analysis

14 Data Science Projects to do During Your 14 Day Quarantine

PCA Algorithm Tutorial in Python

A Comprehensive Roadmap for An Aspiring Data Scientist in 2021

What I learnt from building Mumbai’s largest offline Data Science Community?

DeRF: Decomposed Radiance Fields

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Paul Singman

Paul Singman

DevRel @lakeFS. Ex-ML Engineering Lead @Equinox. Whisperer of data and productivity wisdom. Standing on the shoulders of giants.

More from Medium

Data as Code — Cutting Things Smaller

Stop Treating Your Data Engineer Like a Data Catalog

Data Quality and Reliability

What is a Feature Store?