Jun 12, 2019 · 34m · mad
Fireside Chat: Solmaz Shahalizadeh, VP of Data Science & Engineering at Shopify (Data Driven NYC)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
At a Data Driven NYC fireside chat hosted by Matt Turck, Solmaz Shahalizadeh, VP of Data Science & Engineering at Shopify, discusses Shopify's data culture, team scaling, infrastructure evolution, practical machine learning deployment, and hiring strategies.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 17.9% of the talking time here. How this is scored →
Matt as informed peer 2.8 Guest teaching 2.7 Guest disagreement 0.4 Matt pushing back 0.2
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 28:57 Rejection of marketplace framing
Solmaz directly rejects the audience member's premise by stating 'So we are not a marketplace. We are actually a platform,' establishing a firm boundary on Shopify's business model.
Hardest push from Matt ▶ 7:06 Probing central data team structureMatt interrupts Solmaz's explanation of embedded teams to test her framing, asking 'So there's no central data science?' which prompts Solmaz to clarify their hybrid reporting structure.
Biggest teaching moment ▶ 28:57 Platform vs marketplace correctionSolmaz explicitly corrects an audience member's core assumption about Shopify's business structure, explaining that Shopify enables individual merchant brands rather than operating a centralized marketplace.
Matt holds his own ▶ 6:03 Citing historical coining of 'data scientist'Matt showcases insider tech knowledge by sharing that Jeff Hammerbacher and DJ Patel coined the job title 'data scientist' during an early event hosted by FirstMark.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Event Title Slate and Speaker Introductions | 2 | 1 | 0 | 0 | Matt sets the stage by welcoming Solmaz and referencing FirstMark's early investment in Shopify alongside its $30B market cap. Solmaz shares her early exposure to punch cards and early neural network research at Sloan Kettering. | |
| Joining Shopify and the Evolution of Data Science | 3 | 2 | 0 | 0 | Matt demonstrates industry knowledge by noting that Jeff Hammerbacher and DJ Patel coined the term 'data scientist' at an early FirstMark event. Solmaz notes that fields like bioinformatics, astrophysics, and social sciences were doing data science long before the job title existed. | |
| Shopify Data Team Structure and Embedded Model | 4 | 3 | 0 | 1 | Matt asks if Shopify operates without a central data team after learning about their embedded model. Solmaz clarifies that while team members are embedded in product units, they maintain a unified organizational reporting structure and a shared central data lake. | |
| Machine Learning Use Case: Shopify Capital | 4 | 3 | 0 | 1 | Matt asks informed questions regarding whether models plateau or get retired due to poor user experience. Solmaz details how moving from rule-based logic to initial machine learning yields the biggest jump, while subsequent improvements yield diminishing returns. | |
| Machine Learning Product Portfolio and Feature Strategy | 2 | 2 | 0 | 0 | Matt asks how Shopify prioritizes new machine learning features across its product suite. Solmaz emphasizes a merchant-first approach, prioritizing tangible product solutions over complex algorithms for their own sake. | |
| Applying Monica Rogatti's Data Science Hierarchy of Needs | 4 | 3 | 0 | 0 | Matt references Monica Rogatti's Data Science Hierarchy of Needs to frame the discussion. Solmaz walks through the layers, explaining how jumping straight to top-tier ML without solid data pipelines generates heavy technical debt. | |
| Recruiting and Onboarding Strategy for Data Scientists | 3 | 3 | 1 | 0 | Matt asks how Shopify successfully recruits top data science talent. Solmaz offers a counter-intuitive strategy, explaining that she actively avoids hiring traditional computer science ML specialists in favor of curious domain experts from astrophysics or economics. | |
| Predictions for AI: Accountability, Human-AI Synergy, and Causal Inference | 2 | 3 | 0 | 0 | Matt asks Solmaz for her forward-looking AI predictions. Solmaz highlights transparent algorithms, human-AI synergy, and causal inference as critical areas that deserve more investment over pure machine learning hype. | |
| Audience Q&A: Platform vs. Marketplace Data Architecture | 1 | 4 | 3 | 0 | An audience member asks about marketplace data architecture. Solmaz immediately rejects the premise of the question, clarifying that Shopify is a merchant platform rather than a marketplace. | |
| Audience Q&A: Applications and Open-Source Tools for Causal Inference | 3 | 3 | 0 | 0 | Solmaz explains how causal inference tools like difference-in-differences help evaluate product impacts when A/B testing is impossible. Matt concludes by asking about practical ML tools, prompting Solmaz to highlight Facebook's Prophet for time-series forecasting. |
Statements from this episode (16)
Shahalizadeh: Shopify supports over 800,000 merchants across 175+ countries
“Last reported number, sort of, is that we are over 800,000 merchants in over a 175 countries.”
Shahalizadeh: Shopify expanded from 300 to over 5,000 employees
“When I joined, we were 300 people, and now we are over 5000, so it's been massive growth, really, over the years.”
Shahalizadeh: Compute advances reduced two-year neural net tasks to one week
“What took me two years along with a team to do, like, right now with the advances over the last few years in compute and in storage, you can do probably over a week.”
Shopify transitioned its ETL platform from Ruby to Python around 2013-2014
“Shopify, the application is still Ruby on Rails, and actually the first version of ETL tooling that was built in house was Ruby, but I think we made this decision of moving to Python around 20 13, 20 14, because we thought it's the language that most of the co…”
Shopify Capital is entirely driven by machine learning
“As of last two years, Shopify Capital is entirely machine learning driven.”
Shahalizadeh: Shopify processes over 600,000 checkouts during peak periods
“At peak time we had over 600,000 checkouts going through”
Shahalizadeh: Moving from rules to machine learning yields highest lift
“The first time you go from any rule base or non-machine learning to machine learning actually is the time that you get the highest lift.”
Shahalizadeh: Shopify admin dashboard home cards are driven by machine learning
“When they log into their Shopify store in the admin, they see a bunch of, like, home cards that tells them do this or, like market for this product or make this change in your theme. And those are all driven by machine learning,”
Shahalizadeh: Shopify avoids asking merchants for data without immediate utility
“We don't ask merchants for data that we are not going to give a benefit to them right off the bat.”
Shahalizadeh: Shopify incurred technical debt by jumping straight to ML
“And there has been times that even us, we've sort of, like, jumped to the top top in one of the areas, and you pay the technical debt that you did not do the work in the beginning.”
Shahalizadeh hires machine learning talent by looking outside traditional ML backgrounds
“I recruit machine learning people by not going after machine learning people.”
Causal inference will solve key data problems ignored by machine learning hype
“There's a lot of hype and focus and good work in machine learning. But I do think if we take a step back and look at causal inference, we're gonna actually solve a lot of problems with just that arm. So I hope over the next few years, like, people invest in th…”
Shopify shadow-tests machine learning models using Apache Kafka before release
“We run them in shadow mode, which means, like, they do all the evaluation, but the results are logged to Kafka, and we see if the time to response, if the sort of distribution of predictions are different or not.”
Shahalizadeh: Shopify ML models are built for individual merchant use cases
“So actually all of our machine learning offerings that I was talking about are actually built for the individual use case of the merchant.”
Shahalizadeh: Answering simple questions drives more data impact than machine learning
“And I think that's sort of a neglected part of data science. Everyone talks about, like machine learning and building all of these data products, but I think lots of the impact, lots of the trust and usage of data actually comes from being able to answer to so…”