What Every Engineer Should Know About FinOps

What Every Engineer Should Know About FinOps

At Duolingo, cutting costs doesn't mean cutting corners.

What is FinOps?

At Duolingo, user experience comes first. If a cost-saving change hurts the user experience, it's not worth it. But outside of that, there are plenty of opportunities to maximize cost transparency, minimize waste, and empower engineering teams with actionable insights.

FinOps bridges the gap between Engineering and Finance, so everyone can work together to reduce waste. A recent example: one of our FinOps dashboards flagged a sudden 10x spike in OpenAI costs tied to a specific category of Max users. Finance identified the trend, reached out to the product team, and collaborated with engineers to quickly revert the costly change.

Without that visibility from FinOps, the spike could've gone unnoticed for weeks. Instead, it was resolved in hours. That quick resolution wasn't luck. It was the result of how we approach FinOps at Duolingo. Here are the principles we use to catch and prevent issues like it.

Cost visibility leads to cost insights

A single dollar amount tells us that we're spending money, not where or why. That's why we tag everything. The more granular the data, the easier it is to find savings.

An engineer on our team added internal cost tracking for our text-to-speech (TTS) costs. By breaking down spend by user type and feature, they discovered that Video Call feature's TTS costs were 60-100x higher than that of the Roleplay feature, even though both had similar TTS functionality.

The key difference? Roleplay had in-memory caching for TTS responses while Video Call didn't. Our team implemented a similar cache for Video Call and cut our overall TTS costs by 40%! The problem wasn't hard to solve; it was just hard to see.

Because we track costs at hourly granularity, we don't just see that something changed; we see exactly when it changed. When we deploy an optimization, we can tie cost drops directly to the rollout window, and when something goes wrong and costs spike, we can narrow the investigation to just a few hours of changes, rather than sifting through an entire day or week of deployments. It makes root cause analysis faster and cost impact more measurable.

Unit economics are another way to make costs visible by putting dollar amounts in terms of units your stakeholders care about. Engineers can think in terms of cost per request or instance. Finance can think in terms of cost per DAU, session, or feature. Everyone's speaking the same language, just mapped to their priorities.

Recommendation

Start by tagging and attributing your costs with as much temporal granularity as possible. Even basic visibility can unlock easy wins.

Reduce waste, not cost

One of our operating principles is Take the Long View, so user experience comes before short-term savings. And if a cost-saving change hurts the user experience, it's not worth it. Duolingo invests in higher-quality, more expensive LLMs for Video Call to ensure that Lily sounds like her deadpan, sarcastic self while still speaking at the learner's level. We could slash costs by switching to cheaper models, but Video Call with Lily wouldn't be the same. Strict budgeting policies that focus solely on cost can stifle innovation and slow long-term growth.

At Duolingo, teams are encouraged to spend when it drives value. The FinOps Team just helps make sure it's cost-efficient. Not all cost savings come with trade-offs. In other cases, we've identified pure waste-opportunities to reduce costs without sacrificing quality or value. We're one of the only teams that celebrate neutral metrics in experiments. Our goal is always to save money without harming learning. We migrated our most expensive CloudWatch metrics to Prometheus, reducing costs with no loss in observability.

Recommendation

Cost efficiency should support innovation, not restrict it. Prioritize long-term value over short-term savings. Spend when it adds value. Save where it doesn't.

Tech debt costs money

Outdated infrastructure doesn't just slow you down, it racks up cloud bills. Wins for our team often come from small efforts with big ROI. We achieved substantial cost savings by implementing small changes, such as switching to more cost-effective instances, eliminating unnecessary API calls, upgrading to Graviton, and updating the Python and Java versions of our services. We empower other engineering teams by solving the tricky parts once and documenting the process, so our high-impact fixes are replicable. We highlight our savings and performance wins, making it easier for teams to justify the work.

Recommendation

Framing tech debt in terms of real dollars can help you get buy-in from managers and leadership.

Let automation do the heavy lifting

FinOps at Duolingo is a small but mighty team of five. Manual number crunching just doesn't scale. AWS bills alone are massive, and that's just one vendor. CloudZero helps us make sense of it all by transforming raw cloud costs into user-friendly dashboards, so we can segment and filter the data to root cause many cost increases and show off our cost decreases.

As a FinOps team, we love our scripts. We use them to automatically detect cost anomalies across our 100+ microservices, so we don't have to manually monitor each one. We also generate weekly and monthly summaries of cloud and AI costs to keep engineering leadership and Finance informed.

This reduces the amount of manual calculations Finance has to perform, gives us visibility into trends vs forecasts, and makes it easy for busy stakeholders to quickly understand what's changing, why, and where we're headed.

Recommendation

Automate as much of your FinOps work as possible. Your future self (and your finance team) will thank you.

In closing

FinOps isn't about saying "no" to spending. It's about making sure every dollar drives value for our learners. Every dollar we save means more resources to support our mission. If you're an engineer who's ready to spend some of Duolingo's money (responsibly) to make high-quality education available to everyone, we're hiring!