In the sterile world of theoretical ML, launching a recommendation system begins with six months of Feature Store design and the construction of monumental data pipelines. In reality, the high season is looming on your calendar, you have only two engineers and one data scientist at your disposal, and the time for ramp-up ran out yesterday. As shown by the case of Tutu MLOps engineer Kristina Kucherova, in these conditions, it is not the "correct" architecture that survives, but the one that manages to generate revenue by the deadline. The team consciously chose technological skepticism, replacing heavy platforms with a streamlined stack of Kafka, MongoDB, and ClickHouse.
Strategic context
TravelTech specifics do not forgive slowness: there is no point in offering "abstractly similar" hotels. If a user has bought a ticket to Sochi for specific dates, the recommendation must account for this context instantly. Collaborative filtering in a vacuum is useless here. To avoid drowning in the design of a full-scale Feature Store (like Feast), the engineers turned MongoDB into an operational feature storage. Real-time events from Kafka and nightly aggregates from ClickHouse fly into a single user document on the fly. The result: recommendation delivery in milliseconds without unnecessary infrastructure entities.
Key points
Speed over perfection: The team prioritized a working product over a complex architectural ideal to meet seasonal demand. Tactical tech stack: Kafka, MongoDB, and ClickHouse provided the necessary performance without the overhead of specialized ML platforms. Ruthless prioritization: Features like "cold start" handling and abandoned cart scenarios were cut from the MVP to ensure a timely launch. Scalability: Despite the rapid build, the system remained extensible, later incorporating session aggregates and new ranking models while keeping the original API.
True MLOps is not about implementing every trendy tool from the CNCF landscape, but the ability to cut away the excess for the sake of immediate ROI.
The bottom line
When business metrics are tied to peak calendar dates, a working solution built on familiar tools is far more valuable than a flawless architecture that arrives too late. The Tutu case proves that a lean approach to MLOps allows a small team to move from hypothesis to live production traffic in just over a month, securing profits when they matter most.