Reaching new levels of in-game engagement during play
Second Dinner uses Databricks to optimize in-game experiences
faster feature launches, from 10 per year to more than 50
faster model development and training
Second Dinner, the studio behind Marvel Snap, a competitive, fast-paced card battle game based on the popular Marvel Universe, relies on a small team and collaboration for rapid feature iterations. Second Dinner significantly improved the personalization of their shop experience by centralizing on the Databricks Data Intelligence Platform, progressing from manual processes with general options to one-to-one, real-time recommendations. With faster data processing, analytic insights and ML capabilities, Second Dinner continues expanding use cases for in-game engagement.
Building a successful deck
Second Dinner values their small team for its close collaboration and rapid game development, but this also limits their capacity for big projects. Despite utilizing the Databricks Data Intelligence Platform for various functions — ingesting, processing, warehousing, analytics, dashboarding and ML — they faced challenges in advancing player personalization for their daily shop experience.
Initially, Second Dinner relied on manual, heuristic-driven processes to collect player data and generate dashboards on Databricks. While they saw potential in using this data for shop personalization, the journey from concept to production was slow and cumbersome.
Ted Li, Associate Director of AI at Second Dinner, noted, “The manual steps needed to interpret visualizations and continuously push code created numerous opportunities for errors. A single feature iteration could take a month, limiting us to 10 use cases yearly. Personalization lagged behind player interest, so they were no longer desired by the time cards were available. It was based on aggregate statistics rather than true personalization.”
To meet player demand in real time and maximize monetization, Second Dinner needed to evolve their personalization system without losing the excitement of a randomly generated shop, which is crucial for card collection. The team aimed to shorten the recommendation update loop by iterating with the advanced tools available on Databricks.
Unifying data workflows from ingest to model with Databricks
“We can’t just hire more people, so we need technologies that can do the work for us — that’s where Databricks comes in. It does everything we need for effective personalization,” Li said. Second Dinner leveraged the Databricks Data Intelligence Platform to transition the recommendation update loop from manual heuristics to one-to-one, real-time recommendations. Marvel Snap generates revenue by selling cosmetic variants attached to cards. The “Daily Offer Shop” presents players with eight new card slots daily, from thousands of variants and 30 to 100 daily card releases. Without personalization, players could wait hundreds of days to find a specific card, negatively impacting their experience and bottom-line revenue for the game. Second Dinner drives monetization by offering cards based on players’ preferences.
Second Dinner optimized the feedback loop in and out of the game to speed up insights. Instead of manually reviewing dashboard data, the team built an ML model in Databricks using MLflow. The teams build and ingest customer engagement data through Delta Lake to feed their models. Using PyFunk, a light wrapper from MLflow, the data team can easily add logic via Python while enabling more data users, regardless of language preferences. Now, team members can deploy their own cadences without aligning with client or server uploads. The model adapts to player behavior and automatically updates in-game.
The first model generated daily population-level recommendations. After a month of iterations, Second Dinner developed models for batch pre-computed personalization and real-time recommendations. By removing manual interventions and inference, pre-calculated shop recommendations are now globally deployed using a low-latency, model-serving endpoint.
Predicting player preferences and delivering in-game personalization
Second Dinner has achieved significant results by enhancing shop and in-game personalization, helping to better predict player preferences in real time. “In less than a week, we set up a relatively personalized system for different segments of players worldwide, updating daily. If a card is popular one week and not the next, the model dynamically adjusts,” Li said.
Machine learning has allowed Second Dinner to go beyond simply analyzing purchases. Their models now integrate data on what players see, when they update or play cards and their card preferences. After training the model on Databricks, the team uploads it to the model registry and retrains it weekly. This setup allows Second Dinner to predict and generate personalized recommendations for every active player globally, based on their unique in-game behaviors.
With the capabilities unlocked by the Databricks Data Intelligence Platform, Second Dinner has also begun using time-sensitive triggers to boost in-game engagement. The game can now identify opportunities for real-time interactions and trigger potential next-best actions based on recent player engagement. For example, new card bundles are personalized using recent player activities that indicate card preferences. Each bundle contains at least one card that the player will want, increasing the likelihood of purchase.
Li concluded, saying, “Leveraging these new capabilities in Databricks, Second Dinner has shifted from launching only 10 experiments a year to running them weekly. Our dashboarding and reporting processes are scaling as the data team continues to expand use cases, optimizing opportunities and enhancing player engagement.”
Learn more about Databricks Data Intelligence Platform for Games
