Rappi Improves Efficiency, Stability, and Speed with Amazon DocumentDB (with MongoDB compatibility)

Rappi’s microservices architecture required multiple databases suited for different jobs. The company turned to Elasticsearch for catalog and search and set up a series of microservices organized by business unit and by country, with each country supported by one Elasticsearch cluster. But because every microservice within a country was using the same cluster, if something went wrong with any single microservice—if it established too many connections to the database or wrong index usage, for example—the resulting blast radius could create performance issues ranging from service degradation to a complete outage of every other microservice in that country. Further, because Rappi was deploying one Elasticsearch cluster per country to cover all nine lines of business in that country, the model used up to 100 nodes per cluster, which drove up the price tag, slowed down performance, and required constant maintenance. In other words, the model simply did not scale the way Rappi needed it to. “We were using really huge instances, and we weren’t getting the performance we wanted,” says Marcelo Soto, Rappi head of data platform.

These early scaling and performance problems generated frequent outages for one of the most profitable services at Rappi: its restaurant deliveries. The most egregious of these was a 3- to 4-hour outage during peak service time caused by increased requests on the Elasticsearch cluster. Meanwhile, the ranking query could take up to 500 ms, a latency that slowed down performance and affected the experience of thousands of users.

As it searched for a new database solution with the goal of drastically improving performance, Rappi’s engineers determined that part of its microservices workload needed to be supported by a document store. Rappi had already taken a first step by moving these workloads from Elasticsearch into document store solutions and was working on upgrading its small MongoDB platform. Now, the Rappi team needed to eliminate the undifferentiated heavy lifting of maintaining infrastructure for so many MongoDB clusters. Compatibility with MongoDB was a factor. “Compatibility was crucial because we needed to move microservices that were already working,” says Soto. “We had to be able to move to a different database without changing much code.” Rappi found the solution it was looking for in a fully managed document database service from AWS that provided the MongoDB compatibility it needed: Amazon DocumentDB (with MongoDB compatibility), a nonrelational database designed to give customers the performance, scalability, and availability needed to operate mission critical MongoDB workloads at scale.

Now, with the migration to Amazon DocumentDB complete, “we have 2,000 different databases in production, with DocumentDB being the second most prominent,” says Soto. Rappi’s new model allots one Amazon DocumentDB cluster per each of its nine lines of business per country, and it can manage the exponential increase in clusters with just two engineers. “We had to make small adjustments to transition from MongoDB to Amazon DocumentDB, but after we did, we found that with a lot less code, we could accomplish the same tasks with a lot more stability,” continues Soto. Rappi’s largest Amazon DocumentDB cluster of five instances, which took 2 weeks to migrate, resulted in a significant improvement in performance and efficiency. In fact, Rappi has not experienced any Amazon DocumentDB–related outages, and ranking-query latency has dropped to 80 ms in the 99th percentile. “Now,” says Soto, “engineers have maybe 5 percent of the problems they had before, and if they have problems, it’s a problem with the logic or on the application side, not because Amazon DocumentDB is doing something wrong.” Along with providing better performance, the managed Amazon DocumentDB database requires substantially less maintenance and firefighting, which frees up Rappi’s data platform team to focus on value-adding activities that help move the app forward.