How Blitz scaled their game coaching app with lower latency and leaner operations How Blitz scaled their game coaching app with lower latency and leaner operations Blitz is a fast-growing startup that provides personalized coaching for games such as League of Legends, Valorant, and Fortnite. They aim to help gamers become League of Legends legends through real-time insights and post-match analysis. Blitz While players play, the app does quite a lot of work. It captures live match data, analyzes it quickly, and uses it for real-time game screen overlays plus personalized post-game coaching. The guidance is based on each player’s current and historic game activity, as well as data collected across billions of matches involving hundreds of millions of users. Thanks to growing awareness of Blitz’s popular stats and game-coaching app, their steadily increasing user base pushed their original Postgres- and Elixir-based architecture to its limits. This blog post explains how they recently overhauled their League of Legends data backend – using Rust and ScyllaDB. TL;DR – In order to provide low latency, high availability, and horizontal scalability to their growing user base, they ultimately: TL;DR Migrated backend services from Elixir to Rust. Replaced Postgres with ScyllaDB Cloud. Heavily reduced their Redis footprint. Removed their Riak cluster. Replaced queue processing with realtime processing. Consolidated infrastructure from over a hundred cores of microservices to four n4‑standard‑4 Google Cloud nodes (plus a small Redis instance for edge caching) Migrated backend services from Elixir to Rust. Replaced Postgres with ScyllaDB Cloud. Heavily reduced their Redis footprint. Removed their Riak cluster. Replaced queue processing with realtime processing. Consolidated infrastructure from over a hundred cores of microservices to four n4‑standard‑4 Google Cloud nodes (plus a small Redis instance for edge caching) As an added bonus, these changes ended up cutting Blitz’s infrastructure costs and reducing the database burden on their engineering staff. Blitz Background As Naveed Khan (Head of Engineering at Blitz) explained, “We collect a lot of data from game publishers and during gameplay. For example, if you’re playing League of Legends, we use Riot’s API to pull match data, and if you install our app we also monitor gameplay in real time. All of this data is stored in our transactional database for initial processing, and most of it eventually ends up in our data lake.” Scaling Past Postgres One key part of Blitz’s system is the Playstyles API, which analyzes pre-game data for both teammates and opponents. This intensive process evaluates up to 20 matches per player and runs nine separate times per game (once for each player in the match). The team strategically refactored and consolidated numerous microservices to improve performance. But the data volume remained intense. According to Brian Morin (Principal Backend Engineer at Blitz), “Finding a database solution capable of handling this query volume was critical.” They originally used Postgres, which served them well early on. However, as their write-heavy workloads scaled, the operational complexity and costs on Google Cloud grew significantly. Moreover, scaling Postgres became quite complex. Naveed shared, “We tried all sorts of things to scale. We built multiple services around Postgres to get the scale we needed: a Redis cluster, a Riak cluster, and Elixir Oban queues that occasionally overflowed. Queue management became a big task.” To stay ahead of the game, they needed to move on. As startups scale, they often switch from “just use Postgres” to “just use NoSQL.” Fittingly, the Blitz team considered moving to MongoDB, but eventually ruled it out. “We had lots of MongoDB experience in the team and some of us really liked it. However, our workload is very write-heavy, with thousands of concurrent players generating a constant stream of data. MongoDB uses a single-writer architecture, so scaling writes means vertically scaling one node.” In other words, MongoDB’s primary-secondary architecture would create a bottleneck for their specific workload and anticipated growth. MongoDB’s primary-secondary architecture They then decided to move forward with RocksDB because of its low latency and cost considerations. Tests showed that it would meet their latency needs, so they performed the required data (re)modeling and migrated a few smaller games over from Postgres to RocksDB. However, they ultimately decided against RocksDB due to scale and high availability concerns. “Based on available data from our testing, it was clear RocksDB wouldn’t be able to handle the load of our bigger games – and we couldn’t risk vertically scaling a single instance, and then having that one instance go down,” Naveed explained. Why ScyllaDB One of their backend engineers suggested ScyllaDB, so they reached out and ran a proof of concept. They were primarily looking for a solution that can handle the write throughput, scales horizontally, and provides high availability.They tested it on their own hardware first, then moved to ScyllaDB Cloud. Per Naveed, “The cost was pretty close to self-hosting, and we got full management for free, so it was a no-brainer. We now have a significantly reduced Redis cluster, plus we got rid of the Riak cluster and Oban queues dependencies. Just write to ScyllaDB and it all just works. The amount of time we spend on infrastructure management has significantly decreased.” Performance-wise, the shift met their goal of leveling up the user experience … and also simplified life for their engineering teams. Brian added, “ScyllaDB proved exceptional, delivering robust performance with capacity to spare after optimization. Our League product peaks at around 5k ops/sec with the cluster reporting under 20% load. Our biggest constraint has been disk usage, which we’ve rolled out multiple updates to mitigate. The new system can now often return results immediately instead of relying on cached data, providing more up-to-date information on other players and even identifying frequent teammates. The results of this migration have been impressive: over a hundred cores of microservices have been replaced by just four n4-standard-4 nodes and a minimal Redis instance for caching. Additionally, a 3xn2-highmem ScyllaDB cluster has effectively replaced the previous relational database infrastructure that required significant computing resources.” High-Level Architecture of Blitz Server with Rust and ScyllaDB High-Level Architecture of Blitz Server with Rust and ScyllaDB Rewriting Elixir Services into Rust As part of a major backend overhaul, the Blitz team began rethinking their entire infrastructure – beyond the previously described shift from Postgres to the high-performance and distributed ScyllaDB. Alongside this database migration, they also chose to sunset their Elixir-based services in favor of a more modern language. After careful evaluation, Rust emerged as the clear choice. “Elixir is great and it served its purpose well,” explained Naveed. “But we wanted to move toward something with broader adoption and a stronger systems-level ecosystem. Rust proved to be a robust and future-proof alternative.” Now that the first batch of Rust rewritten services are in production, Naveed and team aren’t looking back: “Rust is fantastic. It’s fast, and the compiler forces you to write memory-safe code upfront instead of debugging garbage-collection issues later. Performance is comparable to C, and the talent pool is also much larger compared to Elixir.”
