Underdog stress-tests stability for real-time sports gaming during traffic peaks

Challenge

Durability and availability during major sporting events

‍Underdog is a sports gaming platform that serves fantasy sports games and prediction markets to users across 49 states, Washington DC, and Canada. As a fast-growing app with more than 5 million users who expect real-time information and speedy transactions, Underdog needs a reliable data streaming infrastructure that operates without issue during major sporting events.

The Underdog team wanted to stress-test a streaming platform to ensure it could handle peak traffic loads during events like the Super Bowl. They wanted to simulate real conditions with rollout constraints, probes, and benchmark traffic to ensure the app’s streaming backbone would remain stable.

Underdog needed a streaming platform that would support Super Bowl–level workloads with low latency, high fan-out, bursty demand, order correctness, and recovery behavior.

Why Redpanda 

Fast, consistent, and durable streaming

While searching for an Apache Kafka^®–compatible streaming infrastructure, Underdog’s software vendor recommended Redpanda due to its ease of use and ability to deploy quickly.

Underdog chose to build and battle-test a Redpanda production environment on Kubernetes for order streaming. Redpanda offered critical capabilities for the stress-testing scenario, including:

• Durability for local NVMe with a hybrid storage model
• Earned availability during restarts
• Sustained high throughput 
• Rollout backpressure
• Correctness under node loss

After modifying Redpanda’s startup replication behavior and failure model, Underdog ran benchmarks that showed healthy client metrics and no obvious bottlenecks. 

“I think Redpanda is the best-of-breed experience for a Kafka self-hosted solution compared to open source, community versions, etc.," says Regis Wilson, Staff Platform Engineer at Underdog.

‍Result 

High performance, reliable streaming, and bottleneck-free 

Using Redpanda’s highly efficient streaming platform, Underdog sustained 32,000 messages per second at 4 kibibytes with P99 latencies below 0.07 milliseconds. The P99 latencies were actually lower than the team’s original p50 targets.

“The latency profile looked far better than we expected for the initial box size. Later, we saw around 32k msg/s territory with client-side metrics still looking healthier than the doom-driven part of the engineering brain would prefer. We were expecting death, destruction, failure, and scrambling for backup solutions. Instead, we found yawns and boredom,” Regis says.

During stress testing, Underdog pinpointed and resolved a failure model issue and switched to the RF=5 availability model to avoid data loss and improve quorum behavior. Intentionally rolling, deleting, and breaking the system also allowed Underdog to validate lifecycle behaviors that would support their prediction markets exchange.

Thanks to Redpanda’s reliable streaming infrastructure, Underdog can sustain peak loads during high-traffic events like the Super Bowl and have peace of mind with real-time streaming—allowing the team to focus on other features for their users. 

“I think the fact that Redpanda worked sufficiently well (up to our highest limits) initially and without a lot of problems definitely gives us one more thing to check off and not worry about as we prepare for launch,” says Regis.