Powering one of Japan’s leading dating apps on a foundation built to last

Tapple is one of Japan’s leading dating apps. A subsidiary of CyberAgent, a major Japanese technology company, Tapple serves 400,000–500,000 active daily users, making service reliability a critical engineering priority.

Tapple’s engineering team relies heavily on AI-assisted development workflows. As a result, the team’s pull request throughput has more than doubled year over year. That development speed also introduced new operational challenges. The platform spans frontend, backend, and infrastructure layers, creating dependencies across multiple systems and services. One major initiative added further operational complexity: migrating an 11-year-old JavaScript monolith to TypeScript microservices on the backend. With multiple features migrating each month, the team needed to minimize disruptions.

Managing those transitions made releases difficult to roll back and offered no way to gradually expose changes to specific user groups. For a team shipping at this pace, the process created operational bottlenecks. “We were shipping feature migrations every month, but our deployment tooling hadn’t kept up,” says Kaito Yamagishi, Engineering Manager. “We needed a way to release changes incrementally and roll them back quickly if issues appeared.”

At the same time, AI-powered features and the LLMs behind them introduced additional production dependencies. Seasonal spikes in user activity, particularly around holidays, increased pressure on systems that the company’s existing monitoring and deployment approaches struggled to manage. Tapple needed a solution that could keep pace with its release velocity while giving teams greater control over production rollouts.

“We needed a way to release changes incrementally and roll them back quickly if issues appeared.”

Safer, faster releases with Datadog Feature Flags

Tapple was already using Datadog before adopting Feature Flags. The team had standardized on Application Performance Monitoring (APM), Log Management, and Infrastructure Monitoring across its stack.

After its previous feature management platform announced a significant price increase, Tapple consolidated feature management onto Datadog. The team had also standardized on OpenFeature, which simplified the migration process.

Previously, deployment and release were treated as a single event that was difficult to reverse. Datadog Feature Flags separated deployment from release, giving the team control over when and how features are activated in production independently of code deployments. “Having everything in one place—flags, traces, monitors, logs—means we can release a feature and immediately check the user flow and any errors, all without switching tools,” says Yamagishi.

Tapple put Feature Flags at the center of its JavaScript-to-TypeScript microservices migration. Feature Flags live in the backend-for-frontend (BFF) layer, gating traffic between the legacy monolith and the new microservices. As each migration becomes ready, traffic shifts gradually through canary rollouts, with an instant rollback available if issues arise.

One example is Tapple’s like-or-dislike workflow, which is one of the most critical interactions in the app and can reach 250 req/sec at peak times. Without Feature Flags, the migration would have required an all-or-nothing release with no easy path to rollback. Instead, the team shipped incrementally, monitored the rollout in real time, and could revert instantly if needed. The same pattern repeated across dozens of migrations.

Shipping AI features with the same control as everything else

Tapple runs two AI-powered features in production: an initial message generator that helps users write a first message to a match, and a profile description generator that turns a user’s prompt answers into a polished bio. Both are built on the Vercel AI SDK, routing requests through Microsoft Azure OpenAI across multiple microservices. Because model behavior is harder to predict than traditional code, the team brings the same release discipline to its AI features that it uses for migrations—gating each feature behind a flag, exposing it to a small segment of users first, and keeping an instant off-switch within reach if responses drift or an upstream dependency degrades.

“When something goes wrong across that many microservices, it helps us immediately understand whether the problem is on our side or Azure's.”

That control is only as valuable as the visibility behind it. Tapple uses Datadog Agent Observability alongside APM to maintain visibility across the full request path, helping engineers identify where problems originate across distributed AI systems. During an early production incident, Azure OpenAI hit a rate limit. Error rates appeared first in Agent Observability, while APM revealed the specific errors that the Vercel SDK was encountering. In another incident, the model began leaking system prompt content into responses. Using Agent Observability, the team identified which users had been affected, traced the failure, and added a guard in both the application logic and the system prompt.

Together, the two form a tight loop: Agent Observability surfaces when and why an AI feature is misbehaving, and Feature Flags give the team the ability to contain it in seconds. “When something goes wrong across that many microservices, it helps us immediately understand whether the problem is on our side or Azure’s,” says Yamagishi.

A decade of backend modernization, shipped safely and on time

With Datadog Feature Flags, Tapple completed the migration of its 11-year-old JavaScript monolith to TypeScript microservices on the backend without any user-visible issues. At its peak velocity, the team was shipping four to five feature migrations per week. Across roughly 100 in-flight migration flags, all have now been fully cut over.

Previously, a full redeployment took up to 30 minutes. With Datadog Feature Flags, a rollback is a 10-second flag flip. That translates to roughly 50 hours of deployment time saved over the three-month period of rapid migration—around four hours per week. And because detection-to-rollback time is dramatically faster, the number of users exposed to any given issue before it is resolved is a fraction of what it once was. A 5% canary rollout with Feature Flags limits the impact of an issue to just 5% of users.

Looking ahead, Tapple plans to expand its use of feature-flag-driven releases. The team is moving toward daily, flag-gated deployments, replacing its weekly release cycle with a CI/CD pipeline built around Feature Flags from the start.