Delivery Hero Moves Most Users to HTTP/3, Cuts Home-Screen Response Time 17%

At the P99 2026 virtual conference, Delivery Hero staff engineer Elvis Chidera said the company moved about 95 percent of users to HTTP/3 and cut its apps' home-screen end-to-end response time by 17 percent. The change also reduced handshake latency by 51 percent, while some devices fell back to HTTP/2 where UDP traffic was blocked.

Why Delivery Hero Needed HTTP/3

Delivery Hero has an estimated 60 million users across Europe, Latin America, Africa, the Middle East, and Asia, with apps such as Foodpanda, Yemeksepeti, Glovo, and PedidosYa. Uber is currently acquiring Delivery Hero for $14.8 billion, with plans to unite it with Uber Eats. Network connectivity varies widely, and last-mile networking was often the bottleneck. In Pakistan, apps could take nearly a second at P90 to establish a secure handshake with a server.

The company looked at HTTP/3 as a way to lower that latency. HTTP/3 uses QUIC over UDP instead of TCP, and can use 0-RTT to start sending eligible application data while the handshake is still underway. It also avoids head-of-line blocking by sending packets in multiple streams.

Rollout and Android Tuning

Because Delivery Hero uses Cloudflare's content delivery network, moving to HTTP/3 involved a simple reconfiguration. Apple has supported HTTP/3 since iOS 15, and most native browsers already support it. Android was trickier: engineers used Google's Cronet library, initially through the Cronet Transport for OkHttp wrapper, to support HTTP/3 while minimizing code changes.

Cronet's engine took about 800 milliseconds to initialize at P90, which ate into some of HTTP/3's speed benefits. The team forked the code into its own bridge using OKCronet and overlapped initialization with other background tasks. Overall, P95 latency fell from 1.2 seconds to 120 milliseconds, a 10x improvement in read latency.

Remaining Packet-Loss Issue

Devices that could not use HTTP/3, largely because networks blocked UDP traffic, automatically fell back to HTTP/2. Pakistan's response time was cut in half. However, secondary screens in some markets prone to packet loss, such as Pakistan, experienced latency issues.

Delivery Hero attributed those issues to Cloudflare's use of the Cubic congestion control algorithm, which has not handled packet loss well. Tests with Google HTTP/3-aware servers using BBR did not get hung up on packet loss. The company is working with Cloudflare to resolve the issue.

Sources: The Register