How Manbang Migrated from Eureka to MSE Nacos for Scalable, High‑Performance Service Discovery

Business Growth Challenges Service Stability

Manbang Group, a leading "Internet+Logistics" platform, serves millions of truck drivers and shippers, processing over 128.3 million orders in 2021 with a total transaction value of CNY 262.3 billion. Rapid growth put extreme pressure on its self‑built service‑registry infrastructure.

Self‑Built Service Gateway and Registry

The company operates a custom microservice gateway that handles north‑south traffic, security, and governance. Requests first hit an ALB, then the internal gateway, which routes to backend services via a service‑registry cluster built on open‑source Eureka and ZooKeeper.

Scaling Issues with Eureka and ZooKeeper

When the number of registered instances exceeded 2,000, Eureka nodes struggled to synchronize registration data, leading to node crashes and service outages. ZooKeeper suffered frequent full GC pauses, lacked authentication, and its high CPU and memory usage threatened overall stability.

Cloud Migration to Alibaba Cloud MSE Nacos & ZooKeeper

To address these problems, Manbang adopted Alibaba Cloud MSE Nacos and MSE ZooKeeper, replacing the original Eureka and ZooKeeper clusters while preserving compatibility with existing Eureka clients.

Compatibility with Eureka Protocol

MSE Nacos fully implements the native Eureka protocol; the data model of Eureka InstanceInfo is mapped one‑to‑one to Nacos Service and Instance objects. Consequently, client code does not need any changes—only the endpoint configuration is switched to the MSE Nacos address.

MSE‑Sync Data Migration Tool

The MSE‑Sync utility (based on an enhanced Nacos‑Sync) performs bi‑directional synchronization, pulling existing Eureka registration data into the new Nacos cluster and keeping incremental updates in sync. After a successful data check, the Eureka endpoint is replaced and the services run on MSE Nacos.

Breaking the Eureka Performance Bottleneck

Eureka’s broadcast replication model creates a write‑load that grows linearly with instance count. MSE Nacos uses a custom AP‑model Distro protocol that hashes service instances and assigns each to a responsible node, drastically reducing synchronization traffic even at massive scale.

High‑Availability and Push‑Empty Protection

Nacos offers push‑empty protection: if the registry cannot provide a valid provider list, consumers receive a safe fallback instead of an empty list, preventing cascading failures. The same protection is exposed to Eureka‑compatible clients.

Incremental Pull for Eureka Clients

Originally, Eureka clients performed full‑registry pulls periodically, consuming up to 40 MB/s of outbound bandwidth. MSE Nacos introduced an incremental pull mechanism—clients fetch the full snapshot once, then receive only delta updates—reducing bandwidth to ~200 KB/s and eliminating gateway‑level GC spikes.

Server‑Side Performance Optimizations

Cache full and incremental registration data; use hash‑based change detection to cut CPU work.

Replace SpringBoot's default StringHttpMessageConverter with an enhanced version for large payloads.

Enable HTTP chunked responses for registration queries.

Adapt Tomcat thread‑pool size dynamically based on container limits.

Results and Metrics

After the upgrades, CPU utilization dropped from 13 % to 2 %, registration read latency fell from 55 ms to under 3 ms, YGC count decreased from >10 to 1, and YGC pause time reduced from 125 ms to under 10 ms. The system now handles 2 million QPS with stable GC behavior.

Post‑Migration Reflections

Manbang’s seamless migration demonstrates that a well‑planned cloud‑native service‑registry upgrade can eliminate legacy bottlenecks, improve stability, and provide a scalable foundation for future business growth.