Diagnosing and Solving a 100 ms Latency Issue in Spring Boot's Embedded Tomcat Using Arthas

The author describes a Spring Boot channel system that suffered a mysterious ~100 ms extra latency between the recorded processing time (≈150 ms) and the client‑observed response time (≈250 ms). Initial checks confirmed the network path (ping to Nginx and from Nginx to the service) was fine, but a direct curl to http://127.0.0.1:7744/send still took ~70 ms on the first request and only a few milliseconds on subsequent requests.

Because the issue could not be reproduced locally, the author turned to Arthas , Alibaba's Java diagnostic tool. Using trace on Spring MVC entry points and then on org.apache.coyote.http11.Http11Processor.service, they discovered a 129 ms hotspot inside Tomcat.

Further tracing of

org.apache.catalina.webresources.TomcatJarInputStream.createZipEntry

revealed repeated loading of JAR entries, especially the swagger‑ui resources. The watch command confirmed that Swagger’s JAR was being accessed on each request, causing the delay.

Removing the Swagger dependencies eliminated the latency, indicating a bug in the Tomcat‑embed version (2.0.2.RELEASE of Spring Boot bundles Tomcat 8.5.31). The fix is to upgrade Tomcat‑embed to 8.5.40+ or upgrade Spring Boot to a newer release where the bug is patched.

Practical solutions include:

Update tomcat.version in the Maven properties to 8.5.40 or higher.

Or upgrade the Spring Boot parent version to 2.1.0.RELEASE or later, which brings a newer Tomcat.

After the upgrade, the 100 ms latency disappears, confirming the root cause was the Tomcat‑embed bug triggered by Swagger resource loading.