Educational cloud Integrated Development Environments (IDEs) must provide language-aware editing support for many concurrent student sessions while remaining operationally manageable. In the current deployment of EduIDE, an Eclipse Theia-based cloud IDE tailored for education at the Technical University of Munich, Java language intelligence is embedded directly into each session container. While this preserves local IDE compatibility, it replicates resource-intensive Java Virtual Machines (JVMs) and language-server state across many otherwise identical workspaces. This redundancy can increase memory overhead and limits cluster scalability during peak usage. Consequently, this thesis investigates whether Java language intelligence can be decoupled from the IDE session and managed as an external infrastructure component.
After showing that a shared-JVM approach for Eclipse JDT is structurally infeasible, this work implements a generalized sidecar architecture for EduIDE. The resulting system extends the Kubernetes operator with sidecar-aware provisioning and supplies runtime endpoint metadata to Eclipse Theia. An experimental validation demonstrates that extraction is technically viable but adds a measurable cold-path penalty. Under persistent storage, median startup rises from 25.8s to 31.4s, median first-error latency from 10.6s to 15.9s, and median warm diagnostic latency from 678ms to 1071ms. Steady-state hover latency, however, remains nearly identical at 839ms versus 838ms. The main benefit of the sidecar design therefore lies in architectural decoupling and lifecycle control rather than in uniformly lower latency.