Software as a Service (SaaS) has been increasingly adopted by software vendors as their main software delivery model, as it provides the opportunity to offer their software applications to a larger market and to benefit from the economies of scale. One of the key enablers to leverage economies of scale is multi-tenancy: resources are shared among multiple customer organizations (the so-called tenants), which leads to higher resource utilization and scalability. The highest degree of resource sharing is achieved with application-level multi-tenancy. However, this focus on increased resource sharing typically results in a one-size-fits-all approach. As a consequence, multi-tenant SaaS applications are inherently limited in terms of flexibility and variability, and cannot be customized to the different and varying requirements of the different tenants.This dissertation presents both a middleware framework and a software engineering method to facilitate the development, operation and management of customizable, multi-tenant SaaS applications. More specifically, the middleware framework improves the flexibility of multi-tenant SaaS applications by enabling tenant-specific customizations, while preserving the economies of scale and limiting the application engineering complexity. The focus is on dynamically composing software variants on a per tenant basis as well as on enforcing tenant-specific performance service level agreements (SLAs) throughout the SaaS application. The service line engineering (SLE) method aims to reduce the management complexity of many co-existing tenant-specific configurations as well as the effort to provision tenants and to update and maintain the customizable SaaS application.This work has been validated and evaluated in the context of two types of industry-relevant SaaS applications, i.e. a request-driven online hotel booking application and a batch-driven document processing application. We have implemented different prototypes on top of existing cloud platforms and the evaluation shows the effectiveness of our solution while introducing only a very limited performance overhead.