The emergence of the Function-as-a-Service (FaaS) paradigm enables software developers to focus on business logic all the while entrusting the service provider in a serverless environment to elastically scale up during a burst of requests or scale down when request rates are low. One of the key enablers of FaaS is the widespread adoption of containerization technology such as Docker, which packages software into images that can be loaded and executed in any other computer running the Docker engine. An encumbering limitation of FaaS is significant overhead when downloading these images to newly instantiated workers from a centralized image repository. Large bursts of requests cause new workers to be instantiated and lead to so-called "cold-start" latencies. Recent works have investigated distributed image registries in the flavor of downloading images from existing workers and the scheduling of functions with knowledge of the server resource availability and function request information separately. However, the scheduling of functions in servers distant from existing images can lead to sub-optimal function completion times, especially in the context of a bursty workload. FaaSFabric addresses these issues by incorporating the information of image locations into the scheduling decisions. In this paper, we present fine-grained overhead breakdowns of workload traces, a Mixed Integer Linear Program (MILP) formulation of incorporating distributed image registry information in request allocation decisions, a justification of the potential of Deep Reinforcement Learning (DRL), and a proposed DRL formulation. We also demonstrate that in a simple workload, image pulling can take up to 80% of the total function completion time on average, which shows the potential for a distributed image registry-aware scheduler to outperform baseline and state-of-the-art schedulers.