Cloud instances are usually virtual machines hosted on shared hardware. Containers are often used to deploy services in cloud instances. However, excessive consumption of shared hardware resources by some VMs may lead to unpredictability in the performance of containers running in co-located VMs. Existing techniques to detect performance interference in applications are either too expensive in terms of profiling or applicable only from the perspective of the infrastructure owner. In this paper, we propose Sherlock, a lightweight subscriber-centric mechanism to detect performance interference and estimate its impact on cloud services. Sherlock does not require access to hardware counters and can work on unmodified clouds without any support from the cloud provider or changes to the hypervisor. Sherlock uses a simple profiling technique which is performed only for a short duration before deployment. When interference from co-resident VMs is observed, Sherlock notifies the cloud subscriber, so that any remedial actions can be taken. We also define a metric IScore, which is an estimate of the impact of interference on a service. Experiments on the real-world web benchmark CloudSuite show that our approach is able to detect interference with accuracy ranging from 89% to 98.4%, and with very less false positives (< 8%). © 2017 IEEE.