The startling development of deepfake generation techniques has raised serious issues about the veracity and reliability of the digital media, making reliable detection mechanisms a top priority research demand. Although many deepfake detection approaches have been proposed, there is a lack of structured evaluations regarding the performance of different algorithms, data sets, and evaluation measures in real world conditions. This paper is a systematic review of deepfake detection research and classifies current methodologies as spatial-based, temporal/spatiotemporal-based, or frequency/audiovisual-based detection paradigms. We also investigate widely used benchmark datasets, evaluation metrics and experimental protocols for reviewing the status of how people are doing their works. Our review uncovers crucial challenges such as poor cross-dataset generalisation, biased training data, susceptibility to compression and low-quality media, sensitivity to adversarial manipulations and scalability issues in deployment. By identifying insights, shortcomings and emergent patterns in the existing works, we aim to draw conclusions about the open research spaces and future directions towards resilient and deployable deepfake detection solutions. The objective of this review is to provide a consolidated guideline that could be used as an anchor for the researchers in developing future generalisable and robust deepfake detection solutions.