With the rapid spread of social media networks, such as Twitter and Facebook, in Arabic societies, it leads to the explosive growth of Arabic posts, reviews, comments, or tweets. Each one of these generates a huge volume of digital opinionated data on different topics such as politics, economics, societies marketing, and businesses. Analyzing valuable subjective information from opinionated data would assist in a better understanding, making decisions, and predicting global issues and events. Therefore, sentiment analysis coincides with social media networks and has become the most interesting research field in performing the analysis process and detecting sentiment polarities to extracted opinionated social-media data. However, there are several challenges faced the sentiment analysis process, especially with Arabic social data. Sentiment analysis of Arabic social media is indeed in its infantile stage and it has not obtained thoroughly attention wherein several challenging issues still need to address. Some of these challenges result from the complexity of Arabic natural language (e.g., complex morphological and lack of lexicon lists and datasets) and other challenges result from social media platform itself (e.g., slang words and colloquial terms). In this manuscript, we first study the impact of social media challenges on the current challenges of Arabic natural language. Our findings show that such challenges add more complexities to the Arabic sentiment analysis process. Based on these findings, we embark to review the emerged and contributed proposals, which give rise on analyzing opinionated data, extracted from Arabic social media networks. Our review methodology is based on a set of criteria, which we propose to assess and highlight the advantages and limitations of these proposals. The interesting point here is to help researchers identify the social sentiment analysis problems along with a comprehensive survey on the sentiment analysis levels (document, sentence, and aspect levels) and classification approaches (supervised, unsupervised, and hybrid approaches). Finally, we compare these contributed proposals in terms of the average accuracy and suggest a new hybrid approach based on our findings.