Is there any reason to use MCC these days? What's the motivation of using MCC these days? Isn't it obsolete? How is it better than ROC? 
I still see people use MCC. Are they too lazy to use a newer method or does MCC give something different? 
 A: @0x90 Lazy? Ah ah ah... you must be kidding ;-)
Matthews correlation coefficient (MCC) obsolete? Well, it was introduced in 1975, while the ROC curves were introduced during the II World War...
Both MCC and ROC curves are not obsolete; but if you want to label "old" something, you should choose the ROC curves.
By the way, let's discuss the motivations of prefering MCC.
If you have a confusion matrix threshold:
If you have a supervised learning binary classification problem, and you have a threshold to use in the confusion matrix to discriminate false positives from true negatives, and false negatives from true positives, then you should use a threhsold-specific rate, such as MCC, F1 score, accuracy, true positive rate (also know as: sensitivity, or recall), true negative rate (also known as: specificity), and other ones.
As I explained in my paper "Ten quick tips for machine learning in computational biology" (Tip 8), among all these rates, the advantage of the MCC is that it generates a high score only if your model is able to predict a high percentage of true positives and a high percentage of true negatives, on any balanced or imbalanced dataset.
The other rates (F1 score, accuracy, true positive rate, true negative rate, precision, etc) do NOT have this feature. As I explained in my paper , they can lead to misleading results.
If you do NOT have a confusion matrix threshold:
In cases when you don't have a specific threshold, you need to compute confusion matrices for all the thresholds. You can analyze these confusion matrices through two main tools: the receiver operating characteristic (ROC) curve, and the precision-recall (PR) curve.
You can then compute the area under the curve (AUC) of both.
As I explained in my paper "Ten quick tips for machine learning in computational biology" (Tip 8), you should calculate both, but then focus on the precision-recall (PR) curve.
EDIT 2020-Apr-07: Regarding this topic, I recently published on BMC Genomics a paper entitled "The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation", that I wrote with my collaborator Giuseppe Jurman. I hope this article can help you understand why using the MCC is the best option in binary classification evaluation.
A: Don't really know in which context they are using MCC as a metric for evaluation. But in classification (and some  regression) tasks MCC has shown a pretty decent performance especially while working with imbalanced data since it takes into account classes proportion in confusion matrix. 
So I would say it depends on what are you working on, but for imbalanced problems MCC could be more informative than classic metrics such as accuracy, or F1-score.
Take a look at this review of metrics in classification tasks.
https://biodatamining.biomedcentral.com/articles/10.1186/s13040-017-0155-3
