Your friend suggests you should randomize the order of your data before splitting into a training and test dataset. Is your friend right?

No$,$ your friend is not right. There is no benefit to doing this, and the detail just makes the process too complex.

No$,$ your friend is not right. You should make sure your work is always repeatable, and there is no way to get repeatability if you randomly reorder the dataset.

Yes, your friend is right. It is possible your data was originally organized in such a way that splitting will give the model a biased view of the dataset $($imagine for example if all the data was ordered by some feature$).$