For machine learning, you almost definitely want to use sklearn.OneHotEncoder. For other tasks like simple analyses, you might be able to use pd.get_dummies, which is a bit more convenient.
Note that sklearn.OneHotEncoder has been updated in the latest version so that it does accept strings for categorical variables, as well as integers.
The crux of it is that the sklearn encoder creates a function which persists and can then be applied to new data sets which use the same categorical variables, with consistent results.
from sklearn.preprocessing import OneHotEncoder
# Create the encoder.
encoder = OneHotEncoder(handle_unknown="ignore")
encoder.fit(X_train) # Assume for simplicity all features are categorical.
# Apply the encoder.
X_train = encoder.transform(X_train)
X_test = encoder.transform(X_test)
Note how we apply the same encoder we created via X_train to the new data set X_test.
Consider what happens if X_test contains different levels than X_train for one of its variables. For example, let’s say X_train["color"] contains only "red" and "green", but in addition to those, X_test["color"] sometimes contains "blue".
If we use pd.get_dummies, X_test will end up with an additional "color_blue" column which X_train doesn’t have, and the inconsistency will probably break our code later on, especially if we are feeding X_test to an sklearn model which we trained on X_train.
And if we want to process the data like this in production, where we’re receiving a single example at a time, pd.get_dummies won’t be of use.
With sklearn.OneHotEncoder on the other hand, once we’ve created the encoder, we can reuse it to produce the same output every time, with columns only for "red" and "green". And we can explicitly control what happens when it encounters the new level "blue": if we think that’s impossible, then we can tell it to throw an error with handle_unknown="error"; otherwise we can tell it to continue and simply set the red and green columns to 0, with handle_unknown="ignore".