Batch learning is the tried and tested method for machine learning. In many ways this makes sense; it allows data scientists to apply the scientific method when training models because data sets are static and well understood. The emergence of Python libraries such as scikit-learn have also contributed to thinking about modeling in terms of reproducible batches. It is also a common practice to train on a dataset that is as large as possible in order to get the best model performance.
However, there are drawbacks to batch learning. First and foremost, training on large datasets incurs higher costs on compute and storage. Second, with the batch learning approach, retraining and redeploying models with new features can only be done periodically and results in a lag between the availability of the features and making inferences from those features. To further complicate matters, in today's fast paced world, new features are available within minutes, if not seconds or microseconds. This results in batch learning models suffering from drift as soon as they are deployed and retraining them on a periodic basis does little to help reduce this problem.
It is due to these limitations that online learning models are beginning to gain in popularity. Online learning models are smaller in size and do not require a lot of data to train on. In fact, online learning models can learn on new features as soon as they arrive so they are able to quickly react to changing data. This means that they are less susceptible to drift. They also require less compute and storage because learning is incremental as opposed to batch learning models where you have to accumulate a lot of data in order to train the model. While online learning models are less susceptible to drift, it is also much easier to put monitoring in place to catch drift as soon as it occurs. In the batch learning scenario, monitoring, catching, and reacting to drift is a lot more manual and time consuming.
Here are some resources for you to get started: