Only a small fraction of workloads, about 5%, can be classified as embarrassingly parallel, such as searching for specific phrases across millions of documents. These tasks are ideally suited for Mapreduce and the traditional Hadoop stack. It's crucial for enthusiasts to ensure their workloads fit this category to achieve optimal performance.
