One of the hard tasks in NLP is answering factoid questions in an open-domain setting. In this paper we consider this problem, by using Wikipedia as unique knowledge source which requires deep understanding of text. Machine comprehension systems cannot solve this problem alone, so our solution integrates search, distant supervision and multitask learning. For this task of machine reading at scale (MRS) we developed DrQA, a strong system for question answering from Wikipedia composed of two parts. First part is Document Retriever, document retrieval system to first narrow our search space and focus on reading only articles that are likely to be relevant for given question, by using bigram hashing and TF-IDF matching. Those articles are then processed by second part of our system - Document Reader, a multi-layer recurrent neural network machine comprehension model trained to detect answer spans in those relevant returned documents. Experiments and results of evaluting system using multiple benchmarks show efficacy of our approach and indicate that MRS is key challenging task for researchers to focus on