Today's organizations collect millions of logs a day containing potentially malicious attacks or machine failures that can easily go undetected. In order to analyze these logs, they must be parsed into a format which only contains relevant information. Traditionally, logs are parsed using complex heuristics and regular expressions, which can be time consuming and hard to maintain as log formats constantly change through time. At Splunk, we believe that we can alleviate these issues by utilizing machine learning techniques to adapt to the shifting data with high accuracy. We explore three different machine learning approaches to log parsing: conditional random fields (CRFs), long-short-term memory neural networks (LSTMs), and two transformer models (BERT and miniBERT). Our experiments on open source log datasets target four key requirements: accuracy, robustness, data efficiency, and performance. From these benchmarks we show that while the shallow LSTM is fast, it lacks robustness while the graph based CRF and the transformers are slower but able to adapt to different domains. Furthermore, since transformers are pre-trained on natural language, we are able to get an accurate and robust model with significantly less data, and with all the optimizations coming out for transformers there is a possibility that transformers can be the best of all worlds.