FYI I am working with Example protos for model input, and I am learning that use tf.parse_example to parse a (shuffled) batch of serialized examples is much faster than using tf.parse_single_example prior to batching. For my particular dataset, using parse_single_example allows me to create feed_dicts with batch size 128 at about 100/min; batching the serialized Example protos and then using parse_example is running at around 3000/min.

You may want to update the documentation to suggest using tf.parse_example everywhere, as is suggested when using sparse input data.