Wiki has a good description of HLS (which may explain the startup delay):
How HLS Works
At a high level, HLS works like all adaptive streaming technologies; you create multiple files for distribution to the player, which can adaptively change streams to optimize the playback experience. As an HTTP-based technology, no streaming server is required, so all the switching logic resides on the player.
To distribute to HLS clients, you encode the source into multiple files at different data rates and divide them into short chunks, usually between 5-10 seconds long. These are loaded onto an HTTP server along with a text-based manifest file with a .M3U8 extension that directs the player to additional manifest files for each of the encoded streams.
See diagrams - http://www.streamingmedia.com/Articles/Editorial/What-Is-.../What-is-HLS-(HTTP-Live-Streaming)-78221.aspx
Roku and HLS are discussed here - http://sdkdocs.roku.com/display/sdkdoc/Video+Encoding+Guidelines
As indicated by Tablo in another thread, the interaction between Tablo and Roku is a pull model which falls in line with the HLS model: