• A chunk is a multi-dimensional rectangle (or "tile"). When a variable is chunked, each chunk is read or written as a single I/O operation, and individually passed from stage to stage of the pipeline and filters.
• Think of how Google Maps accesses map data, as rectangular tiles rather than thin rows of pixels along latitude lines. To access a subset of the data, you only read the subset of tiles ("chunks") containing the data you want, rather than extracting data from long rows along latitude lines.
• Contiguous layout is the default for fixed sized variables without compression or other filters.
• Chunking is required for any variable that uses one or more unlimited dimensions, compression, or other filter (such as check-summing).
• Chunking is set when the variable is created, and is never changed after that. To change the chunksize you have to copy the data into a new file.
• Chunking is transparent to the reader (unlike with Google Maps).

2012 Unidata NetCDF Workshop > Chunking and Deflating Data with NetCDF-4