It is a Fastly Compute
application that retrieves fixed-sized blocks of an object from a backend, and
then it assembles a response using the blocks. Depending on the size of the
object and on which part of the object the client requests (Range in the
request), the response may be assembled from one or more blocks and some blocks
may be included in their entirety while the extremities may use partial blocks.
The blocks are individually cacheable. Blocks have a fixed size (except for the last block in the object), and their position in the object starts at a multiple of the block size.
As Fastly compute does not yet have direct programmatic access to caching, this
application sits in front of a VCL service, which handles caching. The VCL
service sets req.enable_segmented_caching to true in vcl_recv.
For best results the block size used by the VCL service in the back should match the block size used by the compute service in the front. The Segmented Caching feature in VCL then always assembles its response from a single block. A description of Segmented Caching in our VCL offering is here. The default block size is 1 MiB.
Unlike the Segmented Caching implementation in the Fastly Varnish, this application retrieves blocks in parallel (5 at a time by default). That gives it higher throughput, especially for cache misses, at the expense of higher CPU utilization overall and higher origin load for misses.