"Speed tables reads a structure definition and emits C code to create and manipulate tables of rows of that structure. This allows operations to be performed on the table at native code speeds, using a custom extension that manages rows of fields as native C structs and emit subroutines for manipulating those rows in an efficient manner."
FlightAware.com deals with extremely high-volume data throughput, and they manage the whole process with Tcl.
I usually frame the data as clientData passed to a managing command (i.e. custom built first-class command), with appropriate allocation/deallocation. As often as not, I'm creating two commands: an anonymous reference template, and a command-creating-command that will instantiate instances of the "anonymous reference template". Then I can create an arbitrary number of arbitrarily named commands that pack and process their data, including freeing on destruction.
To the uninitiated, this sounds like object-oriented programming, which is arguably true, but doesn't automatically come with subclassing, inheritance, etc, etc. Tcl does have first class objects, but what I'm describing here isn't them.
I'm a Tcl fan, but the above is actually somewhat problematic, because to hold a reference to the C struct, you have to then manually free it later, losing some of the benefits of a language with automatic memory management.
Real dicts, and the new object stuff are big improvements though.