The data templating language

C++ Implementation

The engine that does the hard work of evaluating Jsonnet programs is written in C++11 (no dependencies) and is exposed by a C API. It implements a handwritten lexer and recursive descent parser. The desugaring is performed by the parser, so the rest of the implementation operates on only the core subset of the language. Then, the ASTs are analysed to perform static checking, and to establish the set of unbound variables in each expression. This optimizes the capturing of the environment that is needed for closures, object construction, and creation of thunks. Finally, the analysed AST is executed in the virtual machine.

The Jsonnet virtual machine has its own stack and garbage collected heap. The garbage collector is a simple mark and sweep garbage collector which is triggered when the heap grows (number of objects) by a certain factor since the last GC cycle. This factor can be tuned, as can the number of objects below which the GC does not trigger.

The Jsonnet stack is not bound by the size of the native stack, it is implemented as an explicit stack in the VM. Tail call optimizations are performed.

Some of the standard library functions are implemented in C++ as builtin functions, e.g. the math routines and reflection. However others are implemented in Jsonnet in the std.jsonnet file. This file is embedded into the library when it is built. The intention is that ports of the Jsonnet virtual machine (e.g. to Java, .NET, etc.) should only need to reimplement the builtin functions. The other parts of the standard library can be implemented by re-using the Jsonnet implementations from the C++ implementation.


Currently supported are Linux and OSX. However the code should be easy to compile on any system with a modern C++ compiler, as it has no dependencies. The build has been tested on gcc and clang.