Death by a thousand cuts. Let’s take JSON parsing as an example. In the simple case a value like “foo” requires no copying during parsing - it can be parsed wi

Death by a thousand cuts. Let’s take JSON parsing as an example. In the simple case a value like “foo” requires no copying during parsing - it can be parsed with zero copies. But escaped values like “foo\”bar\”” cause issues. The actual string you want is foo”bar” without the escape slashes. Ok, so for this you need to copy the unescaped string into a scratch buffer somewhere. How big is this buffer? Is the JSON parsing API you’re using flexible enough to let you re-use buffers between calls to avoid extra allocations? Cool. Now how do you express “this parsed struct field may reference the input bytes or it may reference some scratch space” in a way that’s memory safe? Does your chosen language even support this? What if your internal string representation is platform specific? Copying uses potentially double the memory but whatever, we are hampered by calling conventions and legacy code and other libraries and opaque unknown callers in different languages and different operating systems and different CPUs. And on top of this mountain of shit you stand and yell “I just want to parse some JSON!”. Screw it. Just always re-allocate and make copies Now imagine this times by a thousand at every layer of the stack and in every process and every library.