Boost logo

Boost :

From: Andrzej Krzemienski (akrzemi1_at_[hidden])
Date: 2020-09-17 18:10:21

Hi Everyone,
I recommend to *ACCEPT* the library into Boost. It meets the standard for
BI am definitely going to use it.

I tried to use the library in two configurations in header-only mode:
g++ 9.2.0 with -std=c++2a on MinGW on Windows,
g++ 10.2.1 on Fedora with -DBOOST_JSON_STANDALONE and -std=c++20
All worked fine.

I studied library docs and played with examples for about 30 hours. I am
not particularly knowledgeable about the problem domain. But my programs
have to consume and produce JSON files.


It is very good, and meets the highest Boost standards. I mean both the
tutorial and reference sections. While doing this review I had to look at
the documentation of the competing libraries, and they are really bad. This
is why I can appreciate the big effort you put in making these docs. There
is a section "Comparison" that compares Boost.JSON to other JSON libraries.
I personally think it should also mention the documentation.

Some observations:

1. Add reference to RFC 7159 ( It
allows implementations to set limits on the range and precision of numbers

2. The documentation for array object and index mention "Satisfies the
requirements of ContiguousContainer, ReversibleContainer, and
SequenceContainer" These requirements are never defined or referenced. The
user will not even know if you defined them or if they are "commonly
known". Maybe add link to

3. String has two sections "Member Functions":

4. Documentation for value::operator= seems corrupt:
* it has sections with no content
* It is an unconstrained template, it look like it can accept any type

5. Examples of input JSON-s contain a lot of escape characters in string
literals, like in
This makes it difficult to read. Consider using raw string literals.
Instead of
assert( serialize( jv ) == "{\"x\":4,\"y\":1,\"z\":4}" );
We would have
assert( serialize( jv ) == R"({"x":4,"y":1,"z":4})" );

6. Some functions (like operator== or swap(x, y)) are naturally part of
class interface, even though they are not member functions. It is annoying
and confusing that rather than being listed along with the class they
belong to they are instead listed separately as "free functions"

The docs are not uniform about this issue. For `json::value` operator== is
listed as friends along with other member functions, but not 2-argument

7. In we read:
" Given the following definition of customer::customer( const value& )
[...] Objects of type customer can be converted *to* and from value".

This "to" is surprising here. How can a constructor converting from `value`
to `customer` could be used to convert `customer` to `value`?

8. As discussed in another thread, documentation would benefit from the
info on implementation-limits-related design choices.

a. Any json::value that you can build can be serialized and then
deserialized, and you are guaranteed that the resulting json::value will be
equal to the original.

b. JSON inputs where number values cannot be represented losslessly in
uint64_t, int64_t and double, may render different values when parsed and
then serialized back, and for extremely big number values can even fail to

c. Whatever JSON output you can produce with this library, we guarantee it
can be passed by any common JSON implementation (probably also based on
uint64_t+int64_t+double implementation.


Clear, sound and practical. The design decisions correspond with my
expectations of a library: fast, easy to use and learn, with value
semantics. The interface is very intuitive, I practically do not have to
consult the documentation: every functionality (serialization, value
alteration) has the interface that I would expect. I appreciate that it
targets embedded environments.

Some observations:

1. The handling of infinities and NaN's needs to be addressed. Currently, I
can do:
and t produces an invalid JSON. You cannot easily make a precondition on
json::value assignment/construction because you expose the double as a
mutable member:

void fun(double x)
  // precondition x is not infinity or nan
  json::value v = x; // ok
  v.as_double() *= 2.0; // this may produce infinity

So maybe you need a precondition in the serializer.

2. Boost.JSON provides its own string type. This begs a lot of design
questions for `json::string`. Is it a better replacement for std::string?
std::string provides the following capability:

void set(std::string& s, char v, char x, char z, char k)
   if (cond)
     s = {v, k};
     s = {v, x, z, k};

json::string does not. How do I do a similar thing with json::string?

3. string::swap() has a strange precondition: you cannot swap an object
with itself

This is a really unexpected and potentially dangerous, and at minimum
requires a very strong rationale. This would be the first swap function in
std or Boost that has a precondition. While discussing a somewhat related
issue (on the guarantees of the moved-from state) in the LEWG, someone
remarked that there was at least one STL implementation that did a
self-swap for convenience of implementation.

Is it really necessary?


I didn't look at it much. Just some quick points I observed.

1. Values of type `double` representing Infinity and NaN can be stored in
json::value and when later serialized, they render a malformed JSON
output. json::value should either validate special values of `double` or
indicate as precondition that they should not be assigned. Alternatively,
the serializer should either validate this or indicate as precondition.

2. tag_invoke() takes tag arguments by reference, does it have to? Passing
by reference requires the constexpr variables to be instantiated and their
address assigned.

3. Customization through ADL-discovered overloads can cause ODR violations.
There is no superior alternative though. But maybe put a warning in the
docs that the user must make sure that all translation units must see the
same set of overloads.

4. I do not like the 2-state-switch mechanics of BOOST_JSON_STANDALONE. It
assumes that a user either wants a no-Boost version or that she wants a
package: all Boost components.

But what if I am using C++14 and I want to use std::error_code and
Boost-emulated memoty_resource and string_view? What If I use Boost version
and C++17 and want to use std::string_view?

Maybe you can additionally provide macros for turning Boost-emulation
libraries one-by-one? I mean, one for string_view, one for error_code, and
one for memory_resource.

5. The header only mode -- I really appreciate it. It is good for initial

Finally, I would like to thank Vinnie and Krystian for writing and sharing
this library.


Boost list run by bdawes at, gregod at, cpdaniel at, john at