ietf
[Top] [All Lists]

Re: [apps-discuss] Last Call: <draft-ietf-appsawg-json-pointer-07.txt> (JSON Pointer) to Proposed Standard

2013-01-08 11:53:11
I'm definitely not trying to expand the scope beyond the current draft.
More trying to share some experience with an existing implementation of
something similar.

A few comments on path syntax based on this experience.

1) It seems like some of the examples raised on this thread conflate JSON
documents with Javascript objects. According to the json spec, an object
always starts with an "{" and ends with a "}". Therefore, while the json
path "/1" could refer to elements in { "1": "a" } and [1,2], the latter is
not a correct JSON object.

2) Should a JSON pointer be able to refer to the "whole document"? In the
spec, it says that the path "" would refer to the whole JSON document. I
don't think this should be possible. If you want to replace the entire
document, you should be using an HTTP PUT to replace the content, rather
than a patch. Eliminating the ability to refer to the whole document would
also release some ambiguity in certain cases.

For example, earlier it was pointed out that if i have this document

{ "a": 1 }

And then apply this patch

{ op: "add", path: "", value: [1,2,3] }

then it would result in the document after patch being

[1,2,3]

however, this is not a valid json document as it is not enclosed by { }.

Instead, I think that the empty path should refer to the element of the
document named with an empty string.

So with { a: 1 }, and the op { op: "add", path: "", value: [1,2,3] }
would result in this document:

{ a: 1, "": [1,2,3] }

This also means that the leading slash is somewhat superfluous and could be
omitted from the syntax (since the paths "", and "/" would both resolve to
the "" named element of the document. This makes parsing paths easier as
well, since you can just split the path on the "/" and take the count of
parts to figure out how deep you're going.

Given { a: 1, "": [1,2,3] }, i would refer to the 2nd array element with
the path "/2". When parsing this i'd split on "/", which gives me two
tokens, the empty string (on the left of hte /) and the number 2 (on the
right of the /).

-j

On Tue, Jan 8, 2013 at 8:19 AM, Robert Sayre <sayrer(_at_)gmail(_dot_)com> 
wrote:

Hi Jared,

Earlier in the thread, I actually directly asked whether software that
operates on arbitrary JSON was in scope for this WG (my example was
CouchDB), after having suggested either changing the path syntax or
renaming the array operations.[0]

The editors didn't respond, except with process points or simple
contradictions without rationale. My conclusion is that software such
as MongoDB must be out of scope here. To their credit, the editors did
point out that anyone is free to try again, and I plan to do just
that. There's no reason to hold up this WG, since they seem to be
burnt out.

- Rob

[0]
http://www.ietf.org/mail-archive/web/apps-discuss/current/msg08552.html


On Tue, Jan 8, 2013 at 12:26 AM, Jared Rosoff <jsr(_at_)10gen(_dot_)com> 
wrote:
hi team, i'm new to the discussion here, but wanted to jump in. i work on
mongodb, a json database, and i wanted to share how we deal with these
issues.

mongodb uses almost the same notation for pointers ("a.b.c" instead of
"/a/b/c"). We also index arrays in the same way as json pointer ( "a.0"
refers to the 0th element of the array called "a"). and this works fine
in
practice. (ref:
http://docs.mongodb.org/manual/core/document/#dot-notation)

our update syntax is different tho. the verbs in mongodb updates for json
documents are more specific:

set / unset / rename (operations on fields)
inc (increment integer values)
push / pop / pull (operations on arrays)
addToSet / removeFromSet (operations on arrays)

(ref http://docs.mongodb.org/manual/reference/operators/#update)

since update operations are more specific and type dependent, it's easy
to
throw an error if an unexpected type is encountered (e.g. try to push
onto a
field that has a non-array value) and to act smartly on empty fields ( if
path to push is empty, we assume it should be an array, create it, and
then
push the value onto it).

i concur the the pointer syntax is fine and ambiguity comes from the
definition of operators in json patch.

-j


On Mon, Jan 7, 2013 at 10:26 PM, Matthew Morley <matt(_at_)mpcm(_dot_)com> 
wrote:

On Mon, Jan 7, 2013 at 5:39 PM, Conal Tuohy 
<conal(_dot_)tuohy(_at_)versi(_dot_)edu(_dot_)au>
wrote:

On 07/01/13 13:23, Matthew Morley wrote:


For me the deficiency is not in the pointer, but patch format being
generated.

One approach is to push that *one* test, structure conformity, into the
pointer syntax. Another is via the type operation.

If a vague patch is generated, vague results are to be expected.

It seems to me, on the contrary, that the deficiency is in the pointer
syntax, and I think it would be a mistake to try to work around that
deficiency in JSON Patch. Because aren't there other things which one
might
do with JSON Pointer than use it with JSON Patch? There's been mention
of
having it registered as a URI fragment identifier syntax for JSON for
example. JSON Pointers could then end up all over the place, outside of
patches. IMHO JSON Pointer needs to be taken seriously as a technology
in
its own right.


Couldn't agree more about it being taken seriously in its own right. :)

JSON Pointer for me exists outside of JSON Patch, always has and will do
the way we think about structures. As it represents both a resolution
path
and an identity string (both ends of the path concept). I see value
from the
identity view, in describing a location that is aware of being inside an
array.

But JSON Pointer should not be changed just because of issues with JSON
Patch, especially when JSON Patch is attempting to address those issues
with
other mechanisms within the specification. That is all I was trying to
express. The syntax change should be for other reasons, if it is going
to be
made.

My personal experience (for what its worth): In the past I've tried a
number of syntaxes like JSON Pointer. Mostly a.b.c.0 and even a.b.c:0 at
times to address the same issues suggested here. Though my experiences
pushed me towards a single syntax using a.b.c.0, and thus my support for
/a/b/c/0 over /a/b/c:0.

The system at first used the . or : syntax, combined with dynamic
tokens,
being pointers themselves, to resolve other pointers. So it was not
reasonable to know ahead of time if an end point was an array or an
object.
"a.b.c.{d.e.f}" could end up in an array or in an object, depending on
the
value at d.e.f at the time of resolution. Especially with many layers of
tokens to resolve, and changing data structures.

I found in practice, it didn't really matter, so the choice of . or :
was
phased out. At the end of the day the two syntaxes point to mutually
exclusive points within the data, so that `meta data` about the
structure
was removed from the syntax we used. It didn't add value, even if it
added
clarity at times. We also had functions at the end of paths, but that
goes
beyond the JSON focus of the JSON Pointer goals, so those points are not
relevant here.

This discussion thread seems to be getting overly complicated, but JSON
Pointer changes should come from the JSON Pointer view point and that
specifications goals, not from short comings in JSON Patch.

--
Matthew P. C. Morley
_______________________________________________
apps-discuss mailing list
apps-discuss(_at_)ietf(_dot_)org
https://www.ietf.org/mailman/listinfo/apps-discuss



_______________________________________________
apps-discuss mailing list
apps-discuss(_at_)ietf(_dot_)org
https://www.ietf.org/mailman/listinfo/apps-discuss


<Prev in Thread] Current Thread [Next in Thread>