Le mar. 1 déc. 2020 à 14:40, Jonathan Gallimore < jonathan.gallim...@gmail.com> a écrit :
> I'll address a few points inline below, but at a high level, what are we > looking to achieve from a spec/tck challenge? > > I can see a case for some clarification and updates to the Javadoc. > > The assertions that /- will return an error (as that references an index to > append to after the *end* of an array - i.e. array.length) are tested in > the TCK, and other implementations must be passing that TCK. It's hard to > see a spec change happening, as there is no spec document beyond the RFCs > that I can find. A TCK change that would enable Johnzon to pass, and > require other currently passing implementations to make a change seems > unlikely. Jakarta EE 8's TCK has been around a while and has > implementations that pass. The Jakarta EE 9 TCK is basically "done" and is > essentially the same as EE8, bar the namespace change. I guess adding a > test exclude is possible, but serves to make this more vague and vendor > dependent (and non-portable) which feels like it defeats the purpose - > surely having it better defined and tested is the way to go. > Well, here the fact is that it does not impact other vendors since it is a johnzon vendor specific feature we put in a shadow of the (javax/jakarta) spec handling in a custom fashion an error case. Typically the case where we can exclude the TCK since it is irrelevant for our impl but I understand also it is not perfect. > > I appreciate that this introduces a backwards incompatible change, and that > there may be other consumers of the library that would have an issue if > this just changed. This seems like a fairly straightforward case that could > be easily and quickly solved with a feature switch, and passing the TCK is > a worthwhile goal, both for Johnzon and TomEE. I suspect the TCK challenge > will take a bit of time, and we'll likely end up back at the feature switch > anyway. > Issue is we dont have a Json.createPointerFactory(mapWithToggle) so it is a global flag which means it breaks some deployments anyway - at least at tomee level - when > 1 app is deployed (or >= 1 app + 1 extension). > > On Fri, Nov 27, 2020 at 10:44 AM Romain Manni-Bucau <rmannibu...@gmail.com > > > wrote: > > > Hi JL, > > > > As discussed together - but sharing for others - we must take into > account > > some points: > > > > 1. reading both spec, JSON-Patch enables to handle /- as your first did > (ie > > consider it is last element). JSON-Patch uses JSON-Pointer but nowhere it > > is written it behaves as JSON-Pointer in all cases and it is typically > > "integration" definition which can extend an underlying spec (otherwise > > most of EE wouldn't be right? ;)) > > > > I think the idea is that it references a non-existent element, *after* the > last element in an array. So if you have an array [0, 1, 2, 3, 4], then > "/-" would reference element _5_ (assuming you start your numbering at 0), > and not the last element in the array (index 4). > This is the jsonpointer spec right, but JSONPatch never requires to not handle the case as we do, it is just not written (and why we used it also). Issue on jsonpointer side being we can't have another character which means "last element". > > > > 2. On johnzon point of view we can't break this feature which was > requested > > by user and transitive users (ie user of products built with johnzon) > > without at least a clear migration path so if we want to break we should > do > > a 1.3 (dont think we need a 1.2 maintenance branch, we can do it lazily), > > document how to migrate from current behavior to new one (i'll detail it > > after) and communicate on it on our website properly (index.html ref and > > dedicated page I guess with the release annoucement). Alternative is to > > challenge the TCK, it is a failure case so it is typically the kind of > case > > we can plug custom/vendor behavior (we do in other parts of the JSON-B > spec > > for ex). Overall idea is to not let users on the road because some TCK > > exist (functional and users over procedural work). > > > > I'd be interested in the history, it helps to be mindful of it when making > changes. > Goal is to be able to work on the last element, there is nothing in specs about this one but it is very common to need that (see it as "length" operator). Indeed we can enrich jsonlogic module to cover that case but most users just bring jsonp+jsonb and not johnzon-jsonlogic. > > > > > > On strict TCK side, we can also do a johnzon-tck module where we wrap the > > provider to handle that case and pass the TCK, this is purely technical > to > > be compliant but would avoid to break anything. > > Now if we really want to be strict in our implementation we must still > > enable this last element case. One option not far from what we have is to > > use our json-logic module and add some jsonpatch operators. Combining > > multiple operators we can manage to fulfill this common patching need - > but > > we break the overall API + require a new module to be added to apps). > > > > Lastly I would note that JSON Pointer *enables* our impl: > > > > > Note that the use of the "-" character to index an array will always > > > > result in such an error condition because by definition it refers to > > a nonexistent array element. Thus, applications of JSON Pointer need > > to specify how that character is to be handled, if it is to be > > useful. > > > > > > > For example, some applications might stop pointer processing upon an > > > > error, while others may attempt to recover from missing values by > > inserting default ones. > > > > > > Literally means "this is a case we consider as an error but your > > application can recover from it" and we do ;). > > > > Sort of. "applications of JSON Pointer need to specify how that character > is to be handled". What's the definition of "application of JSON pointer"? > In the case of TomEE, I'd suggest the "application" is Jakarta EE, which > has specified that an error should be thrown. In a standalone case, is the > application whatever is consuming Johnzon, or Johnzon itself? > Well TCK define it but not the JSON-P spec and I'm more than happy to request to drop that TCK since it was completely passed under the radar - guess TCK were never really reviewed). Also note that the JsonPointer javadoc - since there is no pdf or spec document - does not mention it must implement the RFC but only that it must respect its syntax and part of its constraints. So really it was under the radar more than anything and we must not assume this TCK was intended originally when JsonPointer class was created IMHO. > > > > Since it is an error case I would start by challenging the TCK to make it > > vendor dependent and exclude it from the passing list for now. > > If really blocking we can go with plan B and try to have a migration path > > but it sounds like a lot of effort for everyone for literally 0 gain > IMHO. > > > > Personally, I'd prefer a switch that enables us to comply with the Jakarta > EE spec behaviour, rather than introducing something vendor specific and > non-portable into the spec. > We would have the factory I would be for that but since we don't I see way more pitfalls than advantages - except passing the TCK there is none actually but in OSGi env or multiapp containers it would be a real pain :( - so if it is the solution you want (and I fully get it is the fastest to pass TCK which is likely current goal) then maybe just wrap JsonProvider with a custom TCKJsonPointer(johnzonProvider.createJsonPointer(...)) and validate the tck case. It is quite trivial to do in tomee-tck setup and will give you a "not risky" flagging. Don't get me wrong, I'm not very happy of that but I just don't want we drop an used feature for a test which is not needed at spec level. Alternative I discussed with JL was to provide a clear migration path and adding some jsonlogic operator to fill the gap, it is also very doable, only point I'm not sure is that adding a module will match other users expectations. Assuming that yes we can do a 1.2.9 keeping this feature (we have some fixes we should let go out before any breaking change), do the changes in jsonlogic module (we can do them before too since it is only additions) and do a 1.3.0 fully compliant with the documentation updated. Small variation of this option is to have our own SPI for JsonPointer factory this way it can be overriden for TCK and we can also keep our impl, I'm less a fan of this one since it will bring a proprietary import in portable code in a nasty way instead of splitting it properly (like jsonlogic module option does). Hope it makes sense. > > Jon >
