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. 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. 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). > 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. > > 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? > 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. Jon
