Hi, On Mon, Jun 22, 2015 at 10:57 PM Alexander Klimetschek <aklim...@adobe.com> wrote:
> On 15.06.2015, at 02:23, Carsten Ziegeler <cziege...@apache.org> wrote: > > > > It really seems that people who are not convinced have never felt the > > current pain - while people who are on the pro side exactly felt this > > pain and ran into the problems which this is trying to solve. I'm > > absolutely unsure on how to solve that situation. > > I was asking this before: what are the pains and specific use cases? > > (Apart from the paging of results) > Apologies for not tracking this discussion, but I wanted to weigh in before things got much further. IIUC, the core problem we are trying to solve is to provide a query syntax indepdent of any particular ResourceResolver implementation. While, to be honest, this is not a problem I have personally run into using Sling for the past 6 years, I can certainly see why it is one. But I do think we have a good answer available which was Alex's original proposal to have Adobe donate the QueryBuilder code to Sling. Now the QueryBuilder code as-is wouldn't solve this problem; it would require a refactoring, but I believe this refactoring is managable. This would have the following benefits: 1) Adopt a syntax many (but certainly not all) Sling developers are famililar with. 2) Provide a path to avoid YAQL. While yes, in the near term we will have "Sling QueryBuilder" and "AEM QueryBuilder", the AEM QueryBuilder could be deprecated (obviously up to AEM Product Management) and eventually removed. 3) An opportunity to fix some of the issues with QueryBuilder (granted, this isn't necessarily Sling's problem to solve). One thing which concerns me about the current Query API is that it appears to be completely non-extensible. How, for example, would one implement something like https://docs.adobe.com/docs/en/cq/5-6-1/javadoc/com/day/cq/search/eval/RelativeDateRangePredicateEvaluator.html ? If I'm reading this correctly, the date math has to be done by the caller. Which isn't that problematic at first, but the code would be significantly more verbose than relativedaterange.property=jcr:lastModified relativedaterange.lowerBound=-1d What is potentially problematic about not having this type of extensibility is that it prevents specific implementations from providing the best implementation possible. For example, let's say that MongoDB has a really efficient way to query for documents modified in the last day. If I do the date math in Java code, I'm making it that much harder for the MongoDB ResourceProvider to opimitize this query (sorry, this isn't a great example, but it's late and I'm getting tired). Plus, the query isn't really expressing what I want -- I want to find resources modified in the last day, not from some absolute date. So someone reading my code later has to figure out what the calls to Calendar.add(Calendar.DAY_OF_MONTH, -1) are there for. Here's a better example: JCR is unable to compare two properties, i.e. give me all nodes where property foo equals the value of property bar. But MongoDB *can* do this (it isn't super-efficient, but it is possible). I can almost see how you would do this with the new Query API, but it would be ugly at best. Or, more broadly, how would the MongoDB $where operator be supported? The advantage of the AEM QueryBuilder's model is that figuring all of this stuff out isn't the responsibility of the platform developer. We just need to provide a solid basis and then let downstream users add their own hooks. As soon as you say that these are the only 8 operations anyone is ever going to do on a property or the 4 operations anyone is ever going to do on a resource, you're into "640k should be enough memory for anyone" territory. So how specifically would the Sling QueryBuilder be different than the AEM QueryBuilder? I think of QueryBuilder queries being processed in these separate steps (FWIW, none of this is proprietary information, it is based on public documentation): 1) A map of key/value pairs is turned into a PredicateGroup object. While technically this step is optional (you can build a PredicateGroup by hand), it is pretty common. This would be common functionality across all ResourceResolvers and the code from AEM could probably be brought over as-is. 2) The PredicateGroup (which is a nested tree) at this point represents the query statement. It is then passed to the ResourceResolver (this part is somewhat different than the AEM QueryBuilder). 3) Each ResourceProvider analyzes the predicates and decides whether or not it knows how to evaluate all of them. If it can't, it should return no results (this is debatable, but I think it makes sense). The only exception is where you had an or clause, i.e. this query: fulltext=Management group.p.or=true group.1_jcrType=dam:Asset group.2_resourceType=some/resource/type If a non-JCR provider didn't know how to evaluate the jcrType predicate type, it could still evaluate the query because it is OR'd with a resourceType predicate (which let's say it does know how to evaluate). But if it didn't know how to evaluate the fulltext predicate type, it shouldn't return any results. 4) The ResourceProvider uses PredicateEvaluators to map each predicate to its native query syntax. For this to work, each ResourceProvider would expose its own PredicateEvaluator interface (in theory, a ResourceProvider doesn't need to do this if the evaluation process isn't intended to be pluggable). IIOW, the current AEM PredicateEvaluator interface would be renamed JcrPredicateEvaluator. 5) At least in JCR (based on current functionality), some Predicates can't be evaluated in a native query (i.e. XPath) and will need to be handled as filters on the result set, but this is an implementation detail left to the ResourceProvider. 6) The ResourceProvider returns results to the ResourceResolver. 7) Sorting is handled (or not) as currently proposed. To be clear, I don't have a concrete proposal for how to replicate (or not) AEM QueryBuilder's facet support. Alex might... Regards, Justin > > Cheers, > Alex >