[infinispan-dev] new Infinispan Query API - ISPN-194

[Emmanuel Bernard]

On 5 avr. 2011, at 13:38, Sanne Grinovero wrote:

> 2011/4/5 Emmanuel Bernard <emmanuel at hibernate.org>:
>> 
>> On 5 avr. 2011, at 12:20, Galder Zamarreño wrote:
>> 
>>> 
>>> On Apr 4, 2011, at 6:23 PM, Sanne Grinovero wrote:
>>> 
>>>> </snip>
>>>> 
>>>> there's one catch:
>>>> when searching for a class type, it will only include results from
>>>> known subtypes. The targeted type is automatically added to the known
>>>> classes, but eventually existing subtypes are not discovered.
>>>> 
>>>> Bringing this issue to an extreme, if the query is not targeting any
>>>> type, and no indexed types where added to the grid (even if some exist
>>>> already as they might have been inserted by other JVMs or previous
>>>> runs), all queries will return no results.
>>>> How to solve this?
>>>> - class scanning?
>>> 
>>> Nope, too expensive.
>>> 
>>>> - explicitly list indexed entities in Infinispan configuration?
>>> 
>>> No
>>> 
>>>> - a metadata cache maintaining a distributed&stored copy of known types
>>> 
>>> That sounds more appealing. It could be a good middle ground until Search can search for types.
>> 
>> Do you have any specific idea in mind?
>> 
>> To magically find types:
>>  - we scan every file system, databases, caches available to the app and look for Lucene metadata => unrealistic
>>  - there is some kind of convention on where the indexes are and we do index scanning at startup => scanning are very likely to be slower that class scanning (potential remote access, bigger dataset etc)
>>  - we enforce one or a fixed number of Lucene indexes for all data in Infinispan => not sure that's a good idea but this can be explored
>>  - we somehow ask the framework using HSearch to fill up classes
>> 
>> other approaches?
> 
> why was class scanning discarded in the first answer? as H. Search can
> auto-discover classes by working on top of JPA entity autodiscovery, I
> guess that each application node could look into it's own known
> classpath.
> After all if some type is not visible to him as it was added from
> another node from a different app, he won't be able to return
> instances of it either.
> We could face the opposite problem of building metadata of classes
> people doesn't mean to index in this cache.

Right. scanning (class or index) will be a bit aggressive and could build unneeded metadata (or even worse, return unexpected classes).