[infinispan-dev] Auto-detection of appropriate key2StringMapperClass

[Dan Berindei]

On Thu, Mar 9, 2017 at 1:09 PM, Sanne Grinovero <sanne at infinispan.org> wrote:
> On 9 March 2017 at 10:22, Dan Berindei <dan.berindei at gmail.com> wrote:
>> On Thu, Mar 9, 2017 at 11:48 AM, Sanne Grinovero <sanne at infinispan.org> wrote:
>>> Hi Dan!
>>>
>>
>> Hey Sanne :)
>>
>>> On 9 March 2017 at 09:38, Dan Berindei <dan.berindei at gmail.com> wrote:
>>>> Every store can have its own Key2StringMapper, so I really don't think
>>>> service discovery is appropriate here.
>>>
>>> I suggested a default, while maintaining the property as an override option.
>>>
>>> It won't make any difference to people in the situation you describe,
>>> but it will avoid trouble for many more users.
>>>
>>> In my experience when one develops a custom type & a custom
>>> Key2StringMapper, you really wouldn't have interest in creating
>>> multiple implementations for the same type. The only case in which I
>>> can see some use for allowing multiple conflicting registrations to
>>> the same type is for upgrade / migration cases..
>>>
>>
>> The problem I see is that each store can only have one
>> Key2StringMapper implementation, so you're going to have conflicts
>> even if your caches use completely different key types. The store
>> doesn't know which key types the application actually uses, so it
>> can't pick the appropriate implementation when there are multiple ones
>> on the classpath.
>
> Good point. I was hoping to allow different services to "contribute"
> their different converters but I realise now that it would introduce a
> risk of ambiguous representations.
> They would need some standard scoping to know to which one to delegate
> to on read operations. I wonder if we could enforce / validate that?
>

We could prefix the string representation with the class and module
name of the mapper ;)

>>>> My favourite solution would be to have a default Key2StringMapper
>>>> implementation that uses Java serialization (not our internal
>>>> marshalling, because that can change between versions). It could even
>>>> do base64 encoding for databases that don't have a VARBINARY/BYTEA
>>>> equivalent.
>>>>
>>
>> Do you see any problem with using serialization everywhere?
>
> I guess we could discuss alternative strategies, but it's an
> orthogonal discussion right? You're proposing a different CacheStore
> implementation, but people would still need a better way to associate
> the adaptors to the types.
>

I don't think it's orthogonal, because having a single key mapper that
supports all key types would make this discussion unnecessary.

> But to answer your question: yes I wouldn't like using Java
> serialization: people need to be able to evolve their domain model.
> When we had to "break" the string format in the Lucene mappers (as I
> added a new field) we managed to do it in an evolutionary way: it
> wasn't too hard to make sure we could read the old string format,
> while with Java serialization this would have been a mess.
>

If your keys are internal it should be possible to use a new key type
and to keep supporting the old type for a while. But yeah, it's a
little more involved than adding support for a new prefix in your key
mapper.

> Also it would be nice to keep the representation as short as possible
> for both efficient processing and storage space efficiency: Java
> Serialization isn't particularly compact.
>

Yeah, I should have just said we should use the same marshalling for
the keys that we use for the values.

I didn't want to bring up our marshalling because we don't have a very
good record for preserving the format of JDK types that get included
in the user types, but Java serialization has its own warts. In the
end, the important point is that what's good for the values must be
good for the keys too.

Cheers
Dan