<div dir="ltr">I think we may be confusing two different features. These changes are what I was hoping could be used as a basis for the transformers work (doesn't mean they will be). These changes were done before there was any discussions about transformers though, so they may not have the use cases covered. I think this might be why there was some confusion.<br><br><div class="gmail_quote"><div dir="ltr">On Thu, Jun 1, 2017 at 8:11 AM Sanne Grinovero <<a href="mailto:sanne@infinispan.org">sanne@infinispan.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Thanks all for the great explanations.<br>
<br>
+1 to try to avoid this. Could the transformation be performed lazily?<br></blockquote><div><br></div><div>I can't think of a reason we can't try to avoid the collection allocations with what I proposed.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
We had a design meeting about transformers and that was a goal we had<br>
agreed on.</blockquote><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
For example in the Search case, I'm doing a getAll( keys ) just to<br>
return the objects in a sorted list back to the user - in whatever<br>
form they are.<br>
In this case the engine doesn't need the objects to be in any specific<br>
form, so the transformation / wrapping phase needs to be moved to the<br>
client endpoint to become an optional, transparent step.<br></blockquote><div><br></div><div>These changes have nothing to do with the client itself. In fact these were mostly added for byte[] requirements. I just added in Compatibility support which required the same type of code.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
Some more comments inline:<br>
<br>
On 1 June 2017 at 11:12, Dan Berindei <<a href="mailto:dan.berindei@gmail.com" target="_blank">dan.berindei@gmail.com</a>> wrote:<br>
> On Wed, May 31, 2017 at 10:05 PM, William Burns <<a href="mailto:mudokonman@gmail.com" target="_blank">mudokonman@gmail.com</a>> wrote:<br>
>> Let me explain why it is there first :) This class was added for two main<br>
>> reasons: as a replacement for compatibility and for supporting equality of<br>
>> byte[] object. What this class does is at the user side is box the given<br>
>> arguments (eg. byte[] -> WrappedByteArray) then the cache only ever deals<br>
>> with the boxed types and then does the unboxing for any values that are<br>
>> returned.<br>
>><br>
>> There are some exceptions with how the boxing/unboxing works for both cases<br>
>> such as Streams and Indexing which have to rebox the data to work properly.<br>
>> But the cost is pretty minimal.<br>
>><br>
>> While compatibility is always on or off unfortunately anyone can pass in a<br>
>> byte[] at any point for a key or value. So we need to have these wrappers<br>
>> there to make sure they work properly.<br>
>><br>
>> We could add a option to the cache, which people didn't show interest<br>
>> before, to have a cache that doesn't support byte[] or compatibility. In<br>
>> this case there would be no need for the wrapper.<br>
<br>
Let's try to avoid a configuration option; if the Transformers design<br>
I mentioned above doesn't fly we could have Infinispan Query use a<br>
dedicated internal API which skips such operations.<br>
<br>
>><br>
>><br>
>> Compatibility<br>
>><br>
>> By using the wrapper around the cache, compatibility becomes quite trivial<br>
>> since we just need a converter in the wrapper and it does everything else<br>
>> for us. My guess is the new encoding changes will utilise these wrapper<br>
>> classes as well as they are quite easy to plug in and have things just work.<br>
<br>
Remember there's more than byte[] vs Pojo, as we also have other<br>
formats which need supporting, and plans to support JSON as well<br>
(which is not a String nor a POJO).<br>
<br>
The query engine will likely need to convert / interpret the "storage<br>
format" into its own format, or like Adrian pointed out being a great<br>
capability of Protobuf is to read only a selection of fields.<br>
<br>
So I wouldn't automatically apply these operations, as they must be<br>
client dependent. Tristan has the design documents from the Infinispan<br>
Query meeting we held in London, I briefly pointed Will to them during<br>
our Konstanz meeting. May I hope that's going to be the next step?</blockquote><div><br></div><div>This is what Gustavo and others are working on.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> </blockquote><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Sounds like it would also avoid this performance issue.<br></blockquote><div><br></div><div>I can't say, but the byte[] is another issue, since it is needed outside of client/server.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
<br>
>> Equality<br>
>><br>
>> With the rewrite for eviction we lost the ability to use custom Equality in<br>
>> the data container. The only option for that is to wrap a byte[] to provide<br>
>> our own equality. Therefore the wrapper does this conversion for us by<br>
>> converting between automatically.<br>
>><br>
>> On Wed, May 31, 2017 at 2:34 PM Sanne Grinovero <<a href="mailto:sanne@infinispan.org" target="_blank">sanne@infinispan.org</a>><br>
>> wrote:<br>
>>><br>
>>> Hi all,<br>
>>><br>
>>> I've been running some benchmarks and for the fist time playing with<br>
>>> Infinispan 9+, so please bear with me as I might shoot some dumb<br>
>>> questions to the list in the following days.<br>
>>><br>
>>> The need for TypeConverterDelegatingAdvancedCache to wrap most<br>
>>> operations - especially "convertKeys" - is highlighet as one of the<br>
>><br>
>><br>
>> It should be wrapping every operation pretty much.<br>
>><br>
>> Unfortunately the methods this hurts the most are putAll, getAll etc as they<br>
>> have to not only box every entry but copy them into a new collection as you<br>
>> saw in "convertKeys". And for getAll it also has to unbox the return as<br>
>> well.<br>
>><br>
>> We could reduce allocations in the collection methods by not creating the<br>
>> new collection until we run into one key or value that required<br>
>> boxing/unboxing. This would still require fully iterating over the<br>
>> collection at best case. This should perform well in majority of cases as I<br>
>> would expect all or almost all entries either require or don't require the<br>
>> boxing. The cases that would be harmed most would be ones that have a sparse<br>
>> number of entries that require boxing.<br>
>><br>
><br>
> +1 for checking the keys/values first and only creating a new<br>
> collection if necessary, although it doesn't seem possible with the<br>
> current TypeConverter interface.<br>
<br>
+1<br>
<br>
> I'm also fine with adding a configuration option to prohibit byte[]<br>
> keys/values, but I'd still want to check the collection and throw an<br>
> exception (at least for writes).<br>
><br>
> Thinking further, do we really need to wrap byte[] values?<br>
> DataContainer only has a compute() method, so we may be able to<br>
> replace the value wrappers with a TypeConverter.valuesEqual(a, b)<br>
> method.<br>
><br>
> Also, Sanne, could you change InternalCacheFactory.createAndWire() to<br>
> return the un-wrapped cache, and then see exactly much the wrapping<br>
> affects performance? With so many moving parts, I wouldn't be<br>
> surprised if the difference in numbers was < 1%.<br>
<br>
Indeed, I don't expect a significant change in throughput: it's an<br>
allocation issue and as such it's very situational if it's going to<br>
affect someone or not. Not worth measuring as my current bottleneck is<br>
caused by other open issues.<br>
<br>
But my main concern now is that apparently this approach does not<br>
address some of our needs as previously agreed, which we need to<br>
improve usability such as transparent transcoding. I guess that's<br>
coming next? In that case we can drop it for now, it's no big deal and<br>
I was just hoping to have identified a low hanging fruit.<br></blockquote><div><br></div><div>Yeah I am looking forward to the new changes as well :)</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
Thanks,<br>
Sanne<br>
<br>
><br>
>>><br>
>>> high allocators in my Search-centric use case.<br>
>>><br>
>>> I'm wondering:<br>
>>> A - Could this implementation be improved?<br>
>><br>
>><br>
>> Most anything can be improved :) The best way would be to add another knob.<br>
>><br>
>>> B - Could I bypass / disable it? Not sure why it's there.<br>
>><br>
>><br>
>> There is no way to bypass currently. Explained above.<br>
>><br>
>><br>
>>><br>
>>><br>
>>> Thanks,<br>
>>> Sanne<br>
>>> _______________________________________________<br>
>>> infinispan-dev mailing list<br>
>>> <a href="mailto:infinispan-dev@lists.jboss.org" target="_blank">infinispan-dev@lists.jboss.org</a><br>
>>> <a href="https://lists.jboss.org/mailman/listinfo/infinispan-dev" rel="noreferrer" target="_blank">https://lists.jboss.org/mailman/listinfo/infinispan-dev</a><br>
>>><br>
>><br>
>> _______________________________________________<br>
>> infinispan-dev mailing list<br>
>> <a href="mailto:infinispan-dev@lists.jboss.org" target="_blank">infinispan-dev@lists.jboss.org</a><br>
>> <a href="https://lists.jboss.org/mailman/listinfo/infinispan-dev" rel="noreferrer" target="_blank">https://lists.jboss.org/mailman/listinfo/infinispan-dev</a><br>
> _______________________________________________<br>
> infinispan-dev mailing list<br>
> <a href="mailto:infinispan-dev@lists.jboss.org" target="_blank">infinispan-dev@lists.jboss.org</a><br>
> <a href="https://lists.jboss.org/mailman/listinfo/infinispan-dev" rel="noreferrer" target="_blank">https://lists.jboss.org/mailman/listinfo/infinispan-dev</a><br>
_______________________________________________<br>
infinispan-dev mailing list<br>
<a href="mailto:infinispan-dev@lists.jboss.org" target="_blank">infinispan-dev@lists.jboss.org</a><br>
<a href="https://lists.jboss.org/mailman/listinfo/infinispan-dev" rel="noreferrer" target="_blank">https://lists.jboss.org/mailman/listinfo/infinispan-dev</a><br>
</blockquote></div></div>