[infinispan-dev] Multi get API

[Radim Vansa]

On 03/24/2015 02:22 PM, William Burns wrote:
> I am nearing completion of the new multi get command [1], allowing for
> more efficient retrieval of multiple keys at the same time.  Thanks
> again to Radim for a lot of the initial work.
>
> In doing so though I want to make sure I get feedback on how we want
> the API to actually look, since this is much more difficult to change
> down the line.
>
> There are a few things in particular I wanted to discuss.
>
> 1. The actual name of the method.  The main suggestions I have seen
> are getAll or getMany.  I do like the naming of the former, however it
> seems a bit misleading (remember API is getAll(Set) since we are
> really getting a subset.  So I am thinking the possibilities at this
> point are getAllOf, getAllFor or getMany.  I am leaning maybe towards
> the last one (getMany).  I am open to any suggestions though.

Actually Tristan suggested the name multiGet() - I would prefer this one 
in the end, and adding multiPut() that would just do putAll (to have the 
API symmetric) and deprecate the putAll() method. multiRemove and others 
can follow later, and the naming is straightforward.
I would object against getAll() since this sounds like retrieving all 
entries from the cache, not just the specified keys.

>
> 2. What should the return value be for this method.  Currently this
> returns a Map<K, V> which makes sense when we retain these values in
> the local transactional context and is pretty nice and clean for end
> users.
>
> The other idea is to use a streaming API possibly returning an
> Iterable<CacheEntry<K, V>>.  The streaming return value doesn't seem
> as intuitive to me, but opens us up for more efficient usage
> especially when there may be a lot of keys passed in (also results can
> be processed concurrently while retrieval is still occurring).
>
> I would lean towards returning the Map<K, V> value, however the next
> point could change that.

I think that Iterable<CacheEntry<K, V>> is too confusing for the end 
user, I would stick to the Map. If you want lazy loading, the Map (and 
it's EntrySet) could be made lazy by a flag.

>
> 3. Where this method should live.  Right now this method is on the
> BasicCache interface which is a parent of both Cache (embedded) and
> RemoteCache (remote).  Unfortunately for remote clients to be as
> efficient as possible we really need a Streaming API so that we don't
> have multiple copies of data in memory at multiple places at the same
> time.  For that reason I suggest we use the streaminig API for both or
> have a different API for remote vs embedded.
>

Radim

-- 
Radim Vansa <rvansa at redhat.com>
JBoss Performance Team