NH mediocre support for distributed caches

[JL Borges]

Hello All,

I've been hacking the NH L2 cache for a few months, and I have come
to the conclusion that it is not really designed for distributed
caches.

Here are the issues I am concerned about:

1) Server Round Trip

With a distributed cache, a lot of the time is spent sending/receiving
data over
the socket, so performance increases dramatically when these
roundtrips are reduced. Many modern distributed caches, like
Memcached and Redis, support client-side pipelining, where multiple
gets or puts are sent in one socket call. The NH ICache interface does
not support multiple puts or gets, so this feature is not available.

2) Locking

a) If a distributed cache supports distributed locks, then there is no
need for an in-process lock, and it just harms performance. And yet,
here is a code snippet from the
ReadWriteCache concurrency strategy

lock (_lockObject)
{
if (log.IsDebugEnabled)
{
log.Debug("Caching: " + key);
}
try
{
cache.Lock(key);
.
.
.

b) distributed locks can fail due to timeout or server crashing after
acquiring the lock.
There is no logic here to support this.

I have a patch to address these issues, in my own NH fork. I've also
submitted a patch to JIRA to add distributed locks to the Enyim
memcached client. No response so far.

Is there any interest in improving support for distributed caches?

Cheers,
JL Borges

[Fabio Maulo]

Hi Jorge/Aaron/JLBorges (same e-mail different name/sign)

http://groups.google.com/group/nhcdevs/browse_thread/thread/ebb6a11601c8365b

http://groups.google.com/group/nhcdevs/browse_thread/thread/8d6b58b141e66395

http://groups.google.com/group/nhcdevs/browse_thread/thread/a3f9bf2ab44d1d0b

http://groups.google.com/group/nhibernate-development/browse_thread/thread/b62a8b7a0a1278d9

and continuing counting...

- Show quoted text -

--
Fabio Maulo

[Fabio Maulo]

Have you some news about this

http://groups.google.com/group/nhcdevs/browse_thread/thread/f9b93ce4400a1d4f
?

--
Fabio Maulo

[Ramon]

Fabio,

What exactly is your problem? He is just asking some questions and you respond with such weird stuff. I dont get it.

[Fabio Maulo]

Only, I would know with who I'm talking.

One of the previous question (using the name Aaron) was about add locks for no-thread-safe cache system.

https://groups.google.com/group/nhibernate-development/browse_thread/thread/88fb6a16adfa7cb7/c883b73b3d16b39c

Cache providers are injectable pieces and can be published whatever where you want and Jorge/Aaron/JLBorges can ask permissions to blog in nhforge.org to promote his providers.

In this case Jorge/Aaron/JLBorges is asking about an "improvement" of ICache: we can discuss it, no problem, but I would be sure that it is really needed.

On the other hand we have to port some other features, about cache, from Hibernate as IsMinimalPutsEnabledByDefault and some other where needed (read/update).

--
Fabio Maulo

[JL Borges]

Let's deal with this issue on its own merits.

This is about cache concurrency strategy and ICache interface.

Cache concurrency strategy is currently fixed to three different
strategies at the moment.
ICache interface is fixed.

Proposal:

1) refactor concurrency strategies to avoid unnecessary locking for
distributed caches
2) refactor ICache to support multiple get/put


Of course, all cache providers will have to do some work to support
the refactor.

Kind Regards,
JL Borges

On Feb 2, 10:49 am, Fabio Maulo <fabioma...@gmail.com> wrote:
> Only, I would know with who I'm talking.
>
> One of the previous question (using the name Aaron) was about add locks for

> no-thread-safe cache system.https://groups.google.com/group/nhibernate-development/browse_thread/...
>
> <https://groups.google.com/group/nhibernate-development/browse_thread/...>Cache

> providers are injectable pieces and can be published whatever where you want
> and Jorge/Aaron/JLBorges can ask permissions to blog in nhforge.org to
> promote his providers.
>
> In this case Jorge/Aaron/JLBorges is asking about an "improvement" of
> ICache: we can discuss it, no problem, but I would be sure that it is really
> needed.
>
> On the other hand we have to port some other features, about cache, from
> Hibernate as IsMinimalPutsEnabledByDefault and some other where needed
> (read/update).
>
>
>
>
>
>
>
>
>
> On Wed, Feb 2, 2011 at 12:19 PM, Ramon <ramon.sm...@gmail.com> wrote:
>
> > Fabio,
>
> > What exactly is your problem? He is just asking some questions and you
> > respond with such weird stuff. I dont get it.
>

> > On Wed, Feb 2, 2011 at 4:13 PM, Fabio Maulo <fabioma...@gmail.com> wrote:
>
> >> Hi Jorge/Aaron/JLBorges (same e-mail different name/sign)
>

> >>http://groups.google.com/group/nhcdevs/browse_thread/thread/ebb6a1160...
>
> >>http://groups.google.com/group/nhcdevs/browse_thread/thread/8d6b58b14...
>
> >>http://groups.google.com/group/nhcdevs/browse_thread/thread/a3f9bf2ab...
>
> >>http://groups.google.com/group/nhibernate-development/browse_thread/t...

>
> >> and continuing counting...
>
> >> - Show quoted text -
>

[Fabio Maulo]

both (1 and 2) needs a modification of ICacheProvider as:

bool SupportsMultipleGets

bool SupportsMultiplePuts

bool NeedsLocks

For multiple get/put I'm not sure if it really needs a modification inside the core or if it can be achieved directly inside the provider implementation (caching n-puts calls before call the underlining multiple-puts).

If we have to change NH the modification may be in the Loader.InitializeEntitiesAndCollections calling the multiple-puts, per persister, for all uploaded objects (perhaps the time to implement a InitializeEntityEventListener has come).

--
Fabio Maulo

[Fabio Maulo]

About cache strategies first of all we should implements the Transactional strategy, then:

1) We should remove the restriction of the 3 already available strategies and give the ability to specify the AssemblyQualifiedName of the concrete implementation of ICacheConcurrencyStrategy.

This way we will have four embedded strategies but the user will have the ability to define his own custom implementation.

2) The other option is to delegate to the ICacheProvider the instantiation of the concrete class for each of the four strategies; this way the provider will have full control of each operation.

3) Injectable ICacheFactory through session-factory configuration

All options can be implemented even together.

Considering the downloads of NHCH (2% of NH itself) I prefer the implementation of the option 2 and the remotion of actual CacheFactory.

--
Fabio Maulo

[JL Borges]

Yes, I think the least invasive change would be to add the following
methods to ICacheProvider

void BeginPutCaching()
void BeginGetCaching()
void Flush()

Providers that don't have pipelining could ignore these.
Those that do support pipelining could store put/get in local
hashtable, then send/request all at once in Flush method.

Calls to BeginPutCaching/Flush would happen in
Loader.InitializeEntitiesAndCollections

Calls to BeginGetCaching/Flush would happen in StandardQueryCache.Get,
to optimize
retrieval of query results.

For locking, it would be nice to have a locking strategy for each
provider.
And we could provide a few ready made strategies:

NoLockStrategy : does nothing
ReadWriteStrategy: in-process locking

Then we would simply call

cache.LockStrategy().Lock()
...
cache.LockStrategy().Unlock()

More performant would be:

cache.LockStrategy().GetReadLock() /
cache.LockStrategy().ReleaseReadLock()
cache.LockStrategy().GetWriteLock() /
cache.LockStrategy().ReleasWriteLock()









////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////

[JL Borges]

On Feb 2, 12:27 pm, Fabio Maulo <fabioma...@gmail.com> wrote:

> > > no-thread-safe cache system.
> >https://groups.google.com/group/nhibernate-development/browse_thread/...
>
> > > <https://groups.google.com/group/nhibernate-development/browse_thread/..

> > .>Cache
> > > providers are injectable pieces and can be published whatever where you
> > want
> > > and Jorge/Aaron/JLBorges can ask permissions to blog in nhforge.org to
> > > promote his providers.
>
> > > In this case Jorge/Aaron/JLBorges is asking about an "improvement" of
> > > ICache: we can discuss it, no problem, but I would be sure that it is
> > really
> > > needed.
>
> > > On the other hand we have to port some other features, about cache, from
> > > Hibernate as IsMinimalPutsEnabledByDefault and some other where needed
> > > (read/update).
>
> > > On Wed, Feb 2, 2011 at 12:19 PM, Ramon <ramon.sm...@gmail.com> wrote:
>
> > > > Fabio,
>
> > > > What exactly is your problem? He is just asking some questions and you
> > > > respond with such weird stuff. I dont get it.
>
> > > > On Wed, Feb 2, 2011 at 4:13 PM, Fabio Maulo <fabioma...@gmail.com>
> > wrote:
>
> > > >> Hi Jorge/Aaron/JLBorges (same e-mail different name/sign)
>

> > > >>http://groups.google.com/group/nhcdevs/browse_thread/thread/ebb6a1160.
> > ..
>
> > > >>http://groups.google.com/group/nhcdevs/browse_thread/thread/8d6b58b14.
> > ..
>
> > > >>http://groups.google.com/group/nhcdevs/browse_thread/thread/a3f9bf2ab.
> > ..
>
> > > >>http://groups.google.com/group/nhibernate-development/browse_thread/t.
> > ..
>

[Fabio Maulo]

I don't think it is viable for gets

--
Fabio Maulo

[JL Borges]

On Feb 2, 2:32 pm, Fabio Maulo <fabioma...@gmail.com> wrote:
> I don't think it is viable for gets
>
>

Yeah, we would need an ICache.MultiGet method as well.

Here is my Prefetch implementation for the standard query cache:



public override void Prefetch(IList cacheable,
IEnumerable<ICacheAssembler> returnTypes)
{
if (cacheable == null || cacheable.Count == 0) return;

var prefetchDict = new Dictionary<ICache, ISet<CacheKey>
>();

//note: first index in cacheable is a timestamp
for (var i = 1; i < cacheable.Count; i++)
{
var cacheOid = cacheable[i];
foreach (var type in returnTypes)
{
if (cacheOid ==
LazyPropertyInitializer.UnfetchedProperty ||
cacheOid == BackrefPropertyAccessor.Unknown)
continue;
var entity = type as EntityType;
if (entity == null) continue;
var entityId =
entity.GetIdentifierType(this).Assemble(cacheOid, this, null);
var entityName =
entity.GetAssociatedEntityName(Factory);
var persister =
Factory.GetEntityPersister(entityName);
var entityType = persister.IdentifierType;

var prefetchKey = new CacheKey(entityId,
entityType, entityName, EntityMode, Factory);
var cache = persister.Cache.Cache;
ISet<CacheKey> cacheKeys = null;
if ( !prefetchDict.ContainsKey(cache))
{
cacheKeys = new HashedSet<CacheKey>();
prefetchDict[cache] = cacheKeys;
}
else
{
cacheKeys = prefetchDict[cache];
}
cacheKeys.Add(prefetchKey);
}
}
//prefetch sent to all persister caches
foreach (var cache in prefetchDict.Keys)
{
var keys = prefetchDict[cache];
if (keys.Count > 1)
{
IDictionary results =
cache.MultiGet(prefetchDict[cache]);
//store in local cache
foreach (var key in results.Keys)
{
_prefetchCache.Put(new
CachePutParameters(null, key, results[key]));

[JL Borges]

Yes, I think #2 is the best option; most flexible

> > On Wed, Feb 2, 2011 at 2:04 PM, JL Borges <boxe...@gmail.com> wrote:
>
> >> Let's deal with this issue on its own merits.
>
> >> This is about cache concurrency strategy and ICache interface.
>
> >> Cache concurrency strategy is currently fixed to three different
> >> strategies at the moment.
> >> ICache interface is fixed.
>
> >> Proposal:
>
> >> 1) refactor concurrency strategies to avoid unnecessary locking for
> >> distributed caches
> >> 2) refactor ICache to support multiple get/put
>
> >> Of course, all cache providers will have to do some work to support
> >> the refactor.
>
> >> Kind Regards,
> >> JL Borges
>
> >> On Feb 2, 10:49 am, Fabio Maulo <fabioma...@gmail.com> wrote:
> >> > Only, I would know with who I'm talking.
>
> >> > One of the previous question (using the name Aaron) was about add locks
> >> for

> >> > no-thread-safe cache system.
> >>https://groups.google.com/group/nhibernate-development/browse_thread/...
>
> >> > <
> >>https://groups.google.com/group/nhibernate-development/browse_thread/..