Cache.put() triggers CacheStore.load()

[aolenev]

Hi,

I'm using Infinispan 7.0.0.Alpha4 and faced with strange behaviour: if I put object in cache it triggers cacheStore.load(). Snippet from my configuration:

<distributed-cache name="string-cache" owners="2" mode="SYNC">
            <state-transfer enabled="true" />
            <persistence>
                <store class="org.sproot_grid.infinispan.cachestore.StringCacheStore">
                    <property name="driver">org.h2.Driver</property>
                    <property name="connection-url">jdbc:h2:mem:testdb</property>
                    <property name="username">sa</property>
                    <property name="password"></property>
                </store>
            </persistence>
        </distributed-cache>

StringCacheStore - is my implementation of CacheLoader and CacheWriter

I found possible reason here (org.infinispan.factories.InterceptorChainFactory):

if (configuration.persistence().usingStores()) {
         if (configuration.persistence().passivation()) {
            if (configuration.clustering().cacheMode().isClustered())
               interceptorChain.appendInterceptor(createInterceptor(new ClusteredActivationInterceptor(), ClusteredActivationInterceptor.class), false);
            else
               interceptorChain.appendInterceptor(createInterceptor(new ActivationInterceptor(), ActivationInterceptor.class), false);
         } else {
            if (configuration.clustering().cacheMode().isClustered())
               interceptorChain.appendInterceptor(createInterceptor(new ClusteredCacheLoaderInterceptor(), ClusteredCacheLoaderInterceptor.class), false);
            else
               interceptorChain.appendInterceptor(createInterceptor(new CacheLoaderInterceptor(), CacheLoaderInterceptor.class), false);
            switch (configuration.clustering().cacheMode()) {
               case DIST_SYNC:
               case DIST_ASYNC:
               case REPL_SYNC:
               case REPL_ASYNC:
                  interceptorChain.appendInterceptor(createInterceptor(new DistCacheWriterInterceptor(), DistCacheWriterInterceptor.class), false);
                  break;
               default:
                  interceptorChain.appendInterceptor(createInterceptor(new CacheWriterInterceptor(), CacheWriterInterceptor.class), false);
                  break;
            }
         }
      }

Why do we need CacheLoaderInterceptor here (despite command is PutKeyValueCommand)?

Moreover it fails with NPE if there is no entry with given key in underlying persistent storage (UPDATE: NPE is fixed, I just had to return null from CacheLoader if nothing found instead of wrapping null into MarshalledEntry).

Can I disable such a workflow by declarative configuration?

[aolenev]

BTW,

If key is missing in cache then CacheStore.load() called twice. Seems like a bug in InterceptorChainFactory of CacheLoaderInterceptor.

[pruivo]

Hi Alexey,

The put contract says it must return the old value for the key. That's why it needs to check the CacheLoader if the value is not found in memory. You can disable this behavior by setting the flag IGNORE_RETURN_VALUES. Example

cache.getAdvancedCache().withFlags(Flag.IGNORE_RETURN_VALUES).put(k, v);

note: this will never return the existing value if it exists

As alternative, if you only want to skip only the CacheLoader, you can use the flag SKIP_CACHE_LOAD

cache.getAdvancedCache().withFlags(Flag.SKIP_CACHE_LOAD).put(k, v);

[pruivo]

Alexey Olenev wrote:

BTW,

If key is missing in cache then CacheStore.load() called twice. Seems like a bug in InterceptorChainFactory of CacheLoaderInterceptor.

What do you mean? It tries to load the key twice for a single operation?

[aolenev]

Thank you, Pedro!

Yes. I saw it in debugger. There were two calls of CacheStore.load() for single Cache.get()

[pruivo]

Hi,

I'm not seeing any reason to try to load it twice. Do you mind to open a JIRA (System Dashboard - JBoss Issue Tracker) about the it?

Thank you

Pedro

[aolenev]

Ok, I'll check it one more time and create JIRA ticket.

Thank you for help

[aolenev]

I checked it once again and there is no bug in ISPN, there was a bug in my test, sorry for bothering.

But I still think that triggering read-through on every put of new key is not an obvious behaviour. May be it would be better to set Flag.SKIP_CACHE_LOAD by default?

[sannegrinovero]

Setting the Flag by default would be a violation of the Map API.

A smart solution is drafted at : infinispan/jokre · GitHub

[genman]

JSR107/JCache has the behavior you want by default.

[aolenev]

Thank you, Elias.

Sanne, I don't think that setting a flag by default is a violation of Map API because as far as I understood Flag.SKIP_CACHE_LOAD just disables CacheLoader.load(), but cache still returns previous value (or null in case there wasn't a value for this key) assigned with key. Moreover, triggering of read-through mechanism on Cache.put() can be a reason of performance degradation and user of ISPN will be surprised by this default behaviour.