On 25 Jun 2014, at 08:45, Dan Berindei <dan.berindei(a)gmail.com> wrote:
On Tue, Jun 24, 2014 at 9:04 PM, William Burns <mudokonman(a)gmail.com> wrote:
I must admit this seems a bit heavy handed to have to enable
transactions, when are using them solely for the purpose of having
implicit transactions.
Can we not instead just tweak the NonTransactionalLockingInterceptor
to obey the FORCE_WRITE_LOCK, so it would guarantee a consistent value
in the face of a concurrent write when doing getCacheEntry?
I suggested that as well, but ISPN-2956 means we can't guarantee the atomicity of
non-tx conditional operations anyway.
And HotRod doesn't work nicely with optimistic locking (yet), so we have to require
pessimistic locking. I'm not sure about total order, though.
Indeed, I agree that transactions is a bit heavy handed, but the fact that transactions
are needed to deal with ISPN-2956 properly, made me lean that way.
Although thinking again on the locking, I don't think it alone
is
sufficient either. As the cache entry is serialized after releasing
the lock, which means there is still a window when only the value may
be changed on an owner node. We really need immutable CacheEntries
returned from getCacheEntry even with locking to work properly.
Galder didn't mention this, but his proposal also copies the entry in
GetKeyValueCommand.perform, so the serialization happens on an ImmutableCacheEntryView.
Yeah, ImmutableCacheEntryView effectively caches the cache entry’s values at construction
time. Assuming there’s lock in place, that should be safe.
The thing to note here is that replaceIfUmodified() needs both the version and value part
to be correct, hence why we added the transaction and force write lock, to avoid one not
matching up with the other. The other piece of the puzzle is getVersioned() operation,
that the client calls to find an entry’s version and use that. Since the only thing sent
to the server is the version (and not the value), I’ve stayed away from making
getVersioned() calling getCacheEntry with FORCE_WRITE_LOCK and a transaction.
Cheers,
- Will
On Tue, Jun 24, 2014 at 1:51 PM, Pedro Ruivo <pedro(a)infinispan.org> wrote:
>
>
> On 06/24/2014 05:11 PM, Mircea Markus wrote:
>>
>> On Jun 24, 2014, at 16:50, Galder Zamarreño <galder(a)redhat.com> wrote:
>>
>>>
>>> On 24 Jun 2014, at 16:51, Mircea Markus <mmarkus(a)redhat.com> wrote:
>>>
>>>>
>>>> On Jun 24, 2014, at 15:27, Galder Zamarreño <galder(a)redhat.com>
wrote:
>>>>
>>>>> To fix this, I’ve been working with Dan on some solutions and we’ve
taken inspiration of the new requirements appearing as a result of ISPN-2956. To be able
to deal with partial application of conditional operations properly, transactional caches
are needed. So, the solution that can be seen in [4] takes that, and creates a transaction
around the replaceIfUmodified and forces the getCacheEntry() call to acquire lock via
FORCE_WRITE_LOCK flag.
>>>>
>>>> so the entire underlaying cache needs to be transactional then?
>>>
>>> Yeah, it needs to be transactional, but the code I’ve written also deals
with the fact that the cache might not be transactional. I’ll probably add a WARN message
when it’s not transactional. This goes in hand with the recommendations for ISPN-2956,
whereby failover for conditional operations relying on return values require transactional
caches to properly deal with failover situations.
>>>
>>> To sum up, if using conditional operations or CRUD methods with
Flag.FORCE_RETURN_VALUE, caches should be transactional. Moreover, to achieve concurrency
guarantees of counter tests such as the one tested for 4424, locking needs to be
pessimistic too. If not using conditional operations or CRUD methods without the flag, the
cache could be non-transactional.
>>
>> Thanks for the analysis. I think we should go with your patch for ISPN 7.0 and
consider the proper solution for the future, as you suggest below.
>> +1 for the warning, users should be made aware for the limitation.
>
> +1 for the patch.
>
> Another suggestion: in *InternalEntryFactory.update()*, you can
I don't know if that would cover all the places since we also set the
value in the various WriteCommands themselves.
I think the WriteCommands only modify wrapped entries.
It would be really bad if a put(k, v) in a read-committed tx cache could modify the cache
entry before the tx is committed. Making this distinction clear would be another reason to
make InternalCacheEntry immutable...
> synchronize in the cache entry, and create a new method
> *InternalCacheEntry copy(InternalCacheEntry)* that makes a copy while it
> also synchronizes in the existing cache entry. I think in this way, you
> don't need the cache to be transactional neither to force the lock on
> reads. Also, I would suggest the copy() to be invoked in your case (or
> in the conditional commands accesses to DataContainer?).
Nitpicking: copy() only exists in the CopyableDeltaAware interface, and the commands
themselves should not access the DataContainer.
This sounds like it might work, and in fact it's quite similar to Takayoshi's
initial proposal (though he targeted it specifically to HotRod's replaceIfUnmodified).
It's a bit scary because CacheEntry.setValue() is called in so many places, but I
think everything but InternalEntryFactoryImpl.update() should be working on context
entries, so they don't need synchronization. Still, ISPN-2956...
Yeah I was thinking of this approach as well, we would either have to
make the copy or synchronize in the serialization. I personally think
making a copy would be better, but the entire copy operation would
have to be synchronized then as you mentioned.
>
>>>
>>>>
>>>>> This solves the issue explained above, but of course it has an
impact of the performance. The test now runs about 1.5 or 2 times slower.
>>>>>
>>>>> This is probably the best that we can do in the 7.0 time frame, but
there’s several things that could improve this:
>>>>>
>>>>> 1. True immutable entries in the cache. If the entries in the cache
were truly immutable, there would be no risk of sending back a partially correct entry
back to the client.
>>>>>
>>>>> 2. A cache replace method that does not compare objects based on
equality (the current replace()), but a replace method that takes a function. The function
could compare that the old entry’s version and the cached entry’s version match. The
function would only be executed inside the inner container, with all the locks held
properly. I already hinted something similar in [5].
>>
>>
>> Shall we raise a new JIRA for the permanent solution then?
>
> +1
>
> I think the immutable entries would be better, but it will destroy
> ReadCommitted isolation. The ReadCommitted is based on the mutability of
> the entry (i.e. it will not invoke the data container if the entry
> exists in transaction context). When (and if) is removed, then we can
> make the entries immutable.
Agree, it would be a significant change. Perhaps context entries in read-committed caches
could hold a reference to the actual EquivalentConcurrentHashMapV8.MapEntry if the entry
was local, but that would add more complexity. So I would rather see it removed
completely.
If we only make getCacheEntry immutable, it wouldn't affect the normal
get operations. I was thinking we could probably make the copy only
in the GetKeyValueCommand.
It would still affect every write operation. And we would need a way to distinguish
between GetKeyValueCommands for getCacheEntry and generated by ClusteredGetCommands, since
both kinds return InternalCacheEntries.
On the other hand, I think we have the same race with the write skew check in optimistic
caches. We read the value and version for write skew checking without holding the key
lock, so in theory it's possible to read an outdated value and the current version. So
we need some kind of synchronization when reading the entry for optimistic mode, anyway.
But instead of a synchronized block, I propose making the value and metadata references
volatile. It would require care to always read the metadata first and write it last, but
reads would still be "free" on x86. HotRod would still need transactions, but it
wouldn't need FORCE_WRITE_LOCK, so this might be a little faster than Galder's
version.
>
>>
>>>>>
>>>>> Finally, this was not a problem when the value stored in Hot Rod was
a ByteArrayValue wrapping the byte array and the version, because the value was treated
atomically, and in hindsight, maybe adding getCacheEntry might have been premature, but
this method has proven useful for other use cases too (rolling upgrades…etc).
>>>>>
>>>>> Cheers,
>>>>>
>>>>> [1]
https://issues.jboss.org/browse/ISPN-4424
>>>>> [2]
https://issues.jboss.org/browse/ISPN-2956
>>>>> [3]
https://github.com/infinispan/infinispan/blob/master/server/core/src/main...
>>>>> [4]
https://github.com/galderz/infinispan/tree/t_4424
>>>>> [5]
https://github.com/infinispan/infinispan/blob/master/core/src/main/java/o...
>>>>> --
>>>>> Galder Zamarreño
>>>>> galder(a)redhat.com
>>>>>
twitter.com/galderz
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> infinispan-dev mailing list
>>>>> infinispan-dev(a)lists.jboss.org
>>>>>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>>
>>>> Cheers,
>>>> --
>>>> Mircea Markus
>>>> Infinispan lead (
www.infinispan.org)
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> infinispan-dev mailing list
>>>> infinispan-dev(a)lists.jboss.org
>>>>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>
>>>
>>> --
>>> Galder Zamarreño
>>> galder(a)redhat.com
>>>
twitter.com/galderz
>>>
>>>
>>> _______________________________________________
>>> infinispan-dev mailing list
>>> infinispan-dev(a)lists.jboss.org
>>>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>
>> Cheers,
>>
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev(a)lists.jboss.org
>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
_______________________________________________
infinispan-dev mailing list
infinispan-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev
_______________________________________________
infinispan-dev mailing list
infinispan-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev
--
Galder Zamarreño
galder(a)redhat.com
twitter.com/galderz