[infinispan-dev] CR3: async stores, again

Fri Aug 27 14:37:51 EDT 2010

Some more news about this:
Mircea attached a patch he was working on to JIRA, but he hadn't time
to fix all tests, so I'm continuing.

There are two more issues:

1)keeping an ordered list of transactions works fine for transactions
only, but modifications done in a transaction are wrongly reordered
with operations not being done in a transaction.

2) Even in the committed current code, I see that a clear() operation
is sent async together with all other operations. Then, when the state
changes are performed, the clear() causes the previously found
operations to be canceled, but "previous" applies here on the iterator
on the state map, not on the order of application of those state
Modifications, so the clear() is removing random Modifications.

3) Also my attempt to reuse some code from AbstractCacheStore though
inheritance is not going to make it as this implementation is doing
too much for the needs of AsyncStore: it cares for async Purge, and
all methods I wanted to commit are slightly different because of that.

Proposal:
1- Out-of-transaction modifications are pushed into a BlockingDeque
instead of a Map.

2- A prepare(tx, list) sets the list aside in the "transactions" Map.

3- A commit(tx) takes the list of operations from the "transactions"
and pushes the list on the same BlockingDeque as all other operations,
guaranteeing same order of application.

4- a clear operation clears the deque and pushes a Clear() operation on it

5- a purge operation is sent though the same deque: only risk I see is
being removed by a clear, but I bet there won't be much to purge after
that :) Also we don't care if the purge is going to be handled async
in the other store, that's a store implementation detail from what I
see in AbstractCacheStore.

6- A single processor of AsyncStore drains blocks of a fixed
configurable amount, puts them in a local map, applies the coalescing
algorithm, and puts the results on a new state map.
It seems to me that this worker needs to be exactly one, or reordering
problems arise again in the order of the blocks taken from the deque.

6/a - When receiving a clear() - exceptional case - we clear the map
of operations, and issue the clear() synchronously: not adding
anything to the map, and aquiring a write lock "clearop" to avoid race
conditions with the workers which might already be busy but happen to
finish after the clear().

6/b - When receiving a purge() this bypasses other logic and is
directly passed to the decorated Store.

7- Now here it becomes tricky, we removed the transactions problem,
the clear operation and the purge from the table. But there still is a
map of Modifications to be applied, by N threads, and to which another
thread might write. This looks like the original problem in simplified
form, but reorderings are still possible as worker A might get some
work on key K1, then worker B kicks in, gets new work about K1 and
happens to perform it before A gets priority again.
Clearly this must be performed either by one thread or using very fine
grained locking: as we're only interested on a per-key lock, we don't
care on the order of operations on other keys: we could use a second
map to store the keys which "are being processed": when a worker takes
some work it uses atomic operations to check if somebody is working on
the same stuff, if it is it could either wait or put it back (if it
wasn't changed, otherwise discard) and pick another work to do.

Comments?

AFAIR the issue I mention on point 7 is a known limitation of Async,
so I could implement that in future as an improvement? I would be fine
in having just one worker.

I wish I had time to do it all soon :|
Sanne

2010/8/26 Sanne Grinovero <sanne.grinovero at gmail.com>:
> 2010/8/26 Manik Surtani <manik at jboss.org>:
>>
>> On 26 Aug 2010, at 12:22, Sanne Grinovero wrote:
>>
>>> Hi manik, thanks for reading through my long mail.
>>>
>>> though I didn't understand what you think I should do now :)
>>> Regarding to "Like I said we should still support commit on a separate thread.",
>>> you mean my approach is not correct?
>>> But then you say "Definitely" relating to my question if this is an improvement.
>>
>> I'm not sure how your approach prevents a commit on a separate thread.  :)
>
> because the caller of AsyncStore.prepare(..) - T1 - will have the
> modifications list added to the "transactions" juc.ConcurrentHashMap,
> using the TX as a key, and this same values will be pushed to "state"
> by AsyncStore.commit(..), nothing else is done, the writing to the
> decorated store is done by T2 using the current logic which takes
> modifications from "state", totally independent and unaware to the
> existing transaction used by T1.
>
>>
>>> BTW why should you support "commit" on a separate thread, if all what
>>> the commit does is move a single value from one concurrent map to
>>> another one? the heavy work of sending the data down to the real store
>>> is performed in the separate thread.
>>
>> Ah sorry - I think I have misunderstood you.  I thought what you meant was that T1 may call prepare() and T2 may call commit(), imposing restrictions on how your impl may make use of, say, a ThreadLocal.
>
> that's what is currently possible as the Executor in asyncStore might
> have >1 threads, on turn taking some of the operations, so some
> prepares() could go to thread X and some commits() to thread Y, so it
> seems to me that thread X might perform the prepare for TX1 and the
> thread Y receive the commit for the same TX1 (they might even try to
> perform them in inverted order).
>
>>
>> You are absolutely correct in that the work done in commit() does not need to be taken offline/performed asynchronously.
>
> Ok
>
>>
>>> If you think this is not going in a good direction, I'll leave this
>>> alone, otherwise I could try cleaning
>>> up my current experiments and propose a patch.
>>
>> No, I think this is definitely a good direction.  :)
>
> Ok, I'll try cleaning up and send you a patch. I assume this should go
> in 4.1.0.Final, so we should handle it with care.. please continue
> supporting me as this code is all new for me, I just make assumptions
> on some parts of it.
>
> Sanne
>
>>>
>>> 2010/8/26 Manik Surtani <manik at jboss.org>:
>>>> Hi Sanne
>>>>
>>>> Thanks for the detailed analysis!
>>>>
>>>> On 26 Aug 2010, at 09:05, Sanne Grinovero wrote:
>>>>
>>>>> After (finally!) being able to isolate the issue with a core-only unit
>>>>> test I changed the name of the issue to
>>>>> "AsyncStore fails to save all values when batching"
>>>>> As:
>>>>> * it's not related to the JDBC stores but the Async wrapper
>>>>> * I can only reproduce it while batching
>>>>>
>>>>> I hacked a fix for me, which appears to work fine both with the unit
>>>>> tests and fix the staging issues in my project, but this changed quite
>>>>> a bit so we should talk about it:
>>>>>
>>>>> I see the current AsyncStore implementation is "coalescing" all
>>>>> changes being part of the same transaction during the prepare-commit
>>>>> phase, and then async-delegating the same prepare-commit to the
>>>>> decorated store. After that, it might also async-delegate a commit or
>>>>> a rollback.
>>>>> So if I understood it well enough, this means that:
>>>>>
>>>>>   a) it's not merging changes from different transactions
>>>>>
>>>>>   b) a commit (or rollback) could be received from the decorated
>>>>> store implementation with a noticeable delay (a full cycle of
>>>>> collected changes), and most importantly it seems to me that it might
>>>>> receive a commit in a different thread than the prepare (is that
>>>>> legal?).
>>>>
>>>> I believe some TMs do allow prepare-and-commit on different threads.  Definitely prepare-and-rollback.
>>>>
>>>>> So I copied the transaction managing code from AbstractCacheStore (as
>>>>> AsyncStore doesn't inherit from it) and replaced the prepare, commit
>>>>> and rollback methods of AbstractCacheStore to the same as the
>>>>> AbstractCacheStore implementations:
>>>>> this way uncommitted changes are never inserted in the "state" map
>>>>> (nor are commit and rollback operations), and it seems that I end up
>>>>> sending to the delegate store only committed changes;
>>>>> best part of it is that I'm aggregating in a single "coalesced" state
>>>>> all changes from all transactions (only committed), this seems to me a
>>>>> good improvement?
>>>>
>>>> Definitely.
>>>>
>>>>> The code I'm having is not that clean, as it copied several lines of
>>>>> code from the AbstractCacheStore; to clean it up I think that the
>>>>> AbstractDelegatingStore should inherit from AbstractCacheStore, and
>>>>> avoid delegating the methods relating to transactions (in fact
>>>>> implementing what I did for the AsyncStore for all  decorated stores).
>>>>
>>>> Yes, much cleaner that way.
>>>>
>>>>> Also some unit tests will need fixing, as there are explicit tests
>>>>> asserting that the commit is done in another thread: could I delete
>>>>> these?
>>>>
>>>> Like I said we should still support commit on a separate thread.
>>>>
>>>>> Finally, as the delegated store is not being used transactionally
>>>>> anymore (it would never receive a prepare or commit), wouldn't it make
>>>>> sense to wrap all blocks of changes in a single transaction? rollback
>>>>> from a failure in the async store is, as far as I understood, not
>>>>> supported even with current code.
>>>>
>>>> Yes, true.
>>>>
>>>>> I'm eager to hear you opinions, especially what I would break with this design.
>>>>>
>>>>> Cheers,
>>>>> Sanne
>>>>>
>>>>>
>>>>> 2010/8/25 Sanne Grinovero <sanne.grinovero at gmail.com>:
>>>>>> 2010/8/25 Manik Surtani <manik at jboss.org>:
>>>>>>>
>>>>>>> On 25 Aug 2010, at 14:48, Sanne Grinovero wrote:
>>>>>>>
>>>>>>>> Hello,
>>>>>>>> I'm testing CR3, I got it so far that I'm testing the stop-restart
>>>>>>>> application cycle now, I'm using the JdbcStringBasedCacheStore in
>>>>>>>> async mode,
>>>>>>>> and every time I stop the application - which issues a clean Stop() on
>>>>>>>> the CacheManager, it appears that some values are lost.
>>>>>>>> This is fixed if I remove the async.
>>>>>>>>
>>>>>>>> Is this expected? I hope not, as this way I can't shutdown the cluster
>>>>>>>> when using async.
>>>>>>>> This time I can reproduce it with a UT.
>>>>>>>
>>>>>>> Cool, a unit test would help.  Is this only with the JDBC store?
>>>>>>
>>>>>> I understand this is not expected and created ISPN-618; you can find
>>>>>> my failing unit test attached to the issue.
>>>>>>
>>>>>> I will try to reproduce the same issue depending on core only to
>>>>>> understand if it's related to JDBC.
>>>>>>
>>>>>> Sanne
>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> Manik Surtani
>>>>>>> manik at jboss.org
>>>>>>> Lead, Infinispan
>>>>>>> Lead, JBoss Cache
>>>>>>> http://www.infinispan.org
>>>>>>> http://www.jbosscache.org
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> infinispan-dev mailing list
>>>>>>> infinispan-dev at lists.jboss.org
>>>>>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>>>>>
>>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> infinispan-dev mailing list
>>>>> infinispan-dev at lists.jboss.org
>>>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>>
>>>> --
>>>> Manik Surtani
>>>> manik at jboss.org
>>>> Lead, Infinispan
>>>> Lead, JBoss Cache
>>>> http://www.infinispan.org
>>>> http://www.jbosscache.org
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> infinispan-dev mailing list
>>>> infinispan-dev at lists.jboss.org
>>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>>
>>>
>>> _______________________________________________
>>> infinispan-dev mailing list
>>> infinispan-dev at lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>
>> --
>> Manik Surtani
>> manik at jboss.org
>> Lead, Infinispan
>> Lead, JBoss Cache
>> http://www.infinispan.org
>> http://www.jbosscache.org
>>
>>
>>
>>
>>
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>
>