[infinispan-dev] Meaning of locking in Infinispan: ISPN-1546 and better general throughput

[Sanne Grinovero]

On 7 December 2011 16:26, Manik Surtani <manik at jboss.org> wrote:
>
> On 7 Dec 2011, at 13:27, Sanne Grinovero wrote:
>
>> So in Infinispan we have a single type of lock associated with each
>> entry; I assume it's designed like it is to save on memory, as are
>> features like lock striping, and as the reason to not support both
>> read and write locks.
>>
>> A problem I'm having more and more often, is that the "lock" is being
>> used both as something that leaks out as a way to control data
>> consistency from a "user" point of view, and as an internal artifact
>> to ensure consistent internal mutations of our data structure.
>>
>> So the same Lock instance can be acquired for long term by a user
>> process, and prevent for example the data to be passivated on a
>> CacheLoader as this will attempt to grab the same lock.
>> These should really be two different things!
>
> Yes agreed.  One is a data lock - a logical, user-centric lock on some data that should not be concurrently mutated - versus an internal, data-structure lock which prevents concurrent modification of parts of the internal data structure.  Is this what you are getting at?

Exactly. It sounds cleaner and more solid to me, and should definitely
perform better; not sure in terms of memory overhead but I guess we'll
have to see that.

>> It might bring up some confusion as well, like with the
>> FineGrainedAtomicMap: the feature I need is to be able to lock values
>> in an atomic map individually, in terms of data consistency. This does
>> NOT imply that Infinispan should not be allowed to acquire locks on
>> the AtomicMap itself for brief moments, to proceed with other internal
>> processes.. be it to create a threadsafe iterator, passivate the entry
>> to a CacheLoader or even transfer the element to a new owned in the
>> grid.
>>
>> Finally, to save memory I think that we don't need to guarantee these
>> lock instances exists all the time; granted it might be more efficient
>> to keep them as part of the Entry to avoid re-creating them too often,
>> but in some cases it might not be the case, so we might even make it
>> possible to create a different kind of Entry to optimise for specific
>> usage patterns.
>>
>> In pratical terms, it's currently quite hard to design a consistent
>> data access using Infinispan's locks from a user point of view if you
>> have to consider that Infinispan might lock the keys for it's own
>> internal needs.
>>
>> To solve ISPN-1546, I think it's totally fine to acquire a lock on the
>> FGAM for the time needed to create an iterator. But this lock needs to
>> be a different instance than the entry itself, and will be very short
>> lived, and not clustered in any way. it's just a means to guarantee we
>> can make a safe copy of the needed Array, and acquiring this lock
>> should have nothing to do with the "data experience" of preventing
>> some entries of the FGAM to be updated.
>>
>> thoughts?
>
> This is all quite a big change though.  Do we have a  JIRA for this?  Perhaps we should start my creating one, and outlining some designs on the wiki.
>
> Cheers
> Manik

Yes it's a big change, and like the "single lock owner" kind of
refactorings it will take some time; but unlike that one this can be
solved by progressive steps, having different subsystems migrate to
using dedicated locking instances.
So the first action is IMHO to design the new components needing locks
using this approach, in primis the FGAM iterator needs it's own lock
and should not interfere with the logical locks: easy, as the issue is
yet to be solved.

We need to pick some solid names. "Logical lock" sounds good for the
user visible locks ? "Internal Lock" for our own?

Sanne