Hi Radim

First of all, I don't think this is feasible. For example, read-committed vs repeatable read changes how the entries are stored in the transaction context, so you can't have a repeatable-read get() in the same transaction after a read-committed get. Write skew check also requires versions, so you couldn't skip updating the version in any optimistic cache just in case some transaction might need it in the future.

We also can't mix non-transactional, transactional asynchronous, and transactional synchronous operations on the same cache, as they would break each other's consistency. In fact, Infinispan 4.x allowed both transactional and non-transactional operations on the same cache, but at some point we realized that there's no way to ensure the consistency of transactions if there are overlapping with non-transactional operations.

I agree that the configuration is very tightly coupled with the code that uses it, so settings that can break the application should be more obvious. We should discuss how we can improve this at the clustering meeting in Berlin.

But I think forgetting to add a flag in some part of the application is just as likely as the administrator making a mistake in the configuration, and having different consistency models in the same cache can also make code harder to understand. So instead of allowing flags to control consistency, I would rather add methods for the user to assert that the cache has certain properties.

Cheers

Dan

On Fri, Nov 21, 2014 at 12:38 PM, Radim Vansa <rvansa@redhat.com> wrote:

Hi,

when thinking about strong/eventual consistency and ease of
configuration, I was considering whether cache configuration should
affect results of operations at all (one example could be read
committed/repeatable read, or write skew check).

It would seem to me that the configuration would be simpler, and user
options more rich if those options that change the result of operation
would be purely API-wise (based on flags or method arguments) and the
configuration could only change the performance (defining cache store
will slow down some operations) or availability of these operations (you
cannot start a transaction when the manager is not defined), not the
outcome.

E.g. is there really a point to be able to change sync/async
configuration of the cache when the code expects strong consistency? If
it can handle that, it should grab cache.withFlags(FORCE_ASYNCHRONOUS)
and work on that.
Another example is in the strong/eventual consistency - if I want to see
the cache as strongly consistent, I can't read from backup owners [1].
Currently there is no option to force reading from primary owner,
therefore, I was wondering whether it should be configurable (together
with staggered gets policy - not that this would be implemented) or
whether that should be specified as a flag - and it seems to me that it
should not be configurable as the administrator could remove the flag
from the config (and see increased performance) but eventually a race
could occur where this flag matters and the application will behave
incorrectly.

WDYT? This question is obviously rather for changes on the roadmap (I'd
say along with leaving ConcurrentMap interface) than any immediate
actions in versions 7.x or 8.x.

Radim

[1] https://issues.jboss.org/browse/ISPN-4995

--
Radim Vansa <rvansa@redhat.com>
JBoss DataGrid QA

_______________________________________________
infinispan-dev mailing list
infinispan-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev