How to aquire cluster-wide lock for non-transactional cache?

[yelin66]

I was testing with simple non-transactional cache with synchronous replication. I tried to update randomly from two cache instances on a range of keys, and there is a big chance for a deadlock. It seems when the two keys being updated fall into the same lock striping, both of them acquired the local lock but failed remotely during replication. I was using the default configuration - that's useLockStriping=true & concurrencyLevel=512  When I disabled  useLockStriping, the chance reduced as I guess it would only happens when they access the exact same key. I would like to lock the key cluster-wide to prevent it, but the cache.lock(key) doesn't apply to non-transactional context. Could you please suggest how to prevent such kind of deadlock?

[manik]

You would need to catch your exception and retry unfortunately.  If you were using transactions, though, you could use deadlock detection.

[mircea.markus]

incresing the concurrencyLevel would reduce the chance of lock overlaps, but as manik suggested I think retrying should be the solution for general case.


[yelin66]

The default concurrencyLevel is 512, and my test has access from only two threads in two separate instances. I also tried to have different prefix for the keys I access from different threads. So basically there is no key collision. However, it happens quite frequently. Could you please suggest why and how to avoid it?

[mircea.markus]

you won't be able to get rid of timeouts even if you disable lock striping; (but this would reduce the chance for this to happen, as you said).

beside disabling lock stipping, you can also increse replication timeout (try 20000 for example).