HazelcastInstance.getFlakeIdGenerator(String)
for an alternative implementation which does not suffer
from this problem.@Deprecated public interface IdGenerator extends DistributedObject
long
IDs in a cluster, under normal operating scenarios.
In theory a IAtomicLong.incrementAndGet()
could be used to provide the same functionality.
The big difference is that the incrementAndGet
requires one or more remote calls for every invocation and therefore
is a performance and scalability bottleneck. The IdGenerator uses an IAtomicLong
under the hood, but instead of
doing a remote call for every call to newId()
, it does it less frequently. It checks out a chunk of 10,000 IDs
and as long as it has not yet consumed all the IDs in its chunk, no remote call is done.
It can be that IDs generated by different cluster members will get out of order because each member will get its own chunk. It can be that member 1 has chunk 1..10000 and member 2 has 10001..20000. Therefore, member 2 will automatically have IDs that are out of order with the IDs generated by member 1. Next, the chunk is kept until all IDs are used.
It should be noted that IdGenerator does not guarantee unique IDs in a split brain situation. During a split brain,
different partitions of the cluster may have the primary or back-up, and will therefore continue in their respective
clusters from the last known value. Other partitions of the cluster may not have the original primary nor the
back-up, in which case the IdGenerator
will be created as new and initialised as per the provided
init(long)
When a cluster heals after a split brain, a default large side merge takes place for the IAtomicLong
that keeps track of the last issued block of IDs, therefore it is possible that already issued IDs could be issued
again after the cluster is repaired. As a precaution when in a new record insert situation try to detect if the key
already exists and then grab another, possibly executing this logic only after a known split brain merge has occurred.
Care should be taken if you are using the IdGenerator to provide keys for IMap
or
ICache
, as after split brain it could be possible for two unique values to share
the same key. When a MapMergePolicy
is applied, one of these unique values will
have to be discarded. Unless configured, the default merge policy is
PutIfAbsentMapMergePolicy
.
It is NOT RECOMMENDED to use IdGenerator
to provide keys that are also used as unique identifiers in
underlying persistent storage, for example in an IMap
that writes to a relational database using MapStore
.
Modifier and Type | Method and Description |
---|---|
boolean |
init(long id)
Deprecated.
Tries to initialize this
IdGenerator instance with the given ID. |
long |
newId()
Deprecated.
Generates and returns a cluster-wide unique ID.
|
destroy, getName, getPartitionKey, getServiceName
boolean init(long id)
IdGenerator
instance with the given ID.
The next generated ID will be 1 greater than the supplied ID.
true
if initialization succeeded, false
otherwise.long newId()
Copyright © 2019 Hazelcast, Inc.. All Rights Reserved.