A Brief Overview of Consistency and Replication in Distributed Systems

Partitioning and replication are the two common techniques used together in distributed databases to achieve scalable, available and transparent data distribution. The data space is divided into partitions, each of which contains a distinct portion of the overall data set. For these partitions, multiple copies called replicas are created. Partition replicas are distributed among the cluster members. Each member is assigned to at most a single replica for a partition. In this setting, different replication techniques can be used to access the data and keep the replicas in sync on updates. The technique being used directly affects the guarantees and properties a distributed data store provides, due to the CAP (Consistency, Availability and Partition Tolerance) principle.

One aspect of replication techniques is about where a replicated data set is accessed and updated. For instance, primary-copy systems first elect a replica, which can be called as primary, master, etc., and use that replica to access the data. Changes in the data on the primary replica are propagated to other replicas. This approach has different namings, such as primary-copy, single-master, passive replication. The primary-copy technique is a powerful model as it prevents conflicts, deadlocks among the replicas. However, primary replicas can become bottlenecks. On the other hand, we can have a different technique by eliminating the primary-copy and treating each replica as equal. These systems can achieve a higher level of availability as a data entry can be accessed and updated using any replica. However, it can become more difficult to keep the replicas in sync with each other.

Replication techniques also differ in how updates are propagated among replicas.

  • One option is to update each replica as part of a single atomic transaction, called as eager replication or synchronous replication. Consensus algorithms apply this approach to achieve strong consistency on a replicated data set. The main drawback is the amount of coordination and communication required while running the replication algorithm. CP systems implement consensus algorithms under the hood.
  • Another option is the lazy replication technique, which is also called as asynchronous replication. Lazy replication algorithms execute updates on replicas with separate transactions. They generally work with best-effort. By this way, the amount of coordination among the replicas are degraded and data can be accessed in a more performant manner. Yet, it can happen that a particular update is executed on some replicas but not on others, which will lead replicas to diverge. Such problems can be resolved with different approaches, such as read-repair, write-repair, anti-entropy. Lazy replication techniques are popular among AP systems.

Hazelcast's Replication Algorithm

The discussion here generally applies to any system that maintains multiple copies of a data set. It applies to Hazelcast as well. In the context of CAP principle, Hazelcast is an AP product, which employs the combination of primary-copy and lazy replication techniques. As briefly described in the Data Partitioning section, each data entry is mapped to a single Hazelcast partition and put into replicas of that partition. One of the replicas is elected as the primary replica, which is responsible for performing operations on that partition. When you read or write a map entry, you transparently talk to the Hazelcast member to which primary replica of the corresponding partition is assigned. By this way, each request hits the most up-to-date version of a particular data entry in a stable cluster. Backup replicas stay in standby mode until the primary replica fails. Upon failure of the primary replica, one of the backup replicas are promoted to the primary role.

With lazy replication, when the primary replica receives an update operation for a key, it executes the update locally, and propagates it to backup replicas. It marks each update with a logical timestamp so that backups apply them in the correct order and converge to the same state with the primary. Backup replicas can be used to scale reads (see the Enabling Backup Reads section) with no strong consistency but monotonic reads guarantee.

Hazelcast offers features such as Quorum, ILock, AtomicLong, etc. In the journey of being a highly elastic, dynamic and easy to use product, Hazelcast tries to provide best-effort consistency guarantees without being a complete CP solution. Therefore, we recommend these features to be used for efficiency purposes in general, instead of correctness. For instance, they can be used to prevent to run a resource-extensive computation multiple times, which would not create any correctness problem if runs more than once. Please refer to Best-Effort Consistency and Network Partitioning sections for more information.

Best-Effort Consistency

Hazelcast's replication technique enables Hazelcast clusters to offer high throughput. However, due to temporary situations in the system, such as network interruption, backup replicas can miss some updates and diverge from the primary. Backup replicas can also hit long GC pauses or VM pauses, and fall behind the primary, which is a situation called as replication lag. If a Hazelcast partition primary replica member crashes while there is a replication lag between itself and the backups, strong consistency of the data can be lost.

Please note that CP systems can have similar problems as well. However, in a CP system, once a replica performs an update locally (i.e., commits the update), the underlying consensus algorithm guarantees durability of the update for the rest of the execution.

On the other hand, in AP systems like Hazelcast, a replica can perform an update locally, even if the update is not to be performed on other replicas. This is a fair trade-off to reduce amount of coordination among replicas, and maintain high throughput & high availability of the system. These systems employ additional measurements to maintain consistency in a best-effort manner. In this regard, Hazelcast tries to minimize effect of such scenarios using an active anti-entropy solution as follows:

  • Each Hazelcast member runs a periodic task in the background.
  • For each primary replica it is assigned, it creates a summary information and sends it to the backups.
  • Then, each backup member compares the summary information with its own data to see if it is up-to-date with the primary.
  • If a backup member detects a missing update, it triggers the synchronization process with the primary.

Invocation Lifecycle

When a write is requested with the methods, such as map.put() or queue.offer(), a write operation is submitted to the Hazelcast member that owns the primary replica of the specific partition. Partition of an operation is determined based on a parameter (key of an entry or name of the data structure, etc.) related to that operation depending on the data structure. Target Hazelcast member is figured out by looking up a local partition assignment/ownership table, which is updated on each partition migration and broadcasted to all cluster eventually.

When a Hazelcast member receives a partition specific operation, it executes the operation and propagates it to backup replica(s) with a logical timestamp. Number of backups for each operation depends on the data structure and its configuration. See Threading Model - Operation Threading for threading details.

Two types of backup replication are available: sync and async. Despite what their names imply, both types are still implementations of the lazy (async) replication model. The only difference between sync and async is that, the former makes the caller block until backup updates are applied by backup replicas and acknowledgments are sent back to the caller, but the latter is just fire & forget. Number of sync and async backups are defined in the data structure configurations, and you can use a combination of sync and async backups.

When backup updates are propagated, response of the execution including number of sync backup updates is sent to the caller and after receiving the response, caller waits to receive the specified number of sync backup acknowledgements for a predefined timeout. This timeout is 5 seconds by default and defined by the system property hazelcast.operation.backup.timeout.millis. (See System Properties section.)

A backup update can be missed because of a few reasons, such as a stale partition table information on a backup replica member, network interruption, or a member crash. That's why sync backup acks require a timeout to give up. Regardless of being a sync or async backup, if a backup update is missed, the periodically running anti-entropy mechanism detects the inconsistency and synchronizes backup replicas with the primary. Also the graceful shutdown procedure ensures that all backup replicas for partitions whose primary replicas are assigned to the shutting down member will be consistent.

In some cases, although the target member of an invocation is assumed to be alive by the failure detector, the target may not execute the operation or send the response back in time. Network splits, long pauses caused by high load, GC or IO (disk, network) can be listed as a few possible reasons. When an invocation doesn't receive any response from the member that owns primary replica, then invocation fails with an OperationTimeoutException. This timeout is 2 minutes by default and defined by the system property hazelcast.operation.call.timeout.millis. (See System Properties section.) When timeout is passed, result of the invocation will be indeterminate.

Exactly-once, At-least-once or At-most-once Execution

Hazelcast, as an AP product, does not provide the exactly-once guarantee. In general, Hazelcast tends to be an at-least-once solution.

In the following failure case, exactly-once guarantee can be broken:

  • When the target member of a pending invocation leaves the cluster while the invocation is waiting for a response, that invocation is re-submitted to its new target due to the new partition table. It can be that, it has already been executed on the leaving member and backup updates are propagated to the backup replicas, but the response is not received by the caller. If that happens, the operation will be executed twice.

In the following failure case, invocation state will become indeterminate:

  • As explained above, when an invocation does not receive a response in time, invocation will fail with an OperationTimeoutException. This exception does not say anything about outcome of the operation, that means operation may not be executed at all, it may be executed once or twice (due to member left case explained above).


As described in the Invocation Lifecycle section, for partition-based mutating invocations, such as map.put(), a caller waits with a timeout for the operation that is executed on corresponding partition's primary replica and backup replicas, based on the sync backup configuration of the distributed data structure. Hazelcast 3.9 introduces a new mechanism to detect indeterminate situations while making such invocations. If hazelcast.operation.fail.on.indeterminate.state system property is enabled, a mutating invocation throws IndeterminateOperationStateException when it encounters the following cases:

  • The operation fails on partition primary replica member with MemberLeftException. In this case, the caller may not determine the status of the operation. It could happen that the primary replica executes the operation, but fails before replicating it to all the required backup replicas. Even if the caller receives backup acks from some backup replicas, it cannot decide if it has received all required ack responses, since it does not know how many acks it should wait for.

  • There is at least one missing ack from the backup replicas for the given timeout duration. In this case, the caller knows that the operation is executed on the primary replica, but some backup may have missed it. It could be also a false-positive, if the backup timeout duration is configured with a very small value. However, Hazelcast's active anti-entropy mechanism eventually kicks in and resolves durability of the write on all available backup replicas as long as the primary replica member is alive.

When an invocation fails with IndeterminateOperationStateException, the system does not try to rollback the changes which are executed on healthy replicas. Effect of a failed invocation may be even observed by another caller, if the invocation has succeeded on the primary replica. Hence, this new behavior does not guarantee linearizability. However, if an invocation completes without IndeterminateOperationStateException when the configuration is enabled, it is guaranteed that the operation has been executed exactly-once on the primary replica and specified number of backup replicas of the partition.

Please note that IndeterminateOperationStateException does not apply to read-only operations, such as map.get(). If a partition primary replica member crashes before replying to a read-only operation, the operation is retried on the new owner of the primary replica.