Configuration

cp-member-count: Number of CP members to initialize CP Subsystem. It is 0 by default, meaning that CP Subsystem is disabled. CP Subsystem is enabled when a positive value is set. After CP Subsystem is initialized successfully, more CP members can be added at run-time and the number of active CP members can go beyond the configured CP member count. The number of CP members can be smaller than the total size of the Hazelcast cluster. For instance, you can run 5 CP members in a Hazelcast cluster of 20 members.

group-size: Number of CP members to form CP groups. If set, it must be an odd number between 3 and 7. Otherwise, cp-member-count is respected while forming CP groups.

session-time-to-live-seconds: Duration for a CP session to be kept alive after the last received heartbeat. A CP session is closed if no session heartbeat is received during this duration. Session TTL must be decided wisely. If a very low value is set, a CP session can be closed prematurely if its owner Hazelcast instance temporarily loses connectivity to CP Subsystem because of a network partition or a GC pause. In such an occasion, all CP resources of this Hazelcast instance, such as FencedLock or ISemaphore, are released. On the other hand, if a very large value is set, CP resources can remain assigned to an actually crashed Hazelcast instance for too long and liveliness problems can occur. CP Subsystem offers an API in CPSessionManagementService to deal with liveliness issues related to CP sessions. In order to prevent premature session expires, session TTL configuration can be set a relatively large value and CPSessionManagementService.forceCloseSession(String, long) can be manually called to close CP session of a crashed Hazelcast instance.

session-heartbeat-interval-seconds: Interval for the periodically-committed CP session heartbeats. A CP session is started on a CP group with the first session-based request of a Hazelcast instance. After that moment, heartbeats are periodically committed to the CP group.

missing-cp-member-auto-removal-seconds: Duration to wait before automatically removing a missing CP member from CP Subsystem. When a CP member leaves the Hazelcast cluster, it is not automatically removed from CP Subsystem, since it could be still alive and left the cluster because of a network partition. On the other hand, if a missing CP member is actually crashed, it creates a danger for its CP groups, because it will be still part of majority calculations. This situation could lead to losing majority of CP groups if multiple CP members leave the cluster over time.

fail-on-indeterminate-operation-state: Offers a choice between at-least-once and at-most-once execution of the operations on top of the Raft consensus algorithm. It is disabled by default and offers at-least-once execution guarantee. If enabled, it switches to at-most-once execution guarantee. When you invoke an API method on a CP data structure proxy, it replicates an internal operation to the corresponding CP group. After this operation is committed to majority of this CP group by the Raft leader node, it sends a response for the public API call. If a failure causes loss of the response, then the calling side cannot determine if the operation is committed on the CP group or not. In this case, if this configuration is disabled, the operation is replicated again to the CP group, and hence could be committed multiple times. If it is enabled, the public API call fails with IndeterminateOperationStateException.

name: Name of the FencedLock.

lock-acquire-limit: Maximum number of reentrant lock acquires. Once a caller acquires the lock this many times, it will not be able to acquire the lock again, until it makes at least one unlock() call.

name: Name of the CP ISemaphore.

jdk-compatible: Enables / disables JDK compatibility of the CP ISemaphore. When it is JDK compatible, just as in the j.u.c.Semaphore.release() method, a permit can be released without acquiring it first, because acquired permits are not bound to threads. However, there is no auto-cleanup of the acquired permits upon Hazelcast server / client failures. If a permit holder fails, its permits must be released manually. When JDK compatibility is disabled, a HazelcastInstance must acquire permits before releasing them and it cannot release a permit that it has not acquired. It means, you can acquire a permit from one thread and release it from another thread using the same HazelcastInstance, but not different HazelcastInstances. In this mode, acquired permits are automatically released upon failure of the holder HazelcastInstance. So there is a minor behavioral difference to the j.u.c.Semaphore.release() method.

initial-permits: Number of permits to initialize the Semaphore. If a positive value is set, the Semaphore is initialized with the given number of permits.

leader-election-timeout-in-millis: Leader election timeout in milliseconds. If a candidate cannot win the majority of the votes in time, a new election round is initiated.

leader-heartbeat-period-in-millis: Duration in milliseconds for a Raft leader node to send periodic heartbeat messages to its followers in order to denote its liveliness. Periodic heartbeat messages are actually append entries requests and can contain log entries for the lagging followers. If a too small value is set, heartbeat messages are sent from Raft leaders to followers too frequently and it can cause an unnecessary usage of CPU and network.

max-missed-leader-heartbeat-count: Maximum number of missed Raft leader heartbeats for a follower to trigger a new leader election round. For instance, if leader-heartbeat-period-in-millis is 1 second and this value is set to 5, then a follower triggers a new leader election round if 5 seconds pass after the last heartbeat message of the current Raft leader node. If this duration is too small, new leader election rounds can be triggered unnecessarily if the current Raft leader temporarily slows down or a network congestion occurs. If it is too large, it takes longer to detect failures of Raft leaders.

append-request-max-entry-count: Maximum number of Raft log entries that can be sent as a batch in a single append entries request. In Hazelcast’s Raft consensus algorithm implementation, a Raft leader maintains a separate replication pipeline for each follower. It sends a new batch of Raft log entries to a follower after the follower acknowledges the last append entries request sent by the leader.

commit-index-advance-count-to-snapshot: Number of new commits to initiate a new snapshot after the last snapshot taken by the local Raft node. This value must be configured wisely as it effects performance of the system in multiple ways. If a small value is set, it means that snapshots are taken too frequently and Raft nodes keep a very short Raft log. If snapshots are large and CP Subsystem Persistence is enabled, this can create an unnecessary overhead on I/O performance. Moreover, a Raft leader can send too many snapshots to followers and this can create an unnecessary overhead on network. On the other hand, if a very large value is set, it can create a memory overhead since Raft log entries are going to be kept in memory until the next snapshot.

uncommitted-entry-count-to-reject-new-appends: Maximum number of uncommitted log entries in the leader’s Raft log before temporarily rejecting new requests of callers. Since Raft leaders send log entries to followers in batches, they accumulate incoming requests in order to improve the throughput. You can configure this field by considering your degree of concurrency in your callers. For instance, if you have at most 1000 threads sending requests to a Raft leader, you can set this field to 1000 so that callers do not get retry responses unnecessarily.

append-request-backoff-timeout-in-millis: Timeout duration in milliseconds to apply backoff on append entries requests. After a Raft leader sends an append entries request to a follower, it will not send a subsequent append entries request either until the follower responds or this timeout occurs. Backoff durations are increased exponentially if followers remain unresponsive.