runtime/me.ahoo.wow.runtime.cluster/ClusterOptions

ClusterOptions

class ClusterOptions @InternalWowApi constructor(val leaseTtl: Duration = 10.seconds, val renewInterval: Duration = 3.seconds, val memberTtl: Duration = 10.seconds, val heartbeatInterval: Duration = 3.seconds, val reconcileInterval: Duration = 1.seconds, val routeRefreshInterval: Duration = 1.seconds, val forwardTimeout: Duration = WaitPlan.DEFAULT_TIMEOUT, val maxInFlightPerPeer: Int, val drainTimeout: Duration = 10.seconds, val leaseParallelism: Int = 16, val reconcileTimeout: Duration = 1.minutes)

The tuning of partition ownership and the cluster channel on one node (–).

leaseTtl, renewed every renewInterval: how long a partition stays with a node that stopped renewing. A renewal round renews the held leases with one call, and acquisitions run at most leaseParallelism at a time: a round must stay well inside leaseTtl − renewInterval. A lease left unrenewed that long is taken as lost, so renewInterval is less than half of leaseTtl: otherwise every lease would be lost between two rounds.
memberTtl, refreshed every heartbeatInterval: how long a node stays a member after its last heartbeat. heartbeatInterval is less than half of memberTtl, so a heartbeat late by up to another interval keeps the node a member.
reconcileInterval: how often a node compares the partitions it holds with its HRW target; a change of the members it notices reconciles at once.
routeRefreshInterval: how often the route table is read from the leases and members; a failed forward refreshes it at once.
forwardTimeout: how long a forward waits for the owner's reply, capped by the command's own deadline; a forward sent and not answered by then is reported unknown, not run locally. Only a forward the owner surely did not receive runs locally. Since a timeout no longer runs the command elsewhere, waiting less than the command would only fail it sooner: the default is the default wait (WaitPlan.DEFAULT_TIMEOUT, 10 seconds), not the 1 second of earlier releases, which covered the owner's tail before falling back.
maxInFlightPerPeer: the forwards one peer may have outstanding; one more runs locally instead.
drainTimeout: how long a hand-off waits for the partition's commands in flight before it moves on.
reconcileTimeout: how long one reconcile may run before the node takes itself as stuck: it stops heartbeating and leaves the members until the reconcile ends, so the others take the partitions it cannot run. Well above what a hand-off takes: the drain, then the stop of each runtime; longer than drainTimeout at least.

Fields may be added: Wow builds it from the host's configuration, so the constructor is internal to Wow.

Constructors

ClusterOptions

@InternalWowApi

constructor(leaseTtl: Duration = 10.seconds, renewInterval: Duration = 3.seconds, memberTtl: Duration = 10.seconds, heartbeatInterval: Duration = 3.seconds, reconcileInterval: Duration = 1.seconds, routeRefreshInterval: Duration = 1.seconds, forwardTimeout: Duration = WaitPlan.DEFAULT_TIMEOUT, maxInFlightPerPeer: Int, drainTimeout: Duration = 10.seconds, leaseParallelism: Int = 16, reconcileTimeout: Duration = 1.minutes)