Skip to main content

Partition Quiescence

Partition quiescence reduces idle heartbeat traffic.

In normal Raft, a leader sends periodic heartbeats to every follower for every partition it leads. That is useful while a partition is active, but expensive when a cluster has many idle partitions.

Quiescence lets an idle partition go quiet without giving up safe failover.

What Quiescence Does​

When a partition leader sees that a partition has been idle for long enough, it sends one final quiesce marker to followers and then stops sending periodic heartbeats for that partition.

Followers that receive the marker stop treating missing per-partition heartbeats as a reason to start an election. Instead, they use SWIM node liveness to decide whether the leader's node is still alive.

The leader is still the leader. The partition is only quiet.

For the failure detector itself, see SWIM Failure Detection.

Why This Matters​

Without quiescence, heartbeat traffic grows with both:

  • node count
  • partition count.

For a cluster with many mostly idle partitions, that can produce a large amount of network and CPU work that carries no application data.

Quiescence changes the idle case from "heartbeat every partition forever" to "watch node liveness once through SWIM."

How A Partition Quiesces​

A leader may quiesce a partition when all of these are true:

  • EnableQuiescence is true
  • the partition is not already quiesced
  • there are no active proposals
  • the partition has been idle longer than QuiesceAfter.

When that happens, the leader sends an AppendLogs message with Quiesce = true. Followers set their local quiesced flag for that partition.

Quiescence is local runtime state. It is not a replicated application log entry and it does not change quorum math.

How A Partition Wakes Up​

Any real write wakes the partition.

On the leader:

  1. a client proposal arrives
  2. the leader clears the local quiesced state
  3. the entry is replicated normally
  4. periodic heartbeats resume.

On followers, any normal append with Quiesce = false clears the quiesced state.

After the burst of writes ends and the partition stays idle past QuiesceAfter, it can quiesce again.

Failover While Quiesced​

Quiesced followers do not rely on the per-partition election timeout while the leader's node is still considered alive.

Instead, they watch SWIM node state:

  • Alive: stay quiet
  • Suspect or Dead: un-quiesce and start the normal pre-vote/election path.

Failover is triggered on Suspect, not Dead, so it is tied to roughly one PingInterval rather than waiting for the full SuspicionTimeout.

Timing Requirements​

Quiescence depends on SWIM.

When EnableQuiescence = true, Kommander validates:

  • PingInterval > 0
  • PingInterval < StartElectionTimeout.

If SWIM is disabled, a quiesced follower would have no signal that the leader's node died. If PingInterval is too high, failover from a quiesced partition would be slower than ordinary election timeout.

The defaults are valid:

  • EnableQuiescence = true
  • QuiesceAfter = 1500 ms
  • PingInterval = 1 s
  • StartElectionTimeout = 2000 ms.

Configuration​

PropertyDefaultDescription
EnableQuiescencetrueEnables idle partition quiescence. Set to false to keep classic per-partition heartbeats on every interval.
QuiesceAfter1500 msHow long a partition must be idle, with no active proposals, before the leader sends a quiesce marker and suppresses heartbeats.
PingInterval1 sSWIM probe cadence. Must be greater than 0 and lower than StartElectionTimeout when quiescence is enabled.
SuspicionTimeout5 sTime from Suspect to Dead. Quiesced failover starts on Suspect, so this does not directly gate quiesced failover latency.
StartElectionTimeout2000 msLower election timeout bound. PingInterval must be below this while quiescence is enabled.

Operational Notes​

  • Quiescence is useful when you run many partitions and many of them are idle.
  • A quiesced partition should wake immediately when a write arrives.
  • If a node stays SWIM-alive but one specific partition executor stops making progress, quiesced followers may not elect a replacement for that partition. This is a known limitation of using node-level liveness for idle partition suppression.
  • Set EnableQuiescence = false in tests or deployments where you intentionally disable SWIM with PingInterval = 0.