» Gossip Protocol

Serf uses a gossip protocol to broadcast messages to the cluster. This page documents the details of this internal protocol. The gossip protocol is based on "SWIM: Scalable Weakly-consistent Infection-style Process Group Membership Protocol", with a few minor adaptations, mostly to increase propagation speed and convergence rate.

» SWIM Protocol Overview

Serf begins by joining an existing cluster or starting a new cluster. If starting a new cluster, additional nodes are expected to join it. New nodes in an existing cluster must be given the address of at least one existing member in order to join the cluster. The new member does a full state sync with the existing member over TCP and begins gossiping its existence to the cluster.

Gossip is done over UDP with a configurable but fixed fanout and interval. This ensures that network usage is constant with regards to number of nodes. Complete state exchanges with a random node are done periodically over TCP, but much less often than gossip messages. This increases the likelihood that the membership list converges properly since the full state is exchanged and merged. The interval between full state exchanges is configurable or can be disabled entirely.

Failure detection is done by periodic random probing using a configurable interval. If the node fails to ack within a reasonable time (typically some multiple of RTT), then an indirect probe is attempted. An indirect probe asks a configurable number of random nodes to probe the same node, in case there are network issues causing our own node to fail the probe. If both our probe and the indirect probes fail within a reasonable time, then the node is marked "suspicious" and this knowledge is gossiped to the cluster. A suspicious node is still considered a member of cluster. If the suspect member of the cluster does not dispute the suspicion within a configurable period of time, the node is finally considered dead, and this state is then gossiped to the cluster.

This is a brief and incomplete description of the protocol. For a better idea, please read the SWIM paper in its entirety, along with the Serf source code.

» SWIM Modifications

As mentioned earlier, the gossip protocol is based on SWIM but includes minor changes, mostly to increase propagation speed and convergence rates.

The changes from SWIM are noted here:

• Serf does a full state sync over TCP periodically. SWIM only propagates changes over gossip. While both are eventually consistent, Serf is able to more quickly reach convergence, as well as gracefully recover from network partitions.

• Serf has a dedicated gossip layer separate from the failure detection protocol. SWIM only piggybacks gossip messages on top of probe/ack messages. Serf uses piggybacking along with dedicated gossip messages. This feature lets you have a higher gossip rate (for example once per 200ms) and a slower failure detection rate (such as once per second), resulting in overall faster convergence rates and data propagation speeds.

• Serf keeps the state of dead nodes around for a set amount of time, so that when full syncs are requested, the requester also receives information about dead nodes. Because SWIM doesn't do full syncs, SWIM deletes dead node state immediately upon learning that the node is dead. This change again helps the cluster converge more quickly.

» Lifeguard Enhancements

SWIM makes the assumption that the local node is healthy in the sense that soft real-time processing of packets is possible. However, in cases where the local node is experiencing CPU or network exhaustion this assumption can be violated. The result is that the node health can occasionally flap, resulting in false monitoring alarms, adding noise to telemetry, and simply causing the overall cluster to waste CPU and network resources diagnosing a failure that may not truly exist.

Serf 0.8 added Lifeguard, which completely resolves this issue with novel enhancements to SWIM.

The first extension introduces a "nack" message to probe queries. If the probing node realizes it is missing "nack" messages then it becomes aware that it may be degraded and slows down its failure detector. As nack messages begin arriving, the failure detector is sped back up.

The second change introduces a dynamically changing suspicion timeout before declaring another node as failured. The probing node will initially start with a very long suspicion timeout. As other nodes in the cluster confirm a node is suspect, the timer accelerates. During normal operations the detection time is actually the same as in previous versions of Serf. However, if a node is degraded and doesn't get confirmations, there is a long timeout which allows the suspected node to refute its status and remain healthy.

These two mechanisms combine to make Serf much more robust to degraded nodes in a cluster, while keeping failure detection performance unchanged. There is no additional configuration for Lifeguard, it tunes itself automatically.

For more details about Lifeguard, please see the Making Gossip More Robust with Lifeguard blog post, which provides a high level overview of the HashiCorp Research paper Lifeguard : SWIM-ing with Situational Awareness.

» Serf-Specific Messages

On top of the SWIM-based gossip layer, Serf sends some custom message types.

Serf makes heavy use of Lamport clocks to maintain some notion of message ordering despite being eventually consistent. Every message sent by Serf contains a Lamport clock time.

When a node gracefully leaves the cluster, Serf sends a leave intent through the gossip layer. Because the underlying gossip layer makes no differentiation between a node leaving the cluster and a node being detected as failed, this allows the higher level Serf layer to detect a failure versus a graceful leave.

When a node joins the cluster, Serf sends a join intent. The purpose of this intent is solely to attach a Lamport clock time to a join so that it can be ordered properly in case a leave comes out of order.

For custom events and queries, Serf sends either a user event, or user query message. This message contains a Lamport time, event name, and event payload. Because user events are sent along the gossip layer, which uses UDP, the payload and entire message framing must fit within a single UDP packet.

UserEventSizeLimit can be configured, but a hard limit of 9KB is applied. It's up to the user to make sure that the "user event"'s network transmission "path" fits their MTU and/or other packet constraints.