SWIM: The scalable membership protocol
In a distributed system we have a group of nodes that need to collaborate and send messages to each other. To achieve that they need to first answer a simple question: Who are my peers?
That’s what membership protocols do. They help each node in this system to maintain a list of nodes that are alive, notifying them when a new node joins the group, when someone intentionally leaves and when a node dies (or at least appears to be dead).
SWIM, or Scalable Weakly-consistent Infection-style Process Group Membership Protocol, is one of these protocols.
Understanding the name
Scalable Weakly-consistent Infection-style Process Group Membership Protocol is quite a long name for a protocol that seems to do such a simple thing, so let’s break down the name and understand what each piece means. This will also help us understand why this protocol was created in the first place.
Scalable: Prior to
SWIM most membership protocols used a heart-beating approach, that is, each node would send a heartbeat (i.e. an empty message that just means “I’m alive!”) to every other node in the cluster, every interval
T. If a node
N1 doesn’t receive a heartbeat from node
N2 after a certain period, it declares this node dead. This works fine for a small cluster, but as the number of nodes in this cluster increases, the number of messages that need to be sent increases quadratically. If you have 10 nodes, it may be fine to send 100 messages every second, but with 1,000 nodes you would need to send 1,000,000 messages, and that would not scale very well.
Weakly-consistent: That means that at a given point in time different nodes can have a different view of the world. They will eventually converge to the same state but we cannot expect strong consistency.
Infection-style: That’s what is also commonly known as a gossip or epidemic protocol. It means that a node shares some information with a subset of its peers, that then share this information with a subset of its own peers, until the entire cluster receives that information. That means a node doesn’t need to send a message to all of its peers, it just tells that to a few nodes that will gossip about it.
Membership: Well, that basically means we will ultimately answer the question “Who are my peers?”
Heart-beating protocols usually solve 2 different problems with the heartbeat messages: They detect when a node fails (because it stops sending the heartbeat), and they keep the list of the peers in the cluster (that is, every node that is sending a heartbeat).
SWIM decided to take the novel approach of dividing these 2 problems in different components, so it has a failure detection and a dissemination module.
Each node in the cluster will choose a node at random (say,
N2) and will send a
ping message, expecting to receive an
ack back. This is simply a probe message and in normal circumstances it would receive this
ack message and confirm that
N2 is still alive.
When that doesn’t happen, though, instead of immediately marking this node as dead, it will try to probe it through other nodes. It will randomly select
k other nodes from its membership list and send a
This helps to prevent false-positives when for some reason
N1 cannot get a response directly from
N2 (maybe because there’s a network congestion between the two), but the node is still alive and accessible by
If the node cannot be accessed by any of the
k members, though, it’s marked as dead.
Upon detecting a node as dead, the protocol can then just multicast this information to all the other nodes in the cluster, and each node would remove
N2 from its local list of peers. Information about nodes that are voluntarily leaving or joining the cluster can be multicast in a similar way.
Making SWIM more robust
This makes for a quite simple protocol, but there a few modification that the original SWIM paper suggests to make it more robust and efficient, they are:
- Change the dissemination component to use an infection-style approach, instead of multicasting (after all, that’s in the name of the protocol);
- Use a suspicion mechanism for the failure detection to reduce the false positives;
- Use a round-robin probe target selection instead of randomly selecting nodes.
Let’s explore each of these points to understand what they mean and why that would be an improvement over the protocol that we have discussed thus far.
There are (at least) two issues that we need to be aware of when using this multicast primitive to disseminate information:
- IP multicast, although generally available, is usually not enabled in most environments. For example, if you are running on Amazon VPC, you are out of luck. You would then need to use a pretty inefficient point-to-point solution;
- Even if you can use this type of multicast, it will usually use
UDP, that’s a best-effort protocol, meaning the network can (and probably will) drop packets, making it hard to maintain a reliable membership list.
A better (and, in my opinion, quite elegant) solution suggested in the SWIM paper is to forget about this multicast idea, and instead use the
ack messages that we use for failure detection to piggyback the information we need to disseminate. We are not adding any new messages, just leveraging the messages that we already send, and “reusing” them to also transport some information about membership updates.
Suspicion mechanism for failure detection
Another optimization is to first suspect a node is dead, before declaring it dead. The goal here is to minimize false positives, as it is usually preferable to take longer to detect a failed node than it is to wrongly mark a healthy node as dead. It is a trade-off, though, and depending on the specific case this might not make sense.
It works like this: When a node
N1 cannot receive an
ack message from node
N2 (neither directly, through a
ping message, nor indirectly, through a
ping-req message), instead of disseminating that
N2 is dead and should be removed from the membership list, it just disseminates that it suspects that
N2 is dead.
This suspected node is treated like a non-faulty node for all effects, and it keeps receiving
ping messages like any other member. If any node can get an
N2, it’s marked again as alive and this information is disseminated.
N2 itself can also receive a message saying it’s suspected to be dead and tell the group that they are wrong and that it never felt better.
If after a predefined timeout we don’t hear from
N2, it’s then confirmed to be dead and this information is disseminated.
Round-Robin probe target selection
In the original protocol definition the node that is selected to be probed (i.e. the node we send a
ping message, expecting an
ack in return) is picked at random order. Although we can guarantee that eventually a node failure will be detected by every non-faulty node, this can take a relatively long time if we are out of luck with our target selection. A way to minimize this problem is to maintain a list of the members we want to probe, and go through this list in a round-robin fashion, and new joiners can be added at a random position in this list.
Using this approach we can have a time-bounded failure detection, where in the worst case it will take
probe interval * number of nodes to select this faulty node.
SWIMis a membership protocol that helps us know which nodes are part of a cluster, maintaining an updated list of healthy peers;
- It divides the membership problem into 2 components: Failure detection and dissemination;
- The failure detection component works by selecting a random node and sending it a
pingmessage, expecting an
- If it does not receive an
ack, it selects
kpeers to probe this node through a
- An optimization of this failure detection is to first mark a node as “suspected”, and mark it as dead just after a timeout;
- We can disseminate membership information by piggybacking the failure detection messages (
ack) instead of using a multicast primitive;
- A way to improve the failure-detection time is to select nodes in a round-robin fashion instead of selecting nodes at random.
The original paper is short and very readable.
Serf, a HashiCorp tool, uses
SWIM and has a nice overview in their documentation. They also talk about some modifications they made to increase propagation speed and convergence rates, and the code is open source.
Armon Dadgar, one of HashiCorp’s co-founders, also gave a very nice talk explaining the protocol and these modifications:
Brian Storti joined AlphaSights in July of 2015 and serves as a software engineer.