Smudge is a minimalist Go implementation of the SWIM (Scalable Weakly-consistent Infection-style Membership) protocol for cluster node membership, status dissemination, and failure detection developed at Cornell University by Motivala, et al. It isn't a distributed data store in its own right, but rather a framework intended to facilitate the construction of such systems.
Smudge also extends the standard SWIM protocol so that in addition to the standard membership status functionality it also allows the transmission of broadcasts containing a small amount (256 bytes) of arbitrary content to all present healthy members. This maximum is related to the limit imposed on maximum safe UDP packet size by RFC 791 and RFC 2460. We recognize that some systems allow larger packets, however, and although that can risk fragmentation and dropped packets the maximum payload size is configurable.
Smudge was conceived with space-sensitive systems (mobile, IoT, containers) in mind, and therefore was developed with a minimalist philosophy of doing a few things well. As such, its feature set is relatively small and mostly limited to functionality around adding and removing nodes and detecting status changes on the cluster.
Complete documentation is available from the associated Godoc.
- Uses gossip (i.e., epidemic) protocol for dissemination, the latency of which grows logarithmically with the number of members.
- Low-bandwidth UDP-based failure detection and status dissemination.
- Imposes a constant message load per group member, regardless of the number of members.
- Member status changes are eventually detected by all non-faulty members of the cluster (strong completeness).
- Supports transmission of short broadcasts that are propagated at most once to all present, healthy members.
- Supports both IPv4 and IPv6.
- Pluggable logging
- Broadcasts are limited to 256 bytes, or 512 bytes when using IPv6.
- No WAN support: only local-network, private IPs are supported.
Deviations from Motivala, et al
- Dead nodes are not immediately removed, but are instead periodically re-tried (with exponential backoff) for a time before finally being removed.
- Smudge allows the transmission of short, arbitrary-content broadcasts to all healthy nodes.
TL;DR a broadcast can be added to the local node by either calling a function or by receiving it from a remote node. A broadcast is send to other nodes a couple of times, piggybacked on membership messages. Then after a while the broadcast is removed from the node.
The emit counter represents the number of times a broadcast message must be send to other nodes. An emit counter is calculated with the following formula: int(2.5 * log(number of nodes) + 0.5). The larger the network the higher the emit counter will be, but the larger the network the slower the emit counter will grow.
Examples:
- 2 nodes:
int(2.5 * log(2) + 0.5) = 2 - 10 nodes:
int(2.5 * log(10) + 0.5) = 6 - 20 nodes:
int(2.5 * log(20) + 0.5) = 8
When a broadcast is added to Smudge, either because it is added locally (by calling a function of the library) or is received from a remote node, an emit counter initialized with the formula above. The emit counter and the broadcast are then saved to a local buffer.
The emit counter is used to track how many times a broadcast must be send to other nodes in the network. When the emit counter gets below a certain, large negative, thresh-hold the broadcast is removed from the buffer. Only broadcasts with a positive emit counter will be send when they are selected.
When Smudge is about to send a membership message it looks for the broadcast with the largest emit counter. If multiple broadcasts have the same emit counter value, one is arbitrarily chosen. The selected broadcast can have a negative emit counter. If the emit counter is larger then 0 Smudge adds that broadcast to the membership message that will be send. In any case the emit counter is lowered by 1.
When a broadcast is received from another node and that broadcast is already in the buffer it will be ignored. To achieve this the origin IP of the node that added the broadcast to the network is saved as part of the broadcast.
Although Smudge is intended to be directly extended, a Dockerfile is provided for testing and proofs-of-function.
The Dockerfile uses a multi-stage build, so Docker 17.05 or higher is required. The build compiles the code in a dedicated Golang container and drops the resulting binary into a scratch image for execution. This makes a Makefile or build.sh largely superfluous and removed the need to configure a local environment.
To run the tests in a containerized environment, which requires only that you have Docker installed (not Go), you can do:
make testOr, if you'd rather not use a Makefile:
go test -v github.com/clockworksoul/smudgeTo execute the build, you simply need to do the following:
make imageOr, if you'd rather not use a Makefile:
docker build -t clockworksoul/smudge:latest .You can test Smudge locally using the Docker image. First create a network to use for your Smudge nodes and then add some nodes to the network.
For IPv4 you can use the following commands:
docker network create smudge
docker run -i -t --network smudge --rm clockworksoul/smudge:latest /smudge
# you can add nodes with the following command
docker run -i -t --network smudge --rm clockworksoul/smudge:latest /smudge -node 172.20.0.2To try out Smudge with IPv6 you can use the following commands:
docker network create --ipv6 --subnet fd02:6b8:b010:9020:1::/80 smudge6
docker run -i -t --network smudge6 --rm clockworksoul/smudge:latest /smudge
# you can add nodes with the following command
docker run -i -t --network smudge6 --rm clockworksoul/smudge:latest /smudge -node [fd02:6b8:b010:9020:1::2]:9999If you already have a $GOPATH set up, you can skip to the following section.
First, you'll need to decide where your Go code and binaries will live. This will be your Gopath. You simply need to export this as GOPATH:
export GOPATH=~/go/Change it to whatever works for you. You'll want to add this to your .bashrc or .bash_profile.
Clone the code into $GOPATH/src/github.com/clockworksoul/smudge. Using the full-qualified path structure makes it possible to import the code into other libraries, as well as Smudge's own main() function.
git clone git@github.com:clockworksoul/smudge.git $GOPATH/src/github.com/clockworksoul/smudgeOnce you have a $GOPATH already configured and the repository correctly cloned into $GOPATH/src/github.com/clockworksoul/smudge, you can execute the following:
make buildWhen using the Makefile, the compiled binary will be present in the /bin directory in the code directory.
If you'd rather not use a Makefile:
go build -a -installsuffix cgo -o smudge github.com/clockworksoul/smudge/smudgeThe binary, compiled for your current environment, will be present in your present working directory.
To use the code, you simply specify a few configuration options (or use the defaults), create and add a node status change listener, and call the smudge.Begin() function.
Perhaps the simplest way of directing the behavior of the SWIM driver is by setting the appropriate system environment variables, which is useful when making use of Smudge inside of a container.
The following variables and their default values are as follows:
Variable | Default | Description
---------------------------------- | --------------- | -------------------------------
SMUDGE_CLUSTER_NAME | smudge | Cluster name for for multicast discovery
SMUDGE_HEARTBEAT_MILLIS | 250 | Milliseconds between heartbeats
SMUDGE_INITIAL_HOSTS | | Comma-delimmited list of known members as IP or IP:PORT
SMUDGE_LISTEN_PORT | 9999 | UDP port to listen on
SMUDGE_LISTEN_IP | 127.0.0.1 | IP address to listen on
SMUDGE_MAX_BROADCAST_BYTES | 256 | Maximum byte length of broadcast payloads
SMUDGE_MULTICAST_ENABLED | true | Multicast announce on startup; listen for multicast announcements
SMUDGE_MULTICAST_ANNOUNCE_INTERVAL | 0 | Seconds between multicast announcements, 0 will disable subsequent anouncements
SMUDGE_MULTICAST_ADDRESS | See description | The multicast broadcast address. Default: `224.0.0.0` (IPv4) or `[ff02::1]` (IPv6)
SMUDGE_MULTICAST_PORT | 9998 | The multicast listen port
If you prefer to direct the behavior of the service using the API, the calls are relatively straight-forward. Note that setting the application properties using this method overrides the behavior of environment variables.
smudge.SetListenPort(9999)
smudge.SetHeartbeatMillis(250)
smudge.SetListenIP(net.ParseIP("127.0.0.1"))
smudge.SetMaxBroadcastBytes(256) // set to 512 when using IPv6Creating a status change listener is very straight-forward:
type MyStatusListener struct {
smudge.StatusListener
}
func (m MyStatusListener) OnChange(node *smudge.Node, status smudge.NodeStatus) {
fmt.Printf("Node %s is now status %s\n", node.Address(), status)
}
func main() {
smudge.AddStatusListener(MyStatusListener{})
}Adding a broadcast listener is very similar to creating a status listener:
type MyBroadcastListener struct {
smudge.BroadcastListener
}
func (m MyBroadcastListener) OnBroadcast(b *smudge.Broadcast) {
fmt.Printf("Received broadcast from %v: %s\n",
b.Origin().Address(),
string(b.Bytes()))
}
func main() {
smudge.AddBroadcastListener(MyBroadcastListener{})
}Adding a new member to your known nodes list will also make that node aware of the adding server. To join an existing cluster without using multicast (or on a network where multicast is disabled) you must use this method to add at least one of that cluster's healthy member nodes.
node, err := smudge.CreateNodeByAddress("localhost:10000")
if err == nil {
smudge.AddNode(node)
}Once everything else is done, starting the server is trivial:
Simply call: smudge.Begin()
To transmit a broadcast to all healthy nodes currenty in the cluster you can use one of the BroadcastBytes(bytes []byte) or BroadcastString(str string) functions.
Be aware of the following caveats:
- Attempting to send a broadcast before the server has been started will cause a panic.
- The broadcast will not be received by the originating member;
BroadcastListeners on the originating member will not be triggered. - Nodes that join the cluster after the broadcast has been fully propagated will not receive the broadcast; nodes that join after the initial transmission but before complete proagation may or may not receive the broadcast.
The AllNodes() can be used to get all known nodes; HealthyNodes() works similarly, but returns only healthy nodes (defined as nodes with a status of "alive").
package main
import "github.com/clockworksoul/smudge"
import "fmt"
import "net"
type MyStatusListener struct {
smudge.StatusListener
}
func (m MyStatusListener) OnChange(node *smudge.Node, status smudge.NodeStatus) {
fmt.Printf("Node %s is now status %s\n", node.Address(), status)
}
type MyBroadcastListener struct {
smudge.BroadcastListener
}
func (m MyBroadcastListener) OnBroadcast(b *smudge.Broadcast) {
fmt.Printf("Received broadcast from %s: %s\n",
b.Origin().Address(),
string(b.Bytes()))
}
func main() {
heartbeatMillis := 500
listenPort := 9999
// Set configuration options
smudge.SetListenPort(listenPort)
smudge.SetHeartbeatMillis(heartbeatMillis)
smudge.SetListenIP(net.ParseIP("127.0.0.1"))
// Add the status listener
smudge.AddStatusListener(MyStatusListener{})
// Add the broadcast listener
smudge.AddBroadcastListener(MyBroadcastListener{})
// Add a new remote node. Currently, to join an existing cluster you must
// add at least one of its healthy member nodes.
node, err := smudge.CreateNodeByAddress("localhost:10000")
if err == nil {
smudge.AddNode(node)
}
// Start the server!
smudge.Begin()
}Smudge comes with a DefaultLogger that writes log messages to stderr. You can plug in your own logger by implementing the functions of the Logger interface and setting the logger by calling smudge.SetLogger(MyCoolLogger).