swisspost/vertx-cluster-watchdog

Checks if all your cluster members are receiveing published messages over the bus.

vertx-cluster-watchdog

Checks if all your hazelcast cluster members are receiveing published messages over the bus.

How to run the watchdog

You can run the watchdog standalone to try different cluster settings or deploy it into an existing vertx instance.

1. Standalone: build the fatjar with mvn install -PshadowJar run the jar with java -jar target/cluster-watchdog-x.x.x-SNAPSHOT-all.jar You can pass configuration to the fatjar with -conf conf.json, check the configuration section for the possible params
2. Existing instance: Deploy the module as every other module

How the watchdog works

1. every verticle gets a unique id
2. The amount of cluster members is read over the hazelcast api Cluster#getMembers()
3. a message is published to the broadcast address, the address to reply to is in the payload with a timestamp
4. the receivers of the broadcast message are sending a message back to the sender of the broadcast address, with the timestamp and its unique id
5. the receivers of the point to point message are counting the received messages with the same timestamp
6. the result of each member is sent to the other members to have consensus
7. every member has a CircularFifoQueue where he stores the result of himself and the received results, the length can be configured
8. if the watchdog is asked for consistency over http://host:port/clusterStatus, the CircularFifoQueue will be consulted, if there is one inconsistent entry, the cluster will be considered as INCONSISTENT if every entry is consistent, the cluster will be considered as CONSISTENT

Rest API

• The status of the cluster can be get over the following URL http://host:port/clusterStatus
• A detailed view of the last watchdog runs can be get over the following URL http://host:port/clusterWatchdogStats

Restrictions

• There must be one vertx-cluster-watchdog verticle instance per vertx instance, the watchdog is relying on the fact that one broadcast message is received by one cluster member
• the verticle can only handle one hazelcast instance. If there are more than one hazelcast instances, the watchdog will not be run

Configuration

{
    "http.port": 7878                       // Port we serve http. Defaults to 7878.
    "cluster.noconf":true                   // pass an empty HazelcastClusterManager instance to the ClusterManager
    "cluster.host":"192.168.26.3"           // the cluster host passed to the VertxOptions
    "hazelcast.group.name":"grp"            // the groupname passed to the Hazelcast GroupConfig
    "hazelcast.group.password":"pw"         // the password passed to the Hazelcast GroupConfig
    "hazelcast.net.port":8981               // the port passed to Hazelcast NetworkConfig
    "hazelcast.multicast.group":"224.2.2.3" // the multicast group passed to the Hazelcast MulticastConfig
    "hazelcast.multicast.port":5559         // the multicast group passed to the Hazelcast MulticastConfig
    "intervalInSec": 30,                    // In which interval the watchdog will be run. Defaults to 30, if the interval is set to 0 the watchdog only run once after deployment.
    "clusterMembers": -1,                   // The amount of the cluster members, defaults to -1, which lets the mod figure out itself the amount of cluster members.
    "resultQueueLength: 100"                // The amount of watchdog runs, that should be kept and considered to figure out the cluster state, defaults to 100.
}

Tests

The tests try to simulate the cluster with multiple instances of the verticle. The amount of cluster members is injected over the config.