org.corfudb.infrastructure.management.failuredetector

Class ClusterGraph

java.lang.Object

org.corfudb.infrastructure.management.failuredetector.ClusterGraph

public class ClusterGraph
extends Object

Graph representation of a cluster state. For instance, asymmetric graph, there is an asymmetric failure between nodes `a` and `b`. NodeConnectivity.ConnectionStatus could be OK/FAILED: {a -> {{a: OK}, {b: FAILED}, {c: OK}}} {b -> {{a: FAILED}, {b: OK}, {c: FAILED}}} {c -> {{a: OK}, {b: false}, {c: OK}}}

This class represents a cluster as a graph, it allows to: - find fully connected node: node that connected to all other nodes exclude unresponsive nodes. - find a decision maker: a node which should decide to add failed node to the unresponsive list. The decision maker is an optimization technique, needed to reduce parallel updates of the cluster state. Still, each node can choose its own decision maker. See this#getDecisionMaker() for details - find a node which should be healed and excluded from the unresponsive list - transform all asymmetric failures to a symmetric ones. If there is an asymmetric failure (node B can connect to node B, but node B can not connect to node A) between two nodes then make it symmetric which means neither node A nor B can connect to each other.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ClusterGraph.ClusterGraphHelper Helper provides methods to build cluster graph

Constructor Summary

Constructors

Constructor and Description
ClusterGraph()

Method Summary

Modifier and Type	Method and Description
FailureDetectorMetrics.ConnectivityGraph	connectivityGraph() Transform ClusterGraph to Connectivity graph
Optional<NodeRank>	findFailedNode() Find failed node in a graph.
Optional<NodeRank>	findFullyConnectedNode(String endpoint, List<String> unresponsiveNodes) See if the node is fully connected.
Optional<NodeRank>	getDecisionMaker() Get a decision maker node to detect a failure.
NodeConnectivity	getNodeConnectivity(String node) Get node by name
int	size() Graph size
static ClusterGraph	toClusterGraph(ClusterState cluster, String localNode) Transform a cluster state to the cluster graph.
String	toJson()
ClusterGraph	toSymmetric() Convert a cluster graph which could have asymmetric failures to a graph with symmetric failures between nodes.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

ClusterGraph

public ClusterGraph()

Method Detail

toClusterGraph

public static ClusterGraph toClusterGraph(ClusterState cluster,
                                          String localNode)

Transform a cluster state to the cluster graph. ClusterState contains some extra information, cluster graph is a pure representation of a graph of nodes. To be able to work efficiently with cluster state to find failed and healed nodes the cluster state should be transformed to cluster graph

Parameters:

cluster - cluster state

Returns:

cluster graph

toSymmetric

public ClusterGraph toSymmetric()

Convert a cluster graph which could have asymmetric failures to a graph with symmetric failures between nodes. See NodeConnectivity.ConnectionStatus for OK/FAILED For instance: {a: [{a: OK, b: FAILED}]} {b: [{a: OK, b: OK}]} Node A believes that node B is disconnected Node B believes that node A is connected. The graph will be changed to: {a: [{a: OK, b: FAILED}]} {b: [{a: FAILED, b: OK}]} Node B is not connected to node A anymore.

Returns:

a graph with symmetric failures between nodes.

getDecisionMaker

public Optional<NodeRank> getDecisionMaker()

Get a decision maker node to detect a failure. It must have: - highest number of successful connections in the graph. There could be many nodes with same number of successful connections, then the decision maker will be a node with smallest name.

The decision maker is an optimization technique, needed to reduce parallel updates of the cluster state. Still, each node can choose its own decision maker. Note: it's not required to choose a particular one for entire cluster to add a node to unresponsive list. In a partitioned scenario, a minority side can choose to have a decision maker just that it's decisions will not be used as it does not have consensus The decision maker always exists. There is always at least the local node, it always has at least one successful connection to itself. We also have two additional checks to prevent all possible incorrect ways. Decision maker not found if: - ClusterGraph is empty, which is an invalid state. - the decision maker doesn't have connections, which is also impossible.

Returns:

a decision maker node

findFailedNode

public Optional<NodeRank> findFailedNode()

Find failed node in a graph. Collect all node ranks in a graph and choose the node with smallest number of successful connections and with highest name (sorted alphabetically)

Returns:

failed node

findFullyConnectedNode

public Optional<NodeRank> findFullyConnectedNode(String endpoint,
                                                 List<String> unresponsiveNodes)

See if the node is fully connected.

Parameters:

endpoint - local node name

unresponsiveNodes - list of unresponsive nodes in the layout

Returns:

local node rank

getNodeConnectivity

public NodeConnectivity getNodeConnectivity(String node)

Get node by name

Parameters:

node - node name

Returns:

a node in the graph

size

public int size()

Graph size

Returns:

size

toJson

public String toJson()

connectivityGraph

public FailureDetectorMetrics.ConnectivityGraph connectivityGraph()

Transform ClusterGraph to Connectivity graph

Returns:

connectivity graph