Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
@@ -0,0 +1,377 @@
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.kafka.streams.processor.internals.assignment;

import java.util.HashMap;
import java.util.HashSet;
import java.util.Map;
import java.util.Map.Entry;
import java.util.Objects;
import java.util.Set;
import java.util.SortedMap;
import java.util.SortedSet;
import java.util.TreeMap;
import java.util.TreeSet;
import java.util.stream.Collectors;

public class Graph<V extends Comparable<V>> {
public class Edge {
final V destination;
final int capacity;
final int cost;
int residualFlow;

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do we actually need this field?

If I read the code right, for a forward edge, it's always capacity - flow and for a backward edge it's always 0. So it seem redundant (and potentially error prone to store it expliclity)? -- Instead we could have a residualFlow() method that compute it on-the-fly (we could also simplify the update logic when modifying flow as we only need to update the flow itself)? Or do I read the update logic inside cancelNegativeCycle incorrectly and those properties are not an invariant?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah. Parameter order... Thought it's , flow, residualFlow, (not sure why you picked the "reverse" order)

int flow;
Edge counterEdge;
boolean forwardEdge;

public Edge(final V destination, final int capacity, final int cost, final int residualFlow, final int flow) {
this(destination, capacity, cost, residualFlow, flow, true);
}

public Edge(final V destination, final int capacity, final int cost, final int residualFlow, final int flow,
final boolean forwardEdge) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: formatting (if it does not fit in one line, we should move each parameter into it's one line to simplify reading)

public Edge(
    final V destination,
    final int capacity,
    final int cost,
    final int residualFlow,
    final int flow,
    final boolean forwardEdge
) {

Objects.requireNonNull(destination);
if (capacity < 0) {
throw new IllegalArgumentException("Edge capacity cannot be negative");
}
if (flow > capacity) {
throw new IllegalArgumentException(String.format("Edge flow %d cannot exceed capacity %d",
flow, capacity));
}

this.destination = destination;
this.capacity = capacity;
this.cost = cost;
this.residualFlow = residualFlow;
this.flow = flow;
this.forwardEdge = forwardEdge;
}

@Override
public boolean equals(final Object other) {
if (this == other) {
return true;
}
if (other == null || other.getClass() != getClass()) {
return false;
}

final Graph<?>.Edge otherEdge = (Graph<?>.Edge) other;

return destination.equals(otherEdge.destination) && capacity == otherEdge.capacity

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit formatting:

return destination.equals(otherEdge.destination)
    && capacity == otherEdge.capacity
    && cost == otherEdge.cost
    && residualFlow == otherEdge.residualFlow
    && flow == otherEdge.flow
    && forwardEdge == otherEdge.forwardEdge;

&& cost == otherEdge.cost && residualFlow == otherEdge.residualFlow && flow == otherEdge.flow
&& forwardEdge == otherEdge.forwardEdge;
}

@Override
public int hashCode() {
return Objects.hash(destination, capacity, cost, residualFlow, flow, forwardEdge);
}

@Override
public String toString() {
return "{destination= " + destination + ", capacity=" + capacity + ", cost=" + cost

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

return "Edge: {...}" ?

+ ", residualFlow=" + residualFlow + ", flow=" + flow + ", forwardEdge=" + forwardEdge;

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

missing }.

Should we also switch to one line per parameter we print to simplify reading?

}
}

private final SortedMap<V, SortedMap<V, Edge>> adjList = new TreeMap<>();

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why do we use nested SortedMap? Is seems SortedMap<V, List<Edge> would be sufficient?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's for getting the edge between two nodes efficiently. e.g. to get edge between node 0 and node 1, we can do adjList.get(0).get(1).

private final SortedSet<V> nodes = new TreeSet<>();
private final boolean isResidualGraph;
private V sourceNode, sinkNode;

public Graph() {
this(false);
}

private Graph(final boolean isResidualGraph) {
this.isResidualGraph = isResidualGraph;
}

public void addEdge(final V u, final V v, final int capacity, final int cost, final int flow) {
addEdge(u, new Edge(v, capacity, cost, capacity - flow, flow));
}

public Set<V> nodes() {
return nodes;
}

public Map<V, Edge> edges(final V node) {
return adjList.get(node);
}

public boolean isResidualGraph() {
return isResidualGraph;
}

public void setSourceNode(final V node) {
sourceNode = node;
}

public void setSinkNode(final V node) {
sinkNode = node;
}

public int totalCost() {
int totalCost = 0;
for (final Map.Entry<V, SortedMap<V, Edge>> nodeEdges : adjList.entrySet()) {
final SortedMap<V, Edge> edges = nodeEdges.getValue();
for (final Entry<V, Edge> nodeEdge : edges.entrySet()) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: we could just iterate over the valueSet ?

totalCost += nodeEdge.getValue().cost * nodeEdge.getValue().flow;
}
}
return totalCost;
}

private void addEdge(final V u, final Edge edge) {
if (!isResidualGraph) {
// Check if there's already an edge from u to v
final Map<V, Edge> edgeMap = adjList.get(edge.destination);
if (edgeMap != null && edgeMap.containsKey(u)) {
throw new IllegalArgumentException(
"There is already an edge from " + edge.destination
+ " to " + u + ". Can not add an edge from " + u + " to " + edge.destination
+ " since there will create a cycle between two nodes");
}
}

adjList.computeIfAbsent(u, set -> new TreeMap<>()).put(edge.destination, edge);
nodes.add(u);
nodes.add(edge.destination);
}

/**
* Get residual graph of this graph.
* Residual graph definition:
* If there is an edge in original graph from u to v with capacity c, cost w and flow f,
* then in the new graph there are two edges e1 and e2. e1 is from u to v with capacity c - f,
* cost w and flow f. e2 is from v to u with capacity f, cost -w and flow 0.
*
* @return Residual graph
*/
public Graph<V> residualGraph() {
if (isResidualGraph) {
return this;
}

final Graph<V> residualGraph = new Graph<>(true);
for (final Map.Entry<V, SortedMap<V, Edge>> nodeEdges : adjList.entrySet()) {
final V node = nodeEdges.getKey();
final SortedMap<V, Edge> edges = nodeEdges.getValue();
for (final Entry<V, Edge> nodeEdge : edges.entrySet()) {
final Edge edge = nodeEdge.getValue();
final Edge forwardEdge = new Edge(edge.destination, edge.capacity, edge.cost, edge.capacity - edge.flow, edge.flow);
final Edge backwardEdge = new Edge(node, edge.capacity, edge.cost * -1, edge.flow, 0, false);
forwardEdge.counterEdge = backwardEdge;
backwardEdge.counterEdge = forwardEdge;
residualGraph.addEdge(node, forwardEdge);
residualGraph.addEdge(edge.destination, backwardEdge);
}
}
return residualGraph;
}

/**
* Solve min cost flow with cycle canceling algorithm.
*/
public void solveMinCostFlow() {
validateMinCostGraph();
final Graph<V> residualGraph = residualGraph();
residualGraph.cancelNegativeCycles();

for (final Entry<V, SortedMap<V, Edge>> nodeEdges : adjList.entrySet()) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not sure what this part does.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is to update the original graph edge to the computed flow. Basically iterating all edges in original graph and change the flow to what's in residualGraph's corresponding edge.

final V node = nodeEdges.getKey();
for (final Entry<V, Edge> nodeEdge : nodeEdges.getValue().entrySet()) {
final V destination = nodeEdge.getKey();
final Edge edge = nodeEdge.getValue();
final Edge residualEdge = residualGraph.adjList.get(node).get(destination);
edge.flow = residualEdge.flow;
edge.residualFlow = residualEdge.residualFlow;
}
}
}

private void populateInOutFlow(final Map<V, Long> inFlow, final Map<V, Long> outFlow) {
for (final Entry<V, SortedMap<V, Edge>> nodeEdges : adjList.entrySet()) {
final V node = nodeEdges.getKey();
if (node.equals(sinkNode)) {
throw new IllegalStateException("Sink node " + sinkNode + " shouldn't have output");
}
for (final Entry<V, Edge> nodeEdge : nodeEdges.getValue().entrySet()) {
final V destination = nodeEdge.getKey();
if (destination.equals(sourceNode)) {
throw new IllegalStateException("Source node " + sourceNode + " shouldn't have input " + node);
}
final Edge edge = nodeEdge.getValue();
Long count = outFlow.get(node);
if (count == null) {
outFlow.put(node, (long) edge.flow);
} else {
outFlow.put(node, count + edge.flow);
}

count = inFlow.get(destination);
if (count == null) {
inFlow.put(destination, (long) edge.flow);
} else {
inFlow.put(destination, count + edge.flow);
}
}
}
}

private void validateMinCostGraph() {
if (isResidualGraph) {
throw new IllegalStateException("Should not be residual graph to solve min cost flow");
}

/*
Check provided flow satisfying below constraints:
1. Input flow and output flow for each node should be the same except for source and destination node
2. Output flow of source and input flow of destination should be the same
*/

final Map<V, Long> inFlow = new HashMap<>();
final Map<V, Long> outFlow = new HashMap<>();
populateInOutFlow(inFlow, outFlow);

for (final Entry<V, Long> in : inFlow.entrySet()) {
if (in.getKey().equals(sourceNode) || in.getKey().equals(sinkNode)) {
continue;
}
final Long out = outFlow.get(in.getKey());
if (!Objects.equals(in.getValue(), out)) {
throw new IllegalStateException("Input flow for node " + in.getKey() + " is " +
in.getValue() + " which doesn't match output flow " + out);
}
}

final Long sourceOutput = outFlow.get(sourceNode);
final Long sinkInput = inFlow.get(sinkNode);
if (!Objects.equals(sourceOutput, sinkInput)) {
throw new IllegalStateException("Output flow for source " + sourceNode + " is " + sourceOutput
+ " which doesn't match input flow " + sinkInput + " for sink " + sinkNode);
}
}

private void cancelNegativeCycles() {
if (!isResidualGraph) {
throw new IllegalStateException("Should be residual graph to cancel negative cycles");
}
boolean cyclePossible = true;
while (cyclePossible) {
cyclePossible = false;
for (final V node : nodes) {
final Map<V, V> parentNodes = new HashMap<>();
final Map<V, Edge> parentEdges = new HashMap<>();
final V possibleNodeInCycle = detectNegativeCycles(node, parentNodes, parentEdges);

if (possibleNodeInCycle == null) {
continue;
}

final Set<V> visited = new HashSet<>();
V nodeInCycle = possibleNodeInCycle;
while (!visited.contains(nodeInCycle)) {
visited.add(nodeInCycle);
nodeInCycle = parentNodes.get(nodeInCycle);
}

cyclePossible = true;
cancelNegativeCycle(nodeInCycle, parentNodes, parentEdges);
}
}
}

private void cancelNegativeCycle(final V nodeInCycle, final Map<V, V> parentNodes, final Map<V, Edge> parentEdges) {
// Start from parentNode since nodeInCyle is used as exit condition in below loops
final V parentNode = parentNodes.get(nodeInCycle);
Edge parentEdge = parentEdges.get(nodeInCycle);

// Find max possible negative flow
int possibleFlow = parentEdge.residualFlow;
for (V curNode = parentNode; curNode != nodeInCycle; curNode = parentNodes.get(curNode)) {
parentEdge = parentEdges.get(curNode);
possibleFlow = Math.min(possibleFlow, parentEdge.residualFlow);
}

// Update graph by removing negative flow
parentEdge = parentEdges.get(nodeInCycle);
Edge counterEdge = parentEdge.counterEdge;
parentEdge.residualFlow -= possibleFlow;
if (parentEdge.forwardEdge) {
parentEdge.flow += possibleFlow;
}
counterEdge.residualFlow += possibleFlow;
if (counterEdge.forwardEdge && counterEdge.flow >= possibleFlow) {
counterEdge.flow -= possibleFlow;
}
for (V curNode = parentNode; curNode != nodeInCycle; curNode = parentNodes.get(curNode)) {
parentEdge = parentEdges.get(curNode);
counterEdge = parentEdge.counterEdge;
parentEdge.residualFlow -= possibleFlow;
if (parentEdge.forwardEdge) {
parentEdge.flow += possibleFlow;
}
counterEdge.residualFlow += possibleFlow;
if (counterEdge.forwardEdge && counterEdge.flow >= possibleFlow) {
counterEdge.flow -= possibleFlow;
}
}
}

/**
* Detect negative cycle using Bellman-ford shortest path algorithm.
* @param source Source node
* @param parentNodes Parent nodes to store negative cycle nodes
* @param parentEdges Parent edges to store negative cycle edges
*
* @return One node which can lead to negative cycle if exists or null if there's no negative cycle
*/
V detectNegativeCycles(final V source, final Map<V, V> parentNodes, final Map<V, Edge> parentEdges) {
// Use long to account for any overflow
final Map<V, Long> distance = nodes.stream().collect(Collectors.toMap(node -> node, node -> (long) Integer.MAX_VALUE));
distance.put(source, 0L);
final int nodeCount = nodes.size();

// Iterate nodeCount iterations since Bellaman-Ford will find shortest path in nodeCount - 1
// iterations. If the distance can still be relaxed in nodeCount iteration, there's a negative
// cycle
for (int i = 0; i < nodeCount; i++) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

From https://en.wikipedia.org/wiki/Bellman%E2%80%93Ford_algorithm it says, so it N-1 times, but we do it N times. Why?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good question. N-1 times will find shortest path, we need to do it one more time to see if the path can be even shorter which mean there's a negative cycle. That's why there's a check on line 335.

// Iterate through all edges
for (final Entry<V, SortedMap<V, Edge>> nodeEdges : adjList.entrySet()) {
final V u = nodeEdges.getKey();
for (final Entry<V, Edge> nodeEdge : nodeEdges.getValue().entrySet()) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking into https://en.wikipedia.org/wiki/Bellman%E2%80%93Ford_algorithm that you linked from the KIP, the algorithms says "go over all edges" -- we need 2 for-loops to this but what is not totally obvious; might be worth to add a coment?

final Edge edge = nodeEdge.getValue();
if (edge.residualFlow == 0) {

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you explain this condition?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This means their can't be a flow in this edge which essentially means these two nodes are disconnected.

continue;
}
final V v = edge.destination;
if (distance.get(v) > distance.get(u) + edge.cost) {
if (i == nodeCount - 1) {
return v;
}
distance.put(v, distance.get(u) + edge.cost);
parentNodes.put(v, u);
parentEdges.put(v, edge);
}
}
}
}
return null;
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.kafka.streams.processor.internals;
package org.apache.kafka.streams.processor.internals.assignment;

import java.util.HashMap;
import java.util.HashSet;
Expand All @@ -30,6 +30,7 @@
import org.apache.kafka.common.TopicPartitionInfo;
import org.apache.kafka.streams.KeyValue;
import org.apache.kafka.streams.processor.TaskId;
import org.apache.kafka.streams.processor.internals.InternalTopicManager;
import org.apache.kafka.streams.processor.internals.TopologyMetadata.Subtopology;
import org.apache.kafka.streams.processor.internals.assignment.AssignorConfiguration.AssignmentConfigs;
import org.slf4j.Logger;
Expand Down
Loading