public String toGraphviz() {
StringBuilder builder = new StringBuilder();
builder.append("digraph QuantileDigest {\n").append("\tgraph [ordering=\"out\"];");
final List<Node> nodes = new ArrayList<>();
postOrderTraversal(
root,
new Callback() {
@Override
public boolean process(Node node) {
nodes.add(node);
return true;
}
});
Multimap<Integer, Node> nodesByLevel =
Multimaps.index(
nodes,
new Function<Node, Integer>() {
@Override
public Integer apply(Node input) {
return input.level;
}
});
for (Map.Entry<Integer, Collection<Node>> entry : nodesByLevel.asMap().entrySet()) {
builder.append("\tsubgraph level_" + entry.getKey() + " {\n").append("\t\trank = same;\n");
for (Node node : entry.getValue()) {
builder.append(
String.format(
"\t\t%s [label=\"[%s..%s]@%s\\n%s\", shape=rect, style=filled,color=%s];\n",
idFor(node),
node.getLowerBound(),
node.getUpperBound(),
node.level,
node.weightedCount,
node.weightedCount > 0 ? "salmon2" : "white"));
}
builder.append("\t}\n");
}
for (Node node : nodes) {
if (node.left != null) {
builder.append(format("\t%s -> %s;\n", idFor(node), idFor(node.left)));
}
if (node.right != null) {
builder.append(format("\t%s -> %s;\n", idFor(node), idFor(node.right)));
}
}
builder.append("}\n");
return builder.toString();
}
private void setChild(Node parent, long branch, Node child) {
if (parent == null) {
root = child;
} else if (branch == 0) {
parent.left = child;
} else {
parent.right = child;
}
}
private Node copyRecursive(Node node) {
Node result = null;
if (node != null) {
result = createNode(node.bits, node.level, node.weightedCount);
result.left = copyRecursive(node.left);
result.right = copyRecursive(node.right);
}
return result;
}
public static QuantileDigest deserialize(DataInput input) {
try {
double maxError = input.readDouble();
double alpha = input.readDouble();
QuantileDigest result = new QuantileDigest(maxError, alpha);
result.landmarkInSeconds = input.readLong();
result.min = input.readLong();
result.max = input.readLong();
result.totalNodeCount = input.readInt();
Deque<Node> stack = new ArrayDeque<>();
for (int i = 0; i < result.totalNodeCount; i++) {
int flags = input.readByte();
Node node = deserializeNode(input);
if ((flags & Flags.HAS_RIGHT) != 0) {
node.right = stack.pop();
}
if ((flags & Flags.HAS_LEFT) != 0) {
node.left = stack.pop();
}
stack.push(node);
result.weightedCount += node.weightedCount;
if (node.weightedCount >= ZERO_WEIGHT_THRESHOLD) {
result.nonZeroNodeCount++;
}
}
if (!stack.isEmpty()) {
Preconditions.checkArgument(
stack.size() == 1, "Tree is corrupted. Expected a single root node");
result.root = stack.pop();
}
return result;
} catch (IOException e) {
throw Throwables.propagate(e);
}
}
private void insert(long bits, double weight) {
long lastBranch = 0;
Node parent = null;
Node current = root;
while (true) {
if (current == null) {
setChild(parent, lastBranch, createLeaf(bits, weight));
return;
} else if (!inSameSubtree(bits, current.bits, current.level)) {
// if bits and node.bits are not in the same branch given node's level,
// insert a parent above them at the point at which branches diverge
setChild(parent, lastBranch, makeSiblings(current, createLeaf(bits, weight)));
return;
} else if (current.level == 0 && current.bits == bits) {
// found the node
double oldWeight = current.weightedCount;
current.weightedCount += weight;
if (current.weightedCount >= ZERO_WEIGHT_THRESHOLD && oldWeight < ZERO_WEIGHT_THRESHOLD) {
++nonZeroNodeCount;
}
weightedCount += weight;
return;
}
// we're on the correct branch of the tree and we haven't reached a leaf, so keep going down
long branch = bits & current.getBranchMask();
parent = current;
lastBranch = branch;
if (branch == 0) {
current = current.left;
} else {
current = current.right;
}
}
}
/**
* Remove the node if possible or set its count to 0 if it has children and it needs to be kept
* around
*/
private Node tryRemove(Node node) {
if (node == null) {
return null;
}
if (node.weightedCount >= ZERO_WEIGHT_THRESHOLD) {
--nonZeroNodeCount;
}
weightedCount -= node.weightedCount;
Node result = null;
if (node.isLeaf()) {
--totalNodeCount;
} else if (node.hasSingleChild()) {
result = node.getSingleChild();
--totalNodeCount;
} else {
node.weightedCount = 0;
result = node;
}
return result;
}
private Node makeSiblings(Node node, Node sibling) {
int parentLevel = MAX_BITS - Long.numberOfLeadingZeros(node.bits ^ sibling.bits);
Node parent = createNode(node.bits, parentLevel, 0);
// the branch is given by the bit at the level one below parent
long branch = sibling.bits & parent.getBranchMask();
if (branch == 0) {
parent.left = sibling;
parent.right = node;
} else {
parent.left = node;
parent.right = sibling;
}
return parent;
}
private Node merge(Node node, Node other) {
if (node == null) {
return copyRecursive(other);
} else if (other == null) {
return node;
} else if (!inSameSubtree(node.bits, other.bits, Math.max(node.level, other.level))) {
return makeSiblings(node, copyRecursive(other));
} else if (node.level > other.level) {
long branch = other.bits & node.getBranchMask();
if (branch == 0) {
node.left = merge(node.left, other);
} else {
node.right = merge(node.right, other);
}
return node;
} else if (node.level < other.level) {
Node result = createNode(other.bits, other.level, other.weightedCount);
long branch = node.bits & other.getBranchMask();
if (branch == 0) {
result.left = merge(node, other.left);
result.right = copyRecursive(other.right);
} else {
result.left = copyRecursive(other.left);
result.right = merge(node, other.right);
}
return result;
}
// else, they must be at the same level and on the same path, so just bump the counts
double oldWeight = node.weightedCount;
weightedCount += other.weightedCount;
node.weightedCount = node.weightedCount + other.weightedCount;
node.left = merge(node.left, other.left);
node.right = merge(node.right, other.right);
if (oldWeight < ZERO_WEIGHT_THRESHOLD && node.weightedCount >= ZERO_WEIGHT_THRESHOLD) {
nonZeroNodeCount++;
}
return node;
}