Add deterministic dependency scheduler

This commit is contained in:
2026-10-19 15:11:28 +00:00
parent aa93aeda3f
commit 167fcd475f
2 changed files with 415 additions and 0 deletions

View File

@@ -0,0 +1,200 @@
package engine
import (
"container/heap"
"sort"
)
// dependencyKey is an immutable tuple used to order dependency-graph nodes.
// Callers provide unique keys and valid edge endpoints.
type dependencyKey []string
// dependencyEdge points from a prerequisite to the node that depends on it.
type dependencyEdge struct {
prerequisite int
dependent int
}
type dependencyReadyHeap struct {
nodes []int
keys []dependencyKey
}
func (h dependencyReadyHeap) Len() int { return len(h.nodes) }
func (h dependencyReadyHeap) Less(i, j int) bool {
return dependencyKeyLess(h.keys[h.nodes[i]], h.keys[h.nodes[j]])
}
func (h dependencyReadyHeap) Swap(i, j int) {
h.nodes[i], h.nodes[j] = h.nodes[j], h.nodes[i]
}
func (h *dependencyReadyHeap) Push(value any) {
h.nodes = append(h.nodes, value.(int))
}
func (h *dependencyReadyHeap) Pop() any {
last := len(h.nodes) - 1
node := h.nodes[last]
h.nodes = h.nodes[:last]
return node
}
// dependencyKeyLess compares tuple components bytewise, with a shorter tuple
// sorting first when all of its components match the corresponding prefix.
func dependencyKeyLess(left, right dependencyKey) bool {
for i := 0; i < len(left) && i < len(right); i++ {
if left[i] != right[i] {
return left[i] < right[i]
}
}
return len(left) < len(right)
}
// scheduleDependencies applies Kahn's algorithm, always selecting the
// smallest currently-ready key. Duplicate edges count only once. If scheduling
// stalls, an explicit depth-first traversal of the residual graph returns an
// actual directed cycle, rotated to its smallest key and closed by repeating
// its first node; blocked non-cycle dependents are excluded.
//
// Nodes are visited by precomputed key rank, which also makes each adjacency
// list deterministic without sorting edges. The work is O(E + V log V), with
// O(V + E) additional memory, apart from key copying and tuple comparisons.
func scheduleDependencies(keys []dependencyKey, edges []dependencyEdge) (order, cycle []int) {
ownedKeys := make([]dependencyKey, len(keys))
for i, key := range keys {
ownedKeys[i] = append(dependencyKey(nil), key...)
}
nodeOrder := make([]int, len(ownedKeys))
for i := range nodeOrder {
nodeOrder[i] = i
}
sort.Slice(nodeOrder, func(i, j int) bool {
return dependencyKeyLess(ownedKeys[nodeOrder[i]], ownedKeys[nodeOrder[j]])
})
incoming := make([][]int, len(ownedKeys))
for _, edge := range edges {
incoming[edge.dependent] = append(incoming[edge.dependent], edge.prerequisite)
}
outgoing := make([][]int, len(ownedKeys))
indegree := make([]int, len(ownedKeys))
for _, dependent := range nodeOrder {
seen := make(map[int]struct{}, len(incoming[dependent]))
for _, prerequisite := range incoming[dependent] {
if _, duplicate := seen[prerequisite]; duplicate {
continue
}
seen[prerequisite] = struct{}{}
indegree[dependent]++
// Dependents are visited by key rank, so each outgoing list is
// already ordered by key.
outgoing[prerequisite] = append(outgoing[prerequisite], dependent)
}
}
ready := dependencyReadyHeap{keys: ownedKeys}
for _, node := range nodeOrder {
if indegree[node] == 0 {
ready.nodes = append(ready.nodes, node)
}
}
heap.Init(&ready)
order = make([]int, 0, len(ownedKeys))
for ready.Len() != 0 {
node := heap.Pop(&ready).(int)
order = append(order, node)
for _, dependent := range outgoing[node] {
indegree[dependent]--
if indegree[dependent] == 0 {
heap.Push(&ready, dependent)
}
}
}
if len(order) == len(ownedKeys) {
return order, nil
}
return nil, dependencyCycle(ownedKeys, nodeOrder, outgoing, indegree)
}
type dependencyDFSFrame struct {
node int
next int
}
// dependencyCycle searches the residual graph iteratively to avoid recursion
// depth depending on input size.
func dependencyCycle(keys []dependencyKey, nodeOrder []int, outgoing [][]int, indegree []int) []int {
const (
unvisited = uint8(0)
active = uint8(1)
finished = uint8(2)
)
state := make([]uint8, len(keys))
position := make([]int, len(keys))
for i := range position {
position[i] = -1
}
path := make([]int, 0, len(keys))
frames := make([]dependencyDFSFrame, 0, len(keys))
for _, root := range nodeOrder {
if indegree[root] == 0 || state[root] != unvisited {
continue
}
state[root] = active
position[root] = len(path)
path = append(path, root)
frames = append(frames, dependencyDFSFrame{node: root})
for len(frames) != 0 {
frame := &frames[len(frames)-1]
if frame.next == len(outgoing[frame.node]) {
state[frame.node] = finished
position[frame.node] = -1
path = path[:len(path)-1]
frames = frames[:len(frames)-1]
continue
}
next := outgoing[frame.node][frame.next]
frame.next++
if indegree[next] == 0 {
continue
}
switch state[next] {
case unvisited:
state[next] = active
position[next] = len(path)
path = append(path, next)
frames = append(frames, dependencyDFSFrame{node: next})
case active:
return canonicalDependencyCycle(keys, path[position[next]:])
}
}
}
// A stalled Kahn traversal always leaves a cycle in the residual graph.
return nil
}
func canonicalDependencyCycle(keys []dependencyKey, path []int) []int {
start := 0
for i := 1; i < len(path); i++ {
if dependencyKeyLess(keys[path[i]], keys[path[start]]) {
start = i
}
}
cycle := make([]int, len(path)+1)
for i := range path {
cycle[i] = path[(start+i)%len(path)]
}
cycle[len(path)] = cycle[0]
return cycle
}

View File

@@ -0,0 +1,215 @@
package engine
import (
"reflect"
"testing"
)
func TestScheduleDependenciesOrdersReadyNodes(t *testing.T) {
tests := []struct {
name string
keys []string
edges [][2]string
want []string
}{
{
name: "independent branches",
keys: []string{"branch-c", "branch-a", "branch-b"},
want: []string{"branch-a", "branch-b", "branch-c"},
},
{
name: "unequal chains",
keys: []string{"chain-d", "chain-c", "chain-b", "chain-a"},
edges: [][2]string{
{"chain-a", "chain-b"},
{"chain-b", "chain-c"},
{"chain-c", "chain-d"},
},
want: []string{"chain-a", "chain-b", "chain-c", "chain-d"},
},
{
name: "multiple dependencies",
keys: []string{"application", "right", "base", "left"},
edges: [][2]string{
{"base", "left"}, {"base", "right"},
{"left", "application"}, {"right", "application"},
},
want: []string{"base", "left", "right", "application"},
},
{
name: "shared dependency at equal depths",
keys: []string{"sink", "right", "root", "left"},
edges: [][2]string{
{"root", "left"}, {"root", "right"},
{"left", "sink"}, {"right", "sink"},
},
want: []string{"root", "left", "right", "sink"},
},
{
name: "shared dependency at unequal depths",
keys: []string{"sink", "deep", "side", "middle", "root"},
edges: [][2]string{
{"root", "middle"}, {"middle", "deep"}, {"root", "side"},
{"deep", "sink"}, {"side", "sink"},
},
want: []string{"root", "middle", "deep", "side", "sink"},
},
{
name: "newly ready small key precedes older large key",
keys: []string{"z-ready", "a-new", "b-ready"},
edges: [][2]string{{"b-ready", "a-new"}},
want: []string{"b-ready", "a-new", "z-ready"},
},
{
name: "duplicate edges count once",
keys: []string{"dependent", "prerequisite"},
edges: [][2]string{
{"prerequisite", "dependent"},
{"prerequisite", "dependent"},
},
want: []string{"prerequisite", "dependent"},
},
}
for _, test := range tests {
t.Run(test.name, func(t *testing.T) {
keys, edges, indexes := dependencyGraph(test.keys, test.edges)
order, cycle := scheduleDependencies(keys, edges)
if len(cycle) != 0 {
t.Fatalf("unexpected cycle: %v", dependencyNames(cycle, indexes))
}
if got := dependencyNames(order, indexes); !reflect.DeepEqual(got, test.want) {
t.Fatalf("schedule = %v, want %v", got, test.want)
}
})
}
}
func TestScheduleDependenciesUsesTupleKeys(t *testing.T) {
keys := []dependencyKey{
{"a", "z"},
{"a"},
{"a", "aa"},
{"A"},
}
order, cycle := scheduleDependencies(keys, nil)
if len(cycle) != 0 {
t.Fatalf("unexpected cycle: %v", cycle)
}
want := []int{3, 1, 2, 0}
if !reflect.DeepEqual(order, want) {
t.Fatalf("schedule = %v, want %v", order, want)
}
}
func TestScheduleDependenciesReportsActualCanonicalCycles(t *testing.T) {
tests := []struct {
name string
keys []string
edges [][2]string
want []string
}{
{
name: "self cycle",
keys: []string{"self"},
edges: [][2]string{{"self", "self"}},
want: []string{"self", "self"},
},
{
name: "multi-node cycle excludes blocked dependent",
keys: []string{"blocked", "cycle-b", "cycle-a"},
edges: [][2]string{
{"cycle-a", "cycle-b"},
{"cycle-b", "cycle-a"},
{"cycle-b", "blocked"},
},
want: []string{"cycle-a", "cycle-b", "cycle-a"},
},
}
for _, test := range tests {
t.Run(test.name, func(t *testing.T) {
keys, edges, indexes := dependencyGraph(test.keys, test.edges)
order, cycle := scheduleDependencies(keys, edges)
if len(order) != 0 {
t.Fatalf("unexpected schedule: %v", dependencyNames(order, indexes))
}
if got := dependencyNames(cycle, indexes); !reflect.DeepEqual(got, test.want) {
t.Fatalf("cycle = %v, want %v", got, test.want)
}
})
}
}
func TestScheduleDependenciesIsStableAcrossInsertionPermutations(t *testing.T) {
edges := [][2]string{
{"base", "left"}, {"base", "right"},
{"left", "sink"}, {"right", "sink"},
}
permutations := [][]string{
{"sink", "right", "base", "left"},
{"left", "base", "sink", "right"},
{"right", "left", "base", "sink"},
}
for i, permutation := range permutations {
keys, permutedEdges, indexes := dependencyGraph(permutation, edges)
if i%2 != 0 {
reverseDependencyEdges(permutedEdges)
}
order, cycle := scheduleDependencies(keys, permutedEdges)
if len(cycle) != 0 {
t.Fatalf("permutation %d returned cycle %v", i, dependencyNames(cycle, indexes))
}
if got, want := dependencyNames(order, indexes), []string{"base", "left", "right", "sink"}; !reflect.DeepEqual(got, want) {
t.Fatalf("permutation %d schedule = %v, want %v", i, got, want)
}
}
cycleEdges := [][2]string{{"b", "a"}, {"a", "b"}, {"b", "blocked"}}
cyclePermutations := [][]string{{"blocked", "b", "a"}, {"a", "blocked", "b"}, {"b", "a", "blocked"}}
for i, permutation := range cyclePermutations {
keys, permutedEdges, indexes := dependencyGraph(permutation, cycleEdges)
if i%2 != 0 {
reverseDependencyEdges(permutedEdges)
}
order, cycle := scheduleDependencies(keys, permutedEdges)
if len(order) != 0 {
t.Fatalf("cycle permutation %d returned schedule %v", i, dependencyNames(order, indexes))
}
if got, want := dependencyNames(cycle, indexes), []string{"a", "b", "a"}; !reflect.DeepEqual(got, want) {
t.Fatalf("cycle permutation %d returned %v, want %v", i, got, want)
}
}
}
func dependencyGraph(nodeNames []string, namedEdges [][2]string) ([]dependencyKey, []dependencyEdge, map[int]string) {
keys := make([]dependencyKey, len(nodeNames))
indexes := make(map[string]int, len(nodeNames))
for i, name := range nodeNames {
keys[i] = dependencyKey{name}
indexes[name] = i
}
edges := make([]dependencyEdge, len(namedEdges))
for i, edge := range namedEdges {
edges[i] = dependencyEdge{prerequisite: indexes[edge[0]], dependent: indexes[edge[1]]}
}
byIndex := make(map[int]string, len(nodeNames))
for name, index := range indexes {
byIndex[index] = name
}
return keys, edges, byIndex
}
func dependencyNames(nodes []int, names map[int]string) []string {
result := make([]string, len(nodes))
for i, node := range nodes {
result[i] = names[node]
}
return result
}
func reverseDependencyEdges(edges []dependencyEdge) {
for i, j := 0, len(edges)-1; i < j; i, j = i+1, j-1 {
edges[i], edges[j] = edges[j], edges[i]
}
}