@diagrammo/dgmo 0.4.1 → 0.4.3

package/src/infra/compute.ts

@@ -0,0 +1,1113 @@
+// ============================================================
+// Infra Chart Computation Engine
+// ============================================================
+//
+// Pure function: computeInfra(parsed, params) → ComputedInfraModel
+//
+// Traverses the DAG from the edge entry point, computing downstream
+// rps through behavior transformations (cache-hit, firewall-block,
+// ratelimit-rps) and split distribution.
+
+import type {
+  ParsedInfra,
+  InfraNode,
+  InfraEdge,
+  InfraComputeParams,
+  ComputedInfraModel,
+  ComputedInfraNode,
+  ComputedInfraEdge,
+  InfraDiagnostic,
+  InfraCbState,
+  InfraLatencyPercentiles,
+  InfraAvailabilityPercentiles,
+  InfraProperty,
+} from './types';
+import { INFRA_BEHAVIOR_KEYS } from './types';
+
+// ============================================================
+// Helpers
+// ============================================================
+
+/** Get a numeric property value from a node, or a default. */
+function getNumProp(node: InfraNode, key: string, fallback: number): number {
+  const prop = node.properties.find((p) => p.key === key);
+  if (!prop) return fallback;
+  if (typeof prop.value === 'number') return prop.value;
+  // Try parsing string (e.g. "1-8" is not a simple number)
+  const num = parseFloat(String(prop.value));
+  return isNaN(num) ? fallback : num;
+}
+
+/** Parse instance range "N-M" → {min, max}, or fixed "N" → {min: N, max: N}. */
+function getInstanceRange(node: InfraNode): { min: number; max: number } {
+  const prop = node.properties.find((p) => p.key === 'instances');
+  if (!prop) return { min: 1, max: 1 };
+  if (typeof prop.value === 'number') return { min: prop.value, max: prop.value };
+  const str = String(prop.value);
+  const dash = str.indexOf('-');
+  if (dash >= 0) {
+    const min = parseInt(str.substring(0, dash), 10);
+    const max = parseInt(str.substring(dash + 1), 10);
+    return { min: isNaN(min) ? 1 : min, max: isNaN(max) ? min : max };
+  }
+  const num = parseInt(str, 10);
+  return { min: isNaN(num) ? 1 : num, max: isNaN(num) ? 1 : num };
+}
+
+/** Check if a node is serverless (has concurrency property). */
+function isServerless(node: InfraNode): boolean {
+  return getNumProp(node, 'concurrency', 0) > 0;
+}
+
+/** Check if a node is a queue (has buffer property). */
+function isQueue(node: InfraNode): boolean {
+  return getNumProp(node, 'buffer', 0) > 0;
+}
+
+/** Compute serverless effective capacity: concurrency / (duration_ms / 1000). */
+function serverlessCapacity(node: InfraNode): number {
+  const concurrency = getNumProp(node, 'concurrency', 0);
+  const durationMs = getNumProp(node, 'duration-ms', 100);
+  return concurrency / (durationMs / 1000);
+}
+
+/** Backward-compatible helper used by overload detection. */
+function getInstances(node: InfraNode): number {
+  return getInstanceRange(node).min;
+}
+
+/** Compute dynamic instance count based on load and max-rps. */
+function computeDynamicInstances(node: InfraNode, computedRps: number): number {
+  const { min, max } = getInstanceRange(node);
+  const maxRps = getNumProp(node, 'max-rps', 0);
+  if (maxRps <= 0 || min === max) return min;
+  const needed = Math.ceil(computedRps / maxRps);
+  return Math.min(Math.max(needed, min), max);
+}
+
+/** Determine circuit breaker state from thresholds and current load. */
+function computeCbState(node: InfraNode, computedRps: number, computedLatencyMs: number, instanceOverride?: number, groupMultiplier = 1): InfraCbState {
+  const errorThreshold = getNumProp(node, 'cb-error-threshold', 0);
+  const latencyThreshold = getNumProp(node, 'cb-latency-threshold-ms', 0);
+  if (errorThreshold <= 0 && latencyThreshold <= 0) return 'closed';
+
+  // Error-rate based: overloaded nodes have error rate proportional to excess
+  let capacity: number;
+  if (isServerless(node)) {
+    capacity = serverlessCapacity(node);
+  } else {
+    const maxRps = getNumProp(node, 'max-rps', 0);
+    const instances = instanceOverride ?? computeDynamicInstances(node, computedRps);
+    capacity = maxRps > 0 ? maxRps * instances * groupMultiplier : 0;
+  }
+
+  if (errorThreshold > 0 && capacity > 0) {
+    const errorRate = computedRps > capacity ? ((computedRps - capacity) / computedRps) * 100 : 0;
+    if (errorRate >= errorThreshold) return 'open';
+  }
+
+  // Latency-based
+  if (latencyThreshold > 0 && computedLatencyMs > latencyThreshold) return 'open';
+
+  return 'closed';
+}
+
+/**
+ * Compute local availability at a node (0-1).
+ * Factors: uptime, overload shed (rps > capacity), rate-limit reject (rps > ratelimit).
+ * Cache-hit and firewall-block are NOT availability reducers — they're
+ * successful traffic reductions (cache serves, firewall correctly blocks).
+ */
+function computeLocalAvailability(node: InfraNode, inboundRps: number, instanceOverride?: number, groupMultiplier = 1, defaultUptime = 100): number {
+  if (node.isEdge) return 1;
+
+  let avail = 1;
+
+  // Uptime factor
+  const uptime = getNumProp(node, 'uptime', defaultUptime) / 100;
+  avail *= uptime;
+
+  // Queue nodes: availability = 1.0 when buffer has headroom, degrades on overflow
+  // drain-rate scales with group instances (more consumers), buffer does NOT
+  if (isQueue(node)) {
+    const buffer = getNumProp(node, 'buffer', 0);
+    const drainRate = getNumProp(node, 'drain-rate', 0) * groupMultiplier;
+    if (drainRate > 0 && buffer > 0) {
+      const fillRate = Math.max(0, inboundRps - drainRate);
+      if (fillRate > 0) {
+        const timeToOverflow = buffer / fillRate;
+        // If overflow is imminent (< 60s sustained), availability degrades
+        if (timeToOverflow < 60) {
+          avail *= drainRate / inboundRps; // proportional to excess being dropped
+        }
+        // Otherwise buffer is absorbing — availability stays at 1.0
+      }
+    }
+    return avail;
+  }
+
+  // Overload shed: if RPS exceeds total capacity, excess requests fail
+  if (isServerless(node)) {
+    const capacity = serverlessCapacity(node);
+    if (capacity > 0 && inboundRps > capacity) {
+      avail *= capacity / inboundRps;
+    }
+  } else {
+    const maxRps = getNumProp(node, 'max-rps', 0);
+    if (maxRps > 0 && inboundRps > 0) {
+      const instances = instanceOverride ?? computeDynamicInstances(node, inboundRps);
+      const capacity = maxRps * instances * groupMultiplier;
+      if (inboundRps > capacity) {
+        avail *= capacity / inboundRps;
+      }
+    }
+  }
+
+  // Rate-limit reject: requests above the rate limit are rejected
+  // Pre-ratelimit RPS = after cache-hit, firewall-block but before ratelimit
+  const rateLimit = getNumProp(node, 'ratelimit-rps', 0);
+  if (rateLimit > 0 && inboundRps > 0) {
+    let preRateLimitRps = inboundRps;
+    const cacheHit = getNumProp(node, 'cache-hit', 0);
+    if (cacheHit > 0) preRateLimitRps *= (100 - cacheHit) / 100;
+    const fwBlock = getNumProp(node, 'firewall-block', 0);
+    if (fwBlock > 0) preRateLimitRps *= (100 - fwBlock) / 100;
+
+    if (preRateLimitRps > rateLimit) {
+      avail *= rateLimit / preRateLimitRps;
+    }
+  }
+
+  return avail;
+}
+
+/**
+ * Compute the effective outbound rps after applying behavior transformations.
+ * Each behavior reduces the traffic that flows downstream:
+ * - cache-hit: N% → only (100 - N)% passes through
+ * - firewall-block: N% → only (100 - N)% passes through
+ * - ratelimit-rps: N → caps outbound at N rps
+ */
+function applyBehaviors(node: InfraNode, inboundRps: number, groupMultiplier = 1): number {
+  let rps = inboundRps;
+
+  const cacheHit = getNumProp(node, 'cache-hit', 0);
+  if (cacheHit > 0) rps *= (100 - cacheHit) / 100;
+
+  const fwBlock = getNumProp(node, 'firewall-block', 0);
+  if (fwBlock > 0) rps *= (100 - fwBlock) / 100;
+
+  const rateLimit = getNumProp(node, 'ratelimit-rps', 0);
+  if (rateLimit > 0 && rps > rateLimit) rps = rateLimit;
+
+  // Queue: cap outbound at drain-rate (scaled by group instances)
+  if (isQueue(node)) {
+    const drainRate = getNumProp(node, 'drain-rate', 0) * groupMultiplier;
+    if (drainRate > 0 && rps > drainRate) rps = drainRate;
+  }
+
+  return rps;
+}
+
+// ============================================================
+// Group Collapse — rewrite parsed model replacing collapsed groups with virtual nodes
+// ============================================================
+
+/**
+ * Collapse specified groups into virtual nodes.
+ * - Child nodes are removed
+ * - Internal edges are removed
+ * - A virtual node is created with aggregated properties
+ * - External edges are re-routed to/from the virtual node
+ */
+interface CollapseResult {
+  parsed: ParsedInfra;
+  /** Per-group child capacities: each child's max-rps × child instances. */
+  childCapacities: Map<string, number[]>;
+  /** Per-group instance count. */
+  groupInstances: Map<string, number>;
+}
+
+function collapseGroups(parsed: ParsedInfra, collapsedIds: Set<string>, defaultLatencyMs = 0, defaultUptime = 100): CollapseResult {
+  if (collapsedIds.size === 0) return { parsed, childCapacities: new Map(), groupInstances: new Map() };
+
+  // Build child sets per collapsed group
+  const groupChildren = new Map<string, InfraNode[]>();
+  for (const node of parsed.nodes) {
+    if (node.groupId && collapsedIds.has(node.groupId)) {
+      const list = groupChildren.get(node.groupId) ?? [];
+      list.push(node);
+      groupChildren.set(node.groupId, list);
+    }
+  }
+
+  // All child node IDs across all collapsed groups
+  const childIds = new Set<string>();
+  for (const children of groupChildren.values()) {
+    for (const c of children) childIds.add(c.id);
+  }
+
+  // Map child node → its collapsed group
+  const nodeToGroup = new Map<string, string>();
+  for (const node of parsed.nodes) {
+    if (node.groupId && collapsedIds.has(node.groupId)) {
+      nodeToGroup.set(node.id, node.groupId);
+    }
+  }
+
+  // Classify edges
+  const internalEdges = new Set<InfraEdge>();
+  // External edges entering a collapsed group (target is a child)
+  const inboundEdges: InfraEdge[] = [];
+  // External edges leaving a collapsed group (source is a child)
+  const outboundEdges: InfraEdge[] = [];
+  // Cross-group edges (source in group A, target in group B)
+  const crossGroupEdges: InfraEdge[] = [];
+  const otherEdges: InfraEdge[] = [];
+
+  for (const edge of parsed.edges) {
+    const srcInside = childIds.has(edge.sourceId);
+    const tgtInside = childIds.has(edge.targetId);
+    if (srcInside && tgtInside) {
+      const srcGroup = nodeToGroup.get(edge.sourceId);
+      const tgtGroup = nodeToGroup.get(edge.targetId);
+      if (srcGroup === tgtGroup) {
+        internalEdges.add(edge);
+      } else {
+        // Different collapsed groups — re-route both ends
+        crossGroupEdges.push(edge);
+      }
+    } else if (!srcInside && tgtInside) {
+      inboundEdges.push(edge);
+    } else if (srcInside && !tgtInside) {
+      outboundEdges.push(edge);
+    } else {
+      otherEdges.push(edge);
+    }
+  }
+
+  // Build virtual nodes for each collapsed group
+  const virtualNodes: InfraNode[] = [];
+  const childCapacities = new Map<string, number[]>();
+  const groupInstancesMap = new Map<string, number>();
+  for (const group of parsed.groups) {
+    if (!collapsedIds.has(group.id)) continue;
+    const children = groupChildren.get(group.id) ?? [];
+    if (children.length === 0) continue;
+
+    // Aggregate properties: sum latencies, bottleneck capacity, compose behaviors
+    let totalLatency = 0;
+    let minEffectiveCapacity = Infinity;
+    let hasMaxRps = false;
+    let composedUptime = 1;
+    const behaviorProps: InfraProperty[] = [];
+    const perChildCapacities: number[] = [];
+
+    for (const child of children) {
+      // Latency: use explicit value, or diagram default (matching BFS behavior)
+      const latencyProp = child.properties.find((p) => p.key === 'latency-ms');
+      const childIsServerless = child.properties.some((p) => p.key === 'concurrency');
+      if (childIsServerless) {
+        // Serverless nodes use duration-ms as latency contribution
+        const durationProp = child.properties.find((p) => p.key === 'duration-ms');
+        totalLatency += durationProp
+          ? (typeof durationProp.value === 'number' ? durationProp.value : parseFloat(String(durationProp.value)) || 100)
+          : 100;
+      } else if (latencyProp) {
+        totalLatency += (typeof latencyProp.value === 'number' ? latencyProp.value : parseFloat(String(latencyProp.value)) || 0);
+      } else {
+        totalLatency += defaultLatencyMs;
+      }
+
+      const maxRps = child.properties.find((p) => p.key === 'max-rps');
+      if (maxRps) {
+        hasMaxRps = true;
+        const val = typeof maxRps.value === 'number' ? maxRps.value : parseFloat(String(maxRps.value)) || 0;
+        // Effective capacity = max-rps × child instances
+        const childInstProp = child.properties.find((p) => p.key === 'instances');
+        const childInst = childInstProp && typeof childInstProp.value === 'number' ? childInstProp.value : 1;
+        const effectiveCapacity = val * childInst;
+        if (effectiveCapacity < minEffectiveCapacity) minEffectiveCapacity = effectiveCapacity;
+        perChildCapacities.push(effectiveCapacity);
+      }
+
+      // Uptime: use explicit value, or diagram default (matching BFS behavior)
+      const uptimeProp = child.properties.find((p) => p.key === 'uptime');
+      const uptimeVal = uptimeProp
+        ? (typeof uptimeProp.value === 'number' ? uptimeProp.value : parseFloat(String(uptimeProp.value)) || 100)
+        : defaultUptime;
+      composedUptime *= uptimeVal / 100;
+
+      // Collect behavior keys (cache-hit, firewall-block, ratelimit-rps)
+      // and queue/serverless properties
+      for (const prop of child.properties) {
+        if (['cache-hit', 'firewall-block', 'ratelimit-rps',
+             'buffer', 'drain-rate', 'retention-hours', 'partitions',
+             'concurrency', 'duration-ms', 'cold-start-ms'].includes(prop.key)) {
+          behaviorProps.push(prop);
+        }
+      }
+    }
+
+    // Build virtual node properties
+    const props: InfraProperty[] = [];
+    if (totalLatency > 0) props.push({ key: 'latency-ms', value: totalLatency, lineNumber: group.lineNumber });
+
+    // Apply group instances multiplier to bottleneck capacity
+    // minEffectiveCapacity already includes child instances; multiply by group instances
+    const groupInstances = typeof group.instances === 'number' ? group.instances : 1;
+    if (hasMaxRps && minEffectiveCapacity < Infinity) {
+      props.push({ key: 'max-rps', value: Math.round(minEffectiveCapacity * groupInstances), lineNumber: group.lineNumber });
+    }
+    if (composedUptime < 1) {
+      props.push({ key: 'uptime', value: Math.round(composedUptime * 10000) / 100, lineNumber: group.lineNumber });
+    }
+    // Note: we do NOT push `instances` as a node property because max-rps is
+    // already multiplied by groupInstances. Adding instances would double-count
+    // in computeDynamicInstances which multiplies capacity = maxRps × instances.
+    // drain-rate scales with group instances (more consumers), buffer does NOT.
+    for (const bp of behaviorProps) {
+      if (bp.key === 'drain-rate') {
+        const val = typeof bp.value === 'number' ? bp.value : parseFloat(String(bp.value)) || 0;
+        props.push({ ...bp, value: val * groupInstances, lineNumber: group.lineNumber });
+      } else {
+        props.push({ ...bp, lineNumber: group.lineNumber });
+      }
+    }
+
+    // Merge tags from children (first non-empty wins per key)
+    const tags: Record<string, string> = {};
+    for (const child of children) {
+      for (const [k, v] of Object.entries(child.tags)) {
+        if (!(k in tags)) tags[k] = v;
+      }
+    }
+
+    const gInst = typeof group.instances === 'number' ? group.instances : 1;
+    groupInstancesMap.set(group.id, gInst);
+    if (perChildCapacities.length > 0) {
+      childCapacities.set(group.id, perChildCapacities);
+    }
+
+    virtualNodes.push({
+      id: group.id,
+      label: group.label,
+      properties: props,
+      groupId: null,
+      tags,
+      isEdge: false,
+      lineNumber: group.lineNumber,
+    });
+  }
+
+  // Re-route edges through virtual nodes
+  const rewrittenEdges: InfraEdge[] = [...otherEdges];
+  const seenEdgeKeys = new Set<string>();
+
+  // Inbound: target was a child → re-target to group virtual node
+  for (const edge of inboundEdges) {
+    const groupId = nodeToGroup.get(edge.targetId)!;
+    const key = `${edge.sourceId}->${groupId}`;
+    if (seenEdgeKeys.has(key)) continue;
+    seenEdgeKeys.add(key);
+    rewrittenEdges.push({ ...edge, targetId: groupId });
+  }
+
+  // Outbound: source was a child → re-source from group virtual node
+  for (const edge of outboundEdges) {
+    const groupId = nodeToGroup.get(edge.sourceId)!;
+    const key = `${groupId}->${edge.targetId}`;
+    if (seenEdgeKeys.has(key)) continue;
+    seenEdgeKeys.add(key);
+    rewrittenEdges.push({ ...edge, sourceId: groupId });
+  }
+
+  // Cross-group: both source and target in different collapsed groups
+  for (const edge of crossGroupEdges) {
+    const srcGroupId = nodeToGroup.get(edge.sourceId)!;
+    const tgtGroupId = nodeToGroup.get(edge.targetId)!;
+    const key = `${srcGroupId}->${tgtGroupId}`;
+    if (seenEdgeKeys.has(key)) continue;
+    seenEdgeKeys.add(key);
+    rewrittenEdges.push({ ...edge, sourceId: srcGroupId, targetId: tgtGroupId });
+  }
+
+  // Build new node list: non-child nodes + virtual nodes
+  const newNodes = parsed.nodes.filter((n) => !childIds.has(n.id));
+  newNodes.push(...virtualNodes);
+
+  // Groups: keep non-collapsed groups
+  const newGroups = parsed.groups.filter((g) => !collapsedIds.has(g.id));
+
+  return {
+    parsed: {
+      ...parsed,
+      nodes: newNodes,
+      edges: rewrittenEdges,
+      groups: newGroups,
+    },
+    childCapacities,
+    groupInstances: groupInstancesMap,
+  };
+}
+
+// ============================================================
+// Split Resolution
+// ============================================================
+
+interface ResolvedSplit {
+  edge: ParsedInfra['edges'][number];
+  split: number; // 0-100
+}
+
+/**
+ * Resolve splits for outbound edges of a node.
+ * - All declared: validate sum = 100
+ * - None declared: even distribution
+ * - Some declared: remainder distributed evenly among undeclared
+ */
+function resolveSplits(
+  outbound: ParsedInfra['edges'],
+  diagnostics: InfraDiagnostic[],
+): ResolvedSplit[] {
+  if (outbound.length === 0) return [];
+  if (outbound.length === 1) {
+    return [{ edge: outbound[0], split: 100 }];
+  }
+
+  const declared = outbound.filter((e) => e.split !== null);
+  const undeclared = outbound.filter((e) => e.split === null);
+
+  if (declared.length === outbound.length) {
+    // All declared — validate sum
+    const sum = declared.reduce((s, e) => s + (e.split ?? 0), 0);
+    if (Math.abs(sum - 100) > 0.01) {
+      diagnostics.push({
+        type: 'SPLIT_SUM',
+        line: declared[0].lineNumber,
+        message: `Split percentages sum to ${sum}%, expected 100%.`,
+      });
+    }
+    return declared.map((e) => ({ edge: e, split: e.split! }));
+  }
+
+  if (declared.length === 0) {
+    // None declared — even distribution
+    const even = 100 / outbound.length;
+    return outbound.map((e) => ({ edge: e, split: even }));
+  }
+
+  // Some declared, some not — distribute remainder
+  const declaredSum = declared.reduce((s, e) => s + (e.split ?? 0), 0);
+  const remainder = 100 - declaredSum;
+
+  if (remainder < 0) {
+    diagnostics.push({
+      type: 'SPLIT_SUM',
+      line: declared[0].lineNumber,
+      message: `Declared splits sum to ${declaredSum}%, exceeding 100%.`,
+    });
+  }
+
+  const evenRemainder = undeclared.length > 0 ? remainder / undeclared.length : 0;
+
+  return outbound.map((e) => ({
+    edge: e,
+    split: e.split !== null ? e.split : Math.max(0, evenRemainder),
+  }));
+}
+
+// ============================================================
+// Main Computation
+// ============================================================
+
+export function computeInfra(
+  parsed: ParsedInfra,
+  params: InfraComputeParams = {},
+): ComputedInfraModel {
+  const diagnostics: InfraDiagnostic[] = [];
+
+  // Chart-level defaults for latency and uptime
+  const defaultLatencyMs = parseFloat(parsed.options['default-latency-ms'] ?? '') || 0;
+  const defaultUptime = parseFloat(parsed.options['default-uptime'] ?? '') || 100;
+
+  // Apply scenario overrides (shallow clone nodes with modified properties)
+  let effectiveNodes = parsed.nodes;
+  if (params.scenario) {
+    const overrides = params.scenario.overrides;
+    effectiveNodes = parsed.nodes.map((node) => {
+      const nodeOverrides = overrides[node.id];
+      if (!nodeOverrides) return node;
+      const props = node.properties.map((p) => {
+        const ov = nodeOverrides[p.key];
+        return ov != null ? { ...p, value: ov } : p;
+      });
+      // Add new properties from scenario that don't exist on the node
+      for (const [key, val] of Object.entries(nodeOverrides)) {
+        if (!props.some((p) => p.key === key)) {
+          props.push({ key, value: val, lineNumber: node.lineNumber });
+        }
+      }
+      return { ...node, properties: props };
+    });
+  }
+  // Apply per-node property overrides (from interactive sliders)
+  // These take precedence over scenario overrides
+  if (params.propertyOverrides) {
+    const propOv = params.propertyOverrides;
+    effectiveNodes = effectiveNodes.map((node) => {
+      const nodeOv = propOv[node.id];
+      if (!nodeOv) return node;
+      const props = node.properties.map((p) => {
+        const ov = nodeOv[p.key];
+        return ov != null ? { ...p, value: ov } : p;
+      });
+      // Add new properties from overrides that don't exist on the node
+      for (const [key, val] of Object.entries(nodeOv)) {
+        if (!props.some((p) => p.key === key)) {
+          props.push({ key, value: val, lineNumber: node.lineNumber });
+        }
+      }
+      return { ...node, properties: props };
+    });
+  }
+
+  let effectiveParsed = effectiveNodes === parsed.nodes ? parsed : { ...parsed, nodes: effectiveNodes };
+
+  // ── Collapse groups into virtual nodes ──
+  const collapsedGroups = params.collapsedGroups ?? new Set(
+    parsed.groups.filter((g) => g.collapsed).map((g) => g.id)
+  );
+  let collapseChildCaps = new Map<string, number[]>();
+  let collapseGroupInst = new Map<string, number>();
+  if (collapsedGroups.size > 0) {
+    const cr = collapseGroups(effectiveParsed, collapsedGroups, defaultLatencyMs, defaultUptime);
+    effectiveParsed = cr.parsed;
+    collapseChildCaps = cr.childCapacities;
+    collapseGroupInst = cr.groupInstances;
+  }
+
+  // Build lookup maps
+  const nodeMap = new Map<string, InfraNode>();
+  for (const node of effectiveParsed.nodes) {
+    nodeMap.set(node.id, node);
+  }
+
+  // Group edges by source for resolving edges that target groups
+  // If an edge targets a [Group], resolve to the first child node in that group
+  const groupChildMap = new Map<string, string[]>();
+  for (const node of effectiveParsed.nodes) {
+    if (node.groupId) {
+      const children = groupChildMap.get(node.groupId) ?? [];
+      children.push(node.id);
+      groupChildMap.set(node.groupId, children);
+    }
+  }
+
+  // Build group instance multiplier: nodeId → parent group's instances
+  // When a group has `instances: N`, each child's effective capacity is multiplied by N
+  const groupInstMultiplier = new Map<string, number>();
+  for (const group of effectiveParsed.groups) {
+    const gi = typeof group.instances === 'number' ? group.instances :
+      typeof group.instances === 'string' ? parseInt(group.instances, 10) || 1 : 1;
+    if (gi > 1) {
+      const children = groupChildMap.get(group.id) ?? [];
+      for (const childId of children) {
+        groupInstMultiplier.set(childId, gi);
+      }
+    }
+  }
+
+  // Build outbound edge map (sourceId -> edges[])
+  const outboundMap = new Map<string, typeof effectiveParsed.edges>();
+  for (const edge of effectiveParsed.edges) {
+    const list = outboundMap.get(edge.sourceId) ?? [];
+    list.push(edge);
+    outboundMap.set(edge.sourceId, list);
+  }
+
+  // Computed rps per node
+  const computedRps = new Map<string, number>();
+  const computedEdgeRps = new Map<string, number>(); // key: `sourceId->targetId`
+  // Latency: cumulative ms from edge to each node (max across incoming paths)
+  const computedLatency = new Map<string, number>();
+  // Uptime: product of uptimes along the path (min across incoming paths)
+  const computedUptime = new Map<string, number>();
+  // Per-node local availability (computed after RPS is known)
+  const localAvailability = new Map<string, number>();
+
+  // Find edge entry point
+  const edgeNode = effectiveParsed.nodes.find((n) => n.isEdge);
+  const baseRps = params.rps ?? (edgeNode ? getNumProp(edgeNode, 'rps', 0) : 0);
+
+  // BFS traversal from edge
+  if (edgeNode) {
+    computedRps.set(edgeNode.id, baseRps);
+    computedLatency.set(edgeNode.id, 0);
+    computedUptime.set(edgeNode.id, 1);
+
+    const queue: string[] = [edgeNode.id];
+    const visited = new Set<string>();
+
+    while (queue.length > 0) {
+      const currentId = queue.shift()!;
+      if (visited.has(currentId)) continue;
+      visited.add(currentId);
+
+      const currentNode = nodeMap.get(currentId);
+      if (!currentNode) continue;
+
+      const inbound = computedRps.get(currentId) ?? 0;
+      const currentLatency = computedLatency.get(currentId) ?? 0;
+      const currentUptime = computedUptime.get(currentId) ?? 1;
+
+      // Add this node's latency and uptime
+      // Serverless nodes use duration-ms as their latency contribution
+      const nodeLatency = currentNode.isEdge ? 0
+        : isServerless(currentNode) ? getNumProp(currentNode, 'duration-ms', 100)
+        : getNumProp(currentNode, 'latency-ms', defaultLatencyMs);
+      const nodeUptime = currentNode.isEdge ? 1 : getNumProp(currentNode, 'uptime', defaultUptime) / 100;
+      const cumulativeLatency = currentLatency + nodeLatency;
+      const cumulativeUptime = currentUptime * nodeUptime;
+
+      // Apply behavior transformations to get outbound rps
+      const gMul = groupInstMultiplier.get(currentId) ?? 1;
+      const outboundRps = currentNode.isEdge ? inbound : applyBehaviors(currentNode, inbound, gMul);
+
+      // Queue latency boundary: downstream latency starts from queue wait time, not cumulative
+      let downstreamLatency = cumulativeLatency;
+      if (isQueue(currentNode)) {
+        const drainRate = getNumProp(currentNode, 'drain-rate', 0) * gMul;
+        const fillRate = drainRate > 0 ? Math.max(0, inbound - drainRate) : 0;
+        const waitTimeMs = fillRate > 0 && drainRate > 0 ? (fillRate / drainRate) * 1000 : 0;
+        downstreamLatency = waitTimeMs; // reset: consumer-side starts fresh from queue wait
+      }
+
+      // Get outbound edges
+      const outbound = outboundMap.get(currentId) ?? [];
+      if (outbound.length === 0) continue;
+
+      // Resolve splits
+      const resolved = resolveSplits(outbound, diagnostics);
+
+      for (const { edge, split } of resolved) {
+        const edgeRps = outboundRps * (split / 100);
+        const edgeKey = `${edge.sourceId}->${edge.targetId}`;
+        computedEdgeRps.set(edgeKey, edgeRps);
+
+        // Resolve target — could be a group or a node
+        let targetIds: string[];
+        const groupChildren = groupChildMap.get(edge.targetId);
+        if (groupChildren && groupChildren.length > 0) {
+          targetIds = groupChildren;
+        } else {
+          targetIds = [edge.targetId];
+        }
+
+        for (const targetId of targetIds) {
+          const perTarget = edgeRps / targetIds.length;
+          const existing = computedRps.get(targetId) ?? 0;
+          computedRps.set(targetId, existing + perTarget);
+
+          // Latency: take max across incoming paths (worst case)
+          const prevLatency = computedLatency.get(targetId) ?? 0;
+          if (downstreamLatency > prevLatency) {
+            computedLatency.set(targetId, downstreamLatency);
+          }
+
+          // Uptime: take min across incoming paths (most conservative)
+          const prevUptime = computedUptime.get(targetId) ?? 1;
+          if (cumulativeUptime < prevUptime) {
+            computedUptime.set(targetId, cumulativeUptime);
+          }
+
+          if (!visited.has(targetId)) {
+            queue.push(targetId);
+          }
+        }
+      }
+    }
+  }
+
+  // ── Per-node local & compound availability ──
+  // Local availability depends on inbound RPS, so computed after BFS.
+  // Compound availability = product of local availabilities from edge to this node.
+  const instanceOverrides = params.instanceOverrides ?? {};
+  for (const node of effectiveParsed.nodes) {
+    const rps = computedRps.get(node.id) ?? 0;
+    localAvailability.set(node.id, computeLocalAvailability(node, rps, instanceOverrides[node.id], groupInstMultiplier.get(node.id) ?? 1, defaultUptime));
+  }
+
+  // Propagate compound availability via second BFS
+  const compoundAvailability = new Map<string, number>();
+  if (edgeNode) {
+    compoundAvailability.set(edgeNode.id, localAvailability.get(edgeNode.id) ?? 1);
+    const queue2: string[] = [edgeNode.id];
+    const visited2 = new Set<string>();
+    while (queue2.length > 0) {
+      const currentId = queue2.shift()!;
+      if (visited2.has(currentId)) continue;
+      visited2.add(currentId);
+      const cumAvail = compoundAvailability.get(currentId) ?? 1;
+      const currentNode = nodeMap.get(currentId);
+      // Queue boundary: downstream availability starts fresh (producer-side decoupled)
+      const propagatedAvail = currentNode && isQueue(currentNode) ? localAvailability.get(currentId) ?? 1 : cumAvail;
+      const outbound = outboundMap.get(currentId) ?? [];
+      for (const edge of outbound) {
+        const groupChildren = groupChildMap.get(edge.targetId);
+        const targetIds = (groupChildren && groupChildren.length > 0) ? groupChildren : [edge.targetId];
+        for (const targetId of targetIds) {
+          const targetLocal = localAvailability.get(targetId) ?? 1;
+          const newCompound = propagatedAvail * targetLocal;
+          // Take min (most conservative) across incoming paths
+          const prev = compoundAvailability.get(targetId) ?? 1;
+          if (newCompound < prev) {
+            compoundAvailability.set(targetId, newCompound);
+          }
+          if (!visited2.has(targetId)) queue2.push(targetId);
+        }
+      }
+    }
+  }
+
+  // ── Per-node latency percentiles ──
+  // For each node, collect all downstream leaf paths (latency, weight) via DFS,
+  // then derive p50/p90/p99 weighted by traffic.
+
+  type LeafPath = { latency: number; uptime: number; availability: number; weight: number };
+  const nodeLeafCache = new Map<string, LeafPath[]>();
+
+  /** Collect leaf paths reachable from `nodeId` with accumulated extra latency/uptime. */
+  function collectLeafPaths(nodeId: string, visited: Set<string>): LeafPath[] {
+    if (nodeLeafCache.has(nodeId)) return nodeLeafCache.get(nodeId)!;
+    if (visited.has(nodeId)) return []; // cycle guard
+    visited.add(nodeId);
+
+    const node = nodeMap.get(nodeId);
+    if (!node) return [];
+
+    // Node latency contribution depends on type:
+    // - Queue: wait time (latency boundary — resets cumulative chain)
+    // - Serverless: duration-ms
+    // - Normal: latency-ms or default
+    let nodeLatency: number;
+    if (node.isEdge) {
+      nodeLatency = 0;
+    } else if (isQueue(node)) {
+      const qGMul = groupInstMultiplier.get(nodeId) ?? 1;
+      const drainRate = getNumProp(node, 'drain-rate', 0) * qGMul;
+      const inbound = computedRps.get(nodeId) ?? 0;
+      const fillRate = drainRate > 0 ? Math.max(0, inbound - drainRate) : 0;
+      nodeLatency = fillRate > 0 && drainRate > 0 ? (fillRate / drainRate) * 1000 : 0;
+    } else if (isServerless(node)) {
+      nodeLatency = getNumProp(node, 'duration-ms', 100);
+    } else {
+      nodeLatency = getNumProp(node, 'latency-ms', defaultLatencyMs);
+    }
+    const nodeUptimeFrac = node.isEdge ? 1 : getNumProp(node, 'uptime', defaultUptime) / 100;
+    const nodeAvail = localAvailability.get(nodeId) ?? 1;
+
+    // Serverless cold-start latency: split into warm (95%) and cold (5%) paths
+    const coldStartMs = isServerless(node) ? getNumProp(node, 'cold-start-ms', 0) : 0;
+    const coldLatency = nodeLatency + coldStartMs;
+
+    const outbound = outboundMap.get(nodeId) ?? [];
+    if (outbound.length === 0) {
+      // Leaf node — return self
+      const rps = computedRps.get(nodeId) ?? 0;
+      let result: LeafPath[];
+      if (rps <= 0) {
+        result = [];
+      } else if (coldStartMs > 0) {
+        // Split into warm/cold paths for percentile accuracy
+        // 95/5 split: cold starts affect ~5% of requests, visible at p99
+        result = [
+          { latency: nodeLatency, uptime: nodeUptimeFrac, availability: nodeAvail, weight: rps * 0.95 },
+          { latency: coldLatency, uptime: nodeUptimeFrac, availability: nodeAvail, weight: rps * 0.05 },
+        ];
+      } else {
+        result = [{ latency: nodeLatency, uptime: nodeUptimeFrac, availability: nodeAvail, weight: rps }];
+      }
+      nodeLeafCache.set(nodeId, result);
+      return result;
+    }
+
+    // Resolve splits for outbound edges
+    const resolved = resolveSplits(outbound, []);
+    const paths: LeafPath[] = [];
+
+    for (const { edge, split } of resolved) {
+      const groupChildren = groupChildMap.get(edge.targetId);
+      const targetIds = (groupChildren && groupChildren.length > 0)
+        ? groupChildren : [edge.targetId];
+
+      for (const targetId of targetIds) {
+        const childPaths = collectLeafPaths(targetId, new Set(visited));
+        for (const cp of childPaths) {
+          if (coldStartMs > 0) {
+            // Warm path (95% of requests)
+            paths.push({
+              latency: nodeLatency + cp.latency,
+              uptime: nodeUptimeFrac * cp.uptime,
+              availability: nodeAvail * cp.availability,
+              weight: cp.weight * (split / 100) / targetIds.length * 0.95,
+            });
+            // Cold path (5% of requests)
+            paths.push({
+              latency: coldLatency + cp.latency,
+              uptime: nodeUptimeFrac * cp.uptime,
+              availability: nodeAvail * cp.availability,
+              weight: cp.weight * (split / 100) / targetIds.length * 0.05,
+            });
+          } else {
+            paths.push({
+              latency: nodeLatency + cp.latency,
+              uptime: nodeUptimeFrac * cp.uptime,
+              availability: nodeAvail * cp.availability,
+              weight: cp.weight * (split / 100) / targetIds.length,
+            });
+          }
+        }
+      }
+    }
+
+    nodeLeafCache.set(nodeId, paths);
+    return paths;
+  }
+
+  /** Compute percentiles from weighted leaf paths. */
+  function percentilesFromPaths(paths: LeafPath[]): InfraLatencyPercentiles {
+    if (paths.length === 0) return { p50: 0, p90: 0, p99: 0 };
+    const sorted = [...paths].sort((a, b) => a.latency - b.latency);
+    const totalWeight = sorted.reduce((s, p) => s + p.weight, 0);
+    if (totalWeight <= 0) return { p50: 0, p90: 0, p99: 0 };
+
+    const getP = (pct: number): number => {
+      const target = totalWeight * (pct / 100);
+      let cumulative = 0;
+      for (const p of sorted) {
+        cumulative += p.weight;
+        if (cumulative >= target) return p.latency;
+      }
+      return sorted[sorted.length - 1].latency;
+    };
+
+    return { p50: getP(50), p90: getP(90), p99: getP(99) };
+  }
+
+  /** Compute availability percentiles from weighted leaf paths. */
+  function availabilityPercentilesFromPaths(paths: LeafPath[]): InfraAvailabilityPercentiles {
+    if (paths.length === 0) return { p50: 1, p90: 1, p99: 1 };
+    // Sort ascending by availability (worst paths first)
+    const sorted = [...paths].sort((a, b) => a.availability - b.availability);
+    const totalWeight = sorted.reduce((s, p) => s + p.weight, 0);
+    if (totalWeight <= 0) return { p50: 1, p90: 1, p99: 1 };
+
+    const getP = (pct: number): number => {
+      const target = totalWeight * (pct / 100);
+      let cumulative = 0;
+      for (const p of sorted) {
+        cumulative += p.weight;
+        if (cumulative >= target) return p.availability;
+      }
+      return sorted[sorted.length - 1].availability;
+    };
+
+    return { p50: getP(50), p90: getP(90), p99: getP(99) };
+  }
+
+  // Pre-compute leaf paths for all nodes
+  for (const node of effectiveParsed.nodes) {
+    collectLeafPaths(node.id, new Set());
+  }
+
+  // System-wide percentiles (from edge node)
+  const edgeLeafPaths = edgeNode ? (nodeLeafCache.get(edgeNode.id) ?? []) : [];
+  const edgeLatency = percentilesFromPaths(edgeLeafPaths);
+  const systemUptime = edgeLeafPaths.length > 0
+    ? (() => {
+        const tw = edgeLeafPaths.reduce((s, p) => s + p.weight, 0);
+        return tw > 0 ? edgeLeafPaths.reduce((s, p) => s + p.uptime * (p.weight / tw), 0) : 1;
+      })()
+    : 1;
+  const systemAvailability = edgeLeafPaths.length > 0
+    ? (() => {
+        const tw = edgeLeafPaths.reduce((s, p) => s + p.weight, 0);
+        return tw > 0 ? edgeLeafPaths.reduce((s, p) => s + p.availability * (p.weight / tw), 0) : 1;
+      })()
+    : 1;
+
+  // Per-node percentiles + downstream availability map
+  const nodePercentiles = new Map<string, InfraLatencyPercentiles>();
+  const nodeAvailPercentiles = new Map<string, InfraAvailabilityPercentiles>();
+  const downstreamAvailability = new Map<string, number>();
+  for (const node of effectiveParsed.nodes) {
+    const paths = nodeLeafCache.get(node.id) ?? [];
+    nodePercentiles.set(node.id, percentilesFromPaths(paths));
+    nodeAvailPercentiles.set(node.id, availabilityPercentilesFromPaths(paths));
+    // Queue nodes decouple availability — use local availability only
+    if (isQueue(node)) {
+      downstreamAvailability.set(node.id, localAvailability.get(node.id) ?? 1);
+    } else {
+      // Weighted average of compound availability across all downstream leaf paths
+      const tw = paths.reduce((s, p) => s + p.weight, 0);
+      downstreamAvailability.set(node.id, tw > 0
+        ? paths.reduce((s, p) => s + p.availability * (p.weight / tw), 0)
+        : localAvailability.get(node.id) ?? 1);
+    }
+  }
+
+  // Build computed nodes
+  const computedNodes: ComputedInfraNode[] = effectiveParsed.nodes.map((node) => {
+    const rps = computedRps.get(node.id) ?? 0;
+    const gMul = groupInstMultiplier.get(node.id) ?? 1;
+    let capacity: number;
+    let dynInstances: number;
+
+    if (isServerless(node)) {
+      capacity = serverlessCapacity(node);
+      dynInstances = 0; // serverless has no instances
+    } else {
+      const maxRps = getNumProp(node, 'max-rps', 0);
+      dynInstances = instanceOverrides[node.id] ?? computeDynamicInstances(node, rps);
+      capacity = maxRps > 0 ? maxRps * dynInstances * gMul : 0;
+    }
+    const overloaded = capacity > 0 && rps > capacity;
+
+    // Rate-limit check: is inbound RPS (after cache/fw/bot) exceeding ratelimit-rps?
+    let rateLimited = false;
+    if (!node.isEdge) {
+      const rl = getNumProp(node, 'ratelimit-rps', 0);
+      if (rl > 0 && rps > 0) {
+        let preRl = rps;
+        const ch = getNumProp(node, 'cache-hit', 0);
+        if (ch > 0) preRl *= (100 - ch) / 100;
+        const fw = getNumProp(node, 'firewall-block', 0);
+        if (fw > 0) preRl *= (100 - fw) / 100;
+        rateLimited = preRl > rl;
+      }
+    }
+
+    // Serverless nodes use duration-ms as their latency contribution
+    const nodeLatency = node.isEdge ? 0
+      : isServerless(node) ? getNumProp(node, 'duration-ms', 100)
+      : getNumProp(node, 'latency-ms', defaultLatencyMs);
+    const cumulativeLatency = (computedLatency.get(node.id) ?? 0) + nodeLatency;
+    const nodeUptime = node.isEdge ? 1 : getNumProp(node, 'uptime', defaultUptime) / 100;
+    const cumulativeUptime = (computedUptime.get(node.id) ?? 1) * nodeUptime;
+
+    const cbState = computeCbState(node, rps, cumulativeLatency, instanceOverrides[node.id], gMul);
+
+    if (overloaded) {
+      const capDetail = isServerless(node)
+        ? `concurrency ${getNumProp(node, 'concurrency', 0)} / ${getNumProp(node, 'duration-ms', 100)}ms`
+        : `${getNumProp(node, 'max-rps', 0)} x ${dynInstances}${gMul > 1 ? ` x ${gMul} group` : ''}`;
+      diagnostics.push({
+        type: 'OVERLOAD',
+        line: node.lineNumber,
+        message: `${node.label} is overloaded: ${Math.round(rps)} rps exceeds capacity ${Math.round(capacity)} rps (${capDetail}).`,
+      });
+    }
+    if (rateLimited) {
+      diagnostics.push({
+        type: 'RATE_LIMITED',
+        line: node.lineNumber,
+        message: `${node.label} is rate-limiting: inbound traffic exceeds ratelimit-rps.`,
+      });
+    }
+
+    // For collapsed group nodes: compute worst child health state
+    let childHealthState: 'normal' | 'warning' | 'overloaded' | undefined;
+    const childCaps = collapseChildCaps.get(node.id);
+    if (childCaps && childCaps.length > 0) {
+      const gInst = collapseGroupInst.get(node.id) ?? 1;
+      const perInstanceRps = rps / gInst;
+      let worst: 'normal' | 'warning' | 'overloaded' = 'normal';
+      for (const cap of childCaps) {
+        if (cap > 0 && perInstanceRps > cap) worst = 'overloaded';
+        else if (cap > 0 && perInstanceRps > cap * 0.7 && worst !== 'overloaded') worst = 'warning';
+      }
+      childHealthState = worst;
+    }
+
+    // Queue metrics — drain-rate scales with group instances, buffer does NOT
+    let queueMetrics: ComputedInfraNode['queueMetrics'];
+    if (isQueue(node)) {
+      const buffer = getNumProp(node, 'buffer', 0);
+      const drainRate = getNumProp(node, 'drain-rate', 0) * gMul;
+      const fillRate = drainRate > 0 ? Math.max(0, rps - drainRate) : 0;
+      const timeToOverflow = fillRate > 0 ? buffer / fillRate : Infinity;
+      const waitTimeMs = fillRate > 0 && drainRate > 0 ? (fillRate / drainRate) * 1000 : 0;
+      queueMetrics = { fillRate, timeToOverflow, waitTimeMs };
+    }
+
+    return {
+      id: node.id,
+      label: node.label,
+      groupId: node.groupId,
+      isEdge: node.isEdge,
+      computedRps: rps,
+      overloaded,
+      rateLimited,
+      computedLatencyMs: cumulativeLatency,
+      computedLatencyPercentiles: nodePercentiles.get(node.id) ?? { p50: 0, p90: 0, p99: 0 },
+      computedUptime: cumulativeUptime,
+      computedAvailability: downstreamAvailability.get(node.id) ?? 1,
+      computedAvailabilityPercentiles: nodeAvailPercentiles.get(node.id) ?? { p50: 1, p90: 1, p99: 1 },
+      computedCbState: cbState,
+      computedInstances: (collapseGroupInst.get(node.id) ?? dynInstances) * gMul,
+      computedConcurrentInvocations: isServerless(node)
+        ? Math.ceil(rps * getNumProp(node, 'duration-ms', 100) / 1000)
+        : 0,
+      childHealthState,
+      queueMetrics,
+      properties: node.properties,
+      tags: node.tags,
+      lineNumber: node.lineNumber,
+    };
+  });
+
+  // Build computed edges with resolved splits and rps
+  const computedEdges: ComputedInfraEdge[] = effectiveParsed.edges.map((edge) => {
+    const edgeKey = `${edge.sourceId}->${edge.targetId}`;
+    const rps = computedEdgeRps.get(edgeKey) ?? 0;
+
+    // Get resolved split
+    const outbound = outboundMap.get(edge.sourceId) ?? [];
+    let resolvedSplit = edge.split ?? 100;
+    if (outbound.length > 1 && edge.split === null) {
+      // Was inferred
+      const declared = outbound.filter((e) => e.split !== null);
+      if (declared.length === 0) {
+        resolvedSplit = 100 / outbound.length;
+      } else {
+        const declaredSum = declared.reduce((s, e) => s + (e.split ?? 0), 0);
+        const undeclared = outbound.filter((e) => e.split === null);
+        resolvedSplit = undeclared.length > 0 ? (100 - declaredSum) / undeclared.length : 0;
+      }
+    }
+
+    return {
+      sourceId: edge.sourceId,
+      targetId: edge.targetId,
+      label: edge.label,
+      computedRps: rps,
+      split: resolvedSplit,
+      lineNumber: edge.lineNumber,
+    };
+  });
+
+  return {
+    nodes: computedNodes,
+    edges: computedEdges,
+    groups: effectiveParsed.groups,
+    tagGroups: effectiveParsed.tagGroups,
+    title: effectiveParsed.title,
+    direction: effectiveParsed.direction,
+    edgeLatency,
+    systemUptime,
+    systemAvailability,
+    options: parsed.options,
+    diagnostics,
+  };
+}