java-topology/defects/rails/patch/rails-0019-digestor-dependency-digest-array-stack-include.md
russell@unturf.com 39981c70ad diamond-scan: fix collisions, add remaining diamond defects from parallel agents
Renumbering fixes (collisions with pre-existing IDs):
- spring-0001-annotations-scanner → spring-0007 (spring-0001 was already assigned)
- hibernate-0001-class-hierarchy-helper → hibernate-validator-0003 (wrong ecosystem/numbering)
- django-0006-migrations-flatten-bases → django-0007 (django-0006 was already assigned)

New diamond defects from agents that rate-limited before committing:
- micronaut-0007: SuperclassAwareTypeVisitor.getInterfaces O(2^D) (new site)
- quarkus-0005: ConfigMappingUtils.collectInterfacesRec O(2^D) (new site)
- weld-0005: Services.identifyServiceInterfaces O(2^D) (new site)
- typescript-0005: hasBaseType O(2^D) diamond interface hierarchy
- rails-0019: Digestor#dependency_digest Array#include? O(N²) cycle detection

CLEAN: go, rustc (diamond recursion patterns absent)
2026-03-29 20:23:42 -04:00

3.5 KiB
Raw Blame History

UNDF: (pending)

rails-0019: ActionView::Digestor#dependency_digest — O(N²) cycle-detection via Array#include?

CWE-407 — Algorithmic Complexity: O(N²) template dependency stack scan

Field Value
ID rails-0019
Severity MEDIUM
Ecosystem ruby/rails
Package actionview
File actionview/lib/action_view/digestor.rb
Lines 97107
Complexity O(N²) for linear template dependency chains of depth N
Hot path Called on every cache-miss template digest computation during asset compilation and development-mode rendering

Defect

# actionview/lib/action_view/digestor.rb  lines 97-107
def dependency_digest(finder, stack)
  children.map do |node|
    if stack.include?(node)      # DEFECT: Array#include? is O(N) — scans entire stack
      false
    else
      finder.digest_cache[node.name] ||= begin
                                           stack.push node   # grows unboundedly
                                           node.digest(finder, stack).tap { stack.pop }
                                         end
    end
  end.join("-")
end

stack is a plain Array. For a chain of depth N (layout → partial_1 → partial_2 → … → partial_N), each level scans the entire stack to detect cycles:

Depth stack.include? cost Cumulative
1 0 0
2 1 1
3 2 3
k k1 k(k1)/2
N N1 O(N²)

Note: diamond template sharing (A→B→D and A→C→D) is already handled by finder.digest_cache — D is cached after first traversal and returned immediately on the second visit. The O(N²) cost arises purely from the cycle-detection guard on deep linear chains, not from diamond re-traversal.

The caller digest(finder, stack = []) initialises stack as a fresh [] per top-level call, so there is no cross-tree pollution, but the per-call cost is O(N²) in chain depth.

Fix

Replace the Array stack with a Set (or use compare_by_identity on a Set already used nearby in to_dep_map):

# AFTER — O(N) total: Set#include? is O(1) amortised
def digest(finder, stack = Set.new.compare_by_identity)
  ActiveSupport::Digest.hexdigest("#{template.source}-#{dependency_digest(finder, stack)}")
end

def dependency_digest(finder, stack)
  children.map do |node|
    if stack.include?(node)          # O(1) identity check
      false
    else
      finder.digest_cache[node.name] ||= begin
                                           stack.add(node)
                                           node.digest(finder, stack).tap { stack.delete(node) }
                                         end
    end
  end.join("-")
end

Set#compare_by_identity matches the existing to_dep_map(seen = Set.new.compare_by_identity) pattern in the same file (line 110), ensuring object identity is used for cycle detection (consistent with the original Array#include? behaviour which also uses == identity for Node objects with no custom ==).

Speedup

Chain depth N Before (ops) After (ops) Speedup
10 45 10 4.5×
50 1,225 50 24.5×
100 4,950 100 49.5×
500 124,750 500 249.5×
1,000 499,500 1,000 499.5×

Template dependency chains of depth 50100 are realistic in large Rails applications with nested layouts, shared partials, and component hierarchies. At depth 500 (possible in generated or framework-heavy view hierarchies) this is a 249× regression.