discourse/lib/onpdiff.rb
Régis Hanol 53668b23a1
DEV: Add comparison budget to ONPDiff (#38063)
The ONP diff algorithm could run for an extremely long time on
pathological inputs (e.g. fully replaced or random content), causing
request timeouts and high CPU usage. The previous safeguard (`p >=
1000`) did not reliably bound work.

Replace it with an explicit comparison budget that caps the number of
element comparisons in the inner `snake` loop. The budget scales
linearly with input size (default factor: 200× combined length, ceiling:
2 000 000 comparisons). When exceeded, a new `DiffLimitExceeded`
exception is raised.

Handle the new exception gracefully at every call site:
- PostsController#revisions / #latest_revision → 422 with message
- StaffActionLogsController#diff → inline fallback message
- PostRevisor#compute_diff_size → treat as Infinity (forces new version)
- discourse-ai UpdateArtifact → fallback diff block

Also adds a pathology benchmark script and comprehensive test coverage
for all affected paths.

Co-authored-by: Sam Saffron <sam.saffron@gmail.com>
2026-02-25 14:25:03 -05:00

269 lines
6.2 KiB
Ruby

# frozen_string_literal: true
# Use "An O(NP) Sequence Comparison Algorithm" as described by Sun Wu, Udi Manber and Gene Myers
# in https://publications.mpi-cbg.de/Wu_1990_6334.pdf
class ONPDiff
class DiffLimitExceeded < StandardError
attr_reader :comparisons_used, :comparison_budget, :left_size, :right_size
def initialize(comparisons_used:, comparison_budget:, left_size:, right_size:)
@comparisons_used = comparisons_used
@comparison_budget = comparison_budget
@left_size = left_size
@right_size = right_size
super(
"Diff comparison budget exceeded (used=#{comparisons_used}, budget=#{comparison_budget}, left_size=#{left_size}, right_size=#{right_size})",
)
end
end
DEFAULT_COMPARISON_BUDGET_FACTOR = 200
MAX_COMPARISON_BUDGET = 2_000_000
attr_reader :comparison_budget, :comparisons_used
def initialize(
a,
b,
comparison_budget_factor: DEFAULT_COMPARISON_BUDGET_FACTOR,
max_comparison_budget: MAX_COMPARISON_BUDGET
)
@a, @b = a, b
@m, @n = a.size, b.size
@backtrack = []
if @reverse = @m > @n
@a, @b = @b, @a
@m, @n = @n, @m
end
@offset = @m + 1
@delta = @n - @m
@comparison_budget = [comparison_budget_factor * (@m + @n), max_comparison_budget].min
@comparisons_used = 0
end
def diff
@diff ||= build_edit_script(compose)
end
def short_diff
@short_diff ||= build_short_edit_script(compose)
end
def paragraph_diff
@paragraph_diff ||= build_paragraph_edit_script(diff)
end
private
def compose
return @shortest_path if @shortest_path
size = @m + @n + 3
fp = Array.new(size, -1)
@path = Array.new(size, -1)
p = -1
begin
p += 1
k = -p
while k <= @delta - 1
fp[k + @offset] = snake(k, fp[k - 1 + @offset] + 1, fp[k + 1 + @offset])
k += 1
end
k = @delta + p
while k >= @delta + 1
fp[k + @offset] = snake(k, fp[k - 1 + @offset] + 1, fp[k + 1 + @offset])
k -= 1
end
fp[@delta + @offset] = snake(@delta, fp[@delta - 1 + @offset] + 1, fp[@delta + 1 + @offset])
end until fp[@delta + @offset] == @n
r = @path[@delta + @offset]
@shortest_path = []
while r != -1
@shortest_path << [@backtrack[r][0], @backtrack[r][1]]
r = @backtrack[r][2]
end
@shortest_path
end
def snake(k, p, pp)
k_offset = k + @offset
if p > pp
r = @path[k_offset - 1]
y = p
else
r = @path[k_offset + 1]
y = pp
end
x = y - k
while x < @m && y < @n
@comparisons_used += 1
raise_diff_limit_exceeded! if @comparisons_used > @comparison_budget
break if @a[x] != @b[y]
x += 1
y += 1
end
@path[k_offset] = @backtrack.size
@backtrack << [x, y, r]
y
end
def build_edit_script(shortest_path)
ses = []
x, y = 1, 1
px, py = 0, 0
i = shortest_path.size - 1
while i >= 0
while px < shortest_path[i][0] || py < shortest_path[i][1]
if shortest_path[i][1] - shortest_path[i][0] > py - px
t = @reverse ? :delete : :add
ses << [@b[py], t]
y += 1
py += 1
elsif shortest_path[i][1] - shortest_path[i][0] < py - px
t = @reverse ? :add : :delete
ses << [@a[px], t]
x += 1
px += 1
else
ses << [@a[px], :common]
x += 1
y += 1
px += 1
py += 1
end
end
i -= 1
end
ses
end
def build_short_edit_script(shortest_path)
ses = []
x, y = 1, 1
px, py = 0, 0
i = shortest_path.size - 1
while i >= 0
while px < shortest_path[i][0] || py < shortest_path[i][1]
if shortest_path[i][1] - shortest_path[i][0] > py - px
t = @reverse ? :delete : :add
if ses.size > 0 && ses[-1][1] == t
ses[-1][0] << @b[py]
else
ses << [@b[py], t]
end
y += 1
py += 1
elsif shortest_path[i][1] - shortest_path[i][0] < py - px
t = @reverse ? :add : :delete
if ses.size > 0 && ses[-1][1] == t
ses[-1][0] << @a[px]
else
ses << [@a[px], t]
end
x += 1
px += 1
else
if ses.size > 0 && ses[-1][1] == :common
ses[-1][0] << @a[px]
else
ses << [@a[px], :common]
end
x += 1
y += 1
px += 1
py += 1
end
end
i -= 1
end
ses
end
def build_paragraph_edit_script(ses)
paragraph_ses = []
i = 0
while i < ses.size
if ses[i][1] == :common
paragraph_ses << ses[i]
else
if ses[i][1] == :add
op_code = :add
opposite_op_code = :delete
else
op_code = :delete
opposite_op_code = :add
end
j = i + 1
j += 1 while j < ses.size && ses[j][1] == op_code
if j >= ses.size
paragraph_ses = paragraph_ses.concat(ses[i..j])
i = j
else
k = j
j -= 1
k += 1 while k < ses.size && ses[k][1] == opposite_op_code
k -= 1
num_before = j - i + 1
num_after = k - j
if num_after > 1
if num_before > num_after
i2 = i + num_before - num_after
paragraph_ses = paragraph_ses.concat(ses[i..i2 - 1])
i = i2
elsif num_after > num_before
k -= num_after - num_before
end
paragraph_ses = paragraph_ses.concat(pair_paragraphs(ses, i, j))
else
paragraph_ses = paragraph_ses.concat(ses[i..k])
end
i = k
end
end
i += 1
end
paragraph_ses
end
def pair_paragraphs(ses, i, j)
pairs = []
num_pairs = j - i + 1
num_pairs.times do
pairs << ses[i]
pairs << ses[i + num_pairs]
i += 1
end
pairs
end
def raise_diff_limit_exceeded!
raise DiffLimitExceeded.new(
comparisons_used: @comparisons_used,
comparison_budget: @comparison_budget,
left_size: @m,
right_size: @n,
)
end
end