gitlab-org--gitlab-foss/lib/declarative_policy/runner.rb
Sean McGivern d771719278 Speed up DeclarativePolicy::Runner#steps_by_score
There were a couple of things here:

1. If the state was already enabled, we don't need to check all the remaining
   steps - only those that can prevent the state. (An enable followed by an
   enable is a no-op.) This logic is in `#run`, but we still did the work of
   scoring and sorting the steps.
2. The sorting is known to be inefficient, but we can make it slightly more
   efficient by stopping once we have a step with zero score, as that means it's
   free.

Neither of these make this _fast_, especially when called lots of times - as we
do when there is lots of activity on an issue - but they do help some.
2017-10-04 13:50:05 +01:00

192 lines
5.2 KiB
Ruby

module DeclarativePolicy
class Runner
class State
def initialize
@enabled = false
@prevented = false
end
def enable!
@enabled = true
end
def enabled?
@enabled
end
def prevent!
@prevented = true
end
def prevented?
@prevented
end
def pass?
!prevented? && enabled?
end
end
# a Runner contains a list of Steps to be run.
attr_reader :steps
def initialize(steps)
@steps = steps
end
# We make sure only to run any given Runner once,
# and just continue to use the resulting @state
# that's left behind.
def cached?
!!@state
end
# used by Rule::Ability. See #steps_by_score
def score
return 0 if cached?
steps.map(&:score).inject(0, :+)
end
def merge_runner(other)
Runner.new(@steps + other.steps)
end
# The main entry point, called for making an ability decision.
# See #run and DeclarativePolicy::Base#can?
def pass?
run unless cached?
@state.pass?
end
# see DeclarativePolicy::Base#debug
def debug(out = $stderr)
run(out)
end
private
def flatten_steps!
@steps = @steps.flat_map { |s| s.flattened(@steps) }
end
# This method implements the semantic of "one enable and no prevents".
# It relies on #steps_by_score for the main loop, and updates @state
# with the result of the step.
def run(debug = nil)
@state = State.new
steps_by_score do |step, score|
return if !debug && @state.prevented?
passed = nil
case step.action
when :enable then
# we only check :enable actions if they have a chance of
# changing the outcome - if no other rule has enabled or
# prevented.
unless @state.enabled? || @state.prevented?
passed = step.pass?
@state.enable! if passed
end
debug << inspect_step(step, score, passed) if debug
when :prevent then
# we only check :prevent actions if the state hasn't already
# been prevented.
unless @state.prevented?
passed = step.pass?
@state.prevent! if passed
end
debug << inspect_step(step, score, passed) if debug
else raise "invalid action #{step.action.inspect}"
end
end
@state
end
# This is the core spot where all those `#score` methods matter.
# It is critical for performance to run steps in the correct order,
# so that we don't compute expensive conditions (potentially n times
# if we're called on, say, a large list of users).
#
# In order to determine the cheapest step to run next, we rely on
# Step#score, which returns a numerical rating of how expensive
# it would be to calculate - the lower the better. It would be
# easy enough to statically sort by these scores, but we can do
# a little better - the scores are cache-aware (conditions that
# are already in the cache have score 0), which means that running
# a step can actually change the scores of other steps.
#
# So! The way we sort here involves re-scoring at every step. This
# is by necessity quadratic, but most of the time the number of steps
# will be low. But just in case, if the number of steps exceeds 50,
# we print a warning and fall back to a static sort.
#
# For each step, we yield the step object along with the computed score
# for debugging purposes.
def steps_by_score(&b)
flatten_steps!
if @steps.size > 50
warn "DeclarativePolicy: large number of steps (#{steps.size}), falling back to static sort"
@steps.map { |s| [s.score, s] }.sort_by { |(score, _)| score }.each do |(score, step)|
yield step, score
end
return
end
remaining_steps = Set.new(@steps)
remaining_enablers, remaining_preventers = remaining_steps.partition(&:enable?).map { |s| Set.new(s) }
loop do
if @state.enabled?
# Once we set this, we never need to unset it, because a single
# prevent will stop this from being enabled
remaining_steps = remaining_preventers
else
# if the permission hasn't yet been enabled and we only have
# prevent steps left, we short-circuit the state here
@state.prevent! if remaining_enablers.empty?
end
return if remaining_steps.empty?
lowest_score = Float::INFINITY
next_step = nil
remaining_steps.each do |step|
score = step.score
if score < lowest_score
next_step = step
lowest_score = score
end
break if lowest_score.zero?
end
[remaining_steps, remaining_enablers, remaining_preventers].each do |set|
set.delete(next_step)
end
yield next_step, lowest_score
end
end
# Formatter for debugging output.
def inspect_step(step, original_score, passed)
symbol =
case passed
when true then '+'
when false then '-'
when nil then ' '
end
"#{symbol} [#{original_score.to_i}] #{step.repr}\n"
end
end
end