Reduce number of String allocations

[nashbridges]

Here's a memory allocation report taken from one of our Rails endpoints with miniprofiler in production:

Total allocated: 36329328 bytes (541884 objects)
Total retained:  4395083 bytes (18004 objects)

allocated memory by gem
-----------------------------------
  14044335  transit-ruby-ff2e3acd071a
   4148691  temple-0.7.6
   3993243  slim-3.0.6
   3794721  activesupport-5.0.7
   2622300  actionview-5.0.7
   1827160  front/app
   1686968  other
    834761  dalli-2.7.6
    787136  transit-rails-6d6b533ba1df
    494528  set
    491982  actionpack-5.0.7
    430072  activerecord-5.0.7
    315464  concurrent-ruby-1.0.5
   ...

You see that transit-ruby is at the top, which is no surprise because the endpoint renders a heavy transit response. But if we look closely on allocations grouped by file, then there's definitely a room for improvement:

We can do nothing with cruby/json.rb, because it's the oj's part:

transit-ruby/lib/transit/marshaler/cruby/json.rb

Lines 31 to 39 in b4973f8

	def emit_array_start(size)
	@state << :array
	@oj.push_array
	end
	def emit_array_end
	@state.pop
	@oj.pop
	end

transit-ruby/lib/transit/marshaler/cruby/json.rb

Lines 59 to 65 in b4973f8

	def emit_value(obj, as_map_key=false)
	if @state.last == :array
	@oj.push_value(obj)
	else
	as_map_key ? @oj.push_key(obj) : @oj.push_value(obj)
	end
	end

Same with the marshaler/base.rb, which produces a string interpolation each time:

transit-ruby/lib/transit/marshaler/base.rb

Lines 87 to 94 in b4973f8

	def emit_string(prefix, tag, value, as_map_key, cache)
	encoded = "#{prefix}#{tag}#{value}"
	if cache.cacheable?(encoded, as_map_key)
	emit_value(cache.write(encoded), as_map_key)
	else
	emit_value(encoded, as_map_key)
	end
	end

But write handlers produce a large amount of static strings for nothing (lines 210, 234):

transit-ruby/lib/transit/write_handlers.rb

Lines 209 to 213 in b4973f8

	class KeywordHandler
	def tag(_) ":" end
	def rep(s) s.to_s end
	def string_rep(s) rep(s) end
	end

transit-ruby/lib/transit/write_handlers.rb

Lines 233 to 237 in b4973f8

	class IntHandler
	def tag(i) i > MAX_INT || i < MIN_INT ? "n" : "i" end
	def rep(i) i > MAX_INT || i < MIN_INT ? i.to_s : i end
	def string_rep(i) i.to_s end
	end

Line 211 also produces a large amount of strings, but it's a symbol to string conversion, which I believe is inevitable.

Another offender is RollingCache:

transit-ruby/lib/transit/rolling_cache.rb

Lines 49 to 51 in b4973f8

	def cacheable?(str, as_map_key=false)
	str.size >= MIN_SIZE_CACHEABLE && (as_map_key || str.start_with?("~#","~$","~:"))
	end

Solution

After adding # frozen_string_literal: true magic comment to the top of the mentioned files, I see the much nicer picture:

Total allocated: 31684710 bytes (425875 objects)
Total retained:  4398650 bytes (18009 objects)

allocated memory by gem
-----------------------------------
   9395815  transit-ruby/lib
   4148546  temple-0.7.6
   3993613  slim-3.0.6
   3796808  activesupport-5.0.7
   2621277  actionview-5.0.7
   1830296  front/app
   1686968  other
    834761  dalli-2.7.6
    787136  transit-rails-6d6b533ba1df
    494528  set
    491790  actionpack-5.0.7
    429992  activerecord-5.0.7
    315464  concurrent-ruby-1.0.5

that is, 20% less object allocations in total.

As for execution speed, I don't see any improvements in my local tests, but having less pressure on GC is always good.

Usually, those comments are being added throughout entire library to be Rails 3 ready.