Sam Saffron / Jeff Atwood - Measuring Ruby - Discourse/StackOverflow

- Always be measuring (don't be "playing it blind")
- Dtrace/Stap
  - Added in Ruby 2.0
  - Using Dtrace call samplers to track method invocation counts
- rbtrace
  - like strace but for ruby
  - very awesome
- manual patching for instrumentation
  - old school
  - precise/clean
- rack-mini-profiler
  - demo (see conference video)
  - in depth debugging information for rack requests (SQL tracing, intermethod timing analysis)
  - memory_profiler support for ruby-head (flamegraphs as well)
    - flamegraphs are much more accurate now, more samples per runtime slice
  - embedded callstack sampler in T*S_2
  - exception performance is poor (0.3 ms per exception) non-local transfer of controls are slow -- rack-mini-profiler shows these
  - require is expensive as well, 0.66 ms per require
- improving dev speed
  - traditional fix is to enable asset bundling in Dev, this compromises the dev experience
  - why is this happening?
    - assets are validated after a full middleware crawl
      - etag calculation
      - ....
    - at discourse, shortcircuit this process w/ custom middleware (code in preso)
- memory and ruby
  - in ruby head, memory profiler is new
  - CoW friendly GC lowers memory overhead
- running benchmarks
  - continuously and with new hooks it's much easier to implement
- what can you do?
  - audit memory usage
  - run rack_mini_profiler to integrate performance metrics into dev and production
  - Use flamegraphs to find your slowest code