Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Sweetie Bot, Part 3 by Cloud Hop (#PonyDev @ Br...

Sweetie Bot, Part 3 by Cloud Hop (#PonyDev @ BronyCon 2018)

A 10 minute lightning talk at BronyCon 2018's Pony Programming panel. Find more slides at http://developer.horse

Presented by Cloud Hop

Inky Collective

July 29, 2018
Tweet

More Decks by Inky Collective

Other Decks in Programming

Transcript

  1. Introduction Erik  McClure  aka  Cloud  Hop From  Redmond,  Washington Works

     At  Microsoft  (again) Made  Pony  Clicker  (for  some  reason) Sometimes  writes  music  and  fanfiction http://cloudhop.horse twitter:  @blackhole0173
  2. Discord § A  new  chat  and  voice  client  for  gamers

     that  emerged  in  late  2015. § Way  better  than  Skype  in  almost  every  way § Honestly  this  slide  should  probably  be  removed  because  everyone  knows  what  Discord  is.
  3. Bot  Origins § When  Discord  first  emerged,  My  Little  Pony

     chats  were  targeted  by  trolls § Building  an  anti-­‐spam  bot  required  using  a  reversed  engineered,  unsupported  api § The  API  itself  would  often  throw  spurious  errors  or  simply  fail  to  perform  actions § Sweetie  Bot  exploded  to  1000+  servers  with  hundreds  of  thousands  of  users § She  is  still  the  top  google  result  for  “discord  anti-­‐spam  bot”,  for  some  reason. § The  old  code  intended  to  deal  with  a  few  hundred  servers  simply  couldn’t  deal  with  this § Last  time,  I  talked  about  the  database  deadlock  issues.
  4. Debugging  With  Cloud  Hop § This  time,  let’s  talk  about

     Go  deadlocks. § Discordgo is  either  single-­‐threaded  or  launches  a  new  goroutine for  every  event. § Single-­‐threaded  is  bad  because  of  a  rate-­‐limiting  edge  case. § Multi-­‐threaded  is  bad  because  discordgo is  awful. § Sweetie’s  method  of  silencing  requires  solving  a  race-­‐condition § In  this  case,  doing  things  the  right  way  was  a  bad  idea.   § Sweetie’s  deadlocking  was  so  bad  I  built  a  deadlock  detector § It  literally  turns  the  bot  off  and  on  again  if  something  goes  wrong
  5. Fix  Code  By  Introducing  Race  Conditions! § For  the  love

     of  all  that  is  holy,  NEVER  ACTUALLY  DO  THIS. § The  deadlock  is  caused  by  DiscordGo’s terrible  architecture. § One  state  object  holds  state  for  all  servers  and  has  only  one  lock § Trying  to  run  discordgo using  Go’s  data  race  detector  tells  you  that  everything  is  a  data  race § Discordgo assumes  that  you  never  take  write  lock  on  the  state  in  an  event  handler. § This  isn’t  actually  mentioned  anywhere § Fixing  this  requires  rewriting  the  entire  library  from  the  ground  up § Even  just  fixing  this  specific  subcase  would  require  a  pull  request. § So  I  just  assign  the  new  slice  to  the  old  slice  without  a  lock  and  hope  it  doesn’t  crash § It  worked
  6. Lessons  Learned § Discordgo is  a  bad  library § Making

     this  bot  was  a  mistake § The  only  way  to  fix  this  is  to  either  build  a   new  Go  API  library  or  use  another  language § But  the  discord  rate-­‐limit  edge  case  makes  this  a   giant  pain  in  the  ass  in  any  language. § Most  of  this  could  probably  have  been   avoided  with  more  granular  locking. § Unfortunately  discordgo has  more  fundamental   architectural  issues.