HN Explorer

Production algorithm tricks anyone?

14 points by kartiksura ↗ HN
For solving the problem of proportioned distribution of ads into sites at internet-scale, we had used a random number generator within a segmented number range. It was elegant and distributed synchronization and such complexities. Any more of such elegant tricks in production software?

19 comments

[ 3.3 ms ] story [ 54.0 ms ] thread
bogomipz
>"we had used a random number generator within a segmented number range."

Could you describe your current implementation and its elegance?

kartiksura
Ad 1 has to be shown 30% of the traffic Ad 2 has to be shown 40% of the traffic Remaining has to be given to RTB

generate a random number between 0:100 if the number is 0-30 the Ad 1 has to be shown if the number is 30-70 the Ad 1 has to be shown else RTB is to be called

The servers can be now independent and no need to share the state, or have a lock based counter. For large number of requests, and with a good random number generator, you will get good results.

packersfan0000
Not to be a hater, but that seems like by far the most obvious way to do that, right?
kartiksura
Yes. It is intuitive to some. But I never learnt it from a text book. Thats why I was curious about such other tricks.
borplk
Yeah not to be an ass but that's pretty standard. And cool too I admit.
bogomipz
>"Yeah not to be an ass but that's pretty standard."

Pretty standard in what context? Can you elaborate? Is there a name for the algorithm?

kartiksura
It might seem standard among some circles. And i have learnt them on the floor. But could never find a book or blog which describes some more examples. I was curious if others had any ideas to share.
bogomipz
>"40% of the traffic Remaining has to be given to RTB"

What is RTB?

kartiksura
Real time bidding. It is used to trade ads with exchanges. When you dont have your own ads from your sales team.
bogomipz
Ah OK thanks, for somer reason I was thinking the TB in RTB was token bucket and I was scratching my head at the first letter. Now I feel silly. Cheers.
lgas
For scaling view counters on the pages of popular videos Youtube increments the counter by N views with 1/N probability for each actual view. Or so I've been told.
ageitgey
I've seen similar approaches used to enforce rate limit checks on APIs that get lots of traffic and are distributed amongst many different servers.
kartiksura
This is really neat. You need to take only 1/N of the locks. Maybe the N can be started small and changed gradually as the views increase.
shaftway
I've always been a fan of the Fast Inverse Square Root hack.

  float Q_rsqrt( float number )
  {
    long i;
    float x2, y;
    const float threehalfs = 1.5F;
  
    x2 = number * 0.5F;
    y  = number;
    i  = * ( long * ) &y;                       // evil floating point bit level hacking
    i  = 0x5f3759df - ( i >> 1 );               // what the fuck? 
    y  = * ( float * ) &i;
    y  = y * ( threehalfs - ( x2 * y * y ) );   // 1st iteration
    // y  = y * ( threehalfs - ( x2 * y * y ) );   // 2nd iteration, this can be removed
  
    return y;
  }
https://en.wikipedia.org/wiki/Fast_inverse_square_root
anon263626
There's a bunch in Hacker's Delight.

One that comes to mind is XOR-swap. It's cute, but it's likely slower than using a temp var.

Some of the integer population count algos (1's counting) are clever.

Also, original Doom source has some neat tricks.

kartiksura
Aah finally a good clue about reading material. Thanks a ton.