Production algorithm tricks anyone?

https://news.ycombinator.com/item?id=14006174

by kartiksura • 9 years ago

14 19 9 years ago

news.ycombinator.com

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?

NathanKP• 9 years ago

99% of the cool, elegant "tricks" I've seen have been applications of the following computer science concepts:

https://en.wikipedia.org/wiki/B-tree

https://en.wikipedia.org/wiki/Hash_table

https://en.wikipedia.org/wiki/Paxos_(computer_science)

tjalfi• 9 years ago

SPITBOL/370 is a SNOBOL4 implementation for IBM mainframes. The following Dr. Dobbs articles describe some of the tricks used in the code. Elegant is the wrong word but both of these techniques are clever hacks.

http://www.drdobbs.com/cpp/misusing-floating-point-arithmeti...

http://www.drdobbs.com/cpp/some-programs-are-poorly-designed...

bogomipz4• 9 years ago

"we had used a random number generator within a segmented number range."

Could you describe your current implementation and its elegance?

kartiksura6• 9 years ago

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• 9 years ago

Not to be a hater, but that seems like by far the most obvious way to do that, right?

borplk• 9 years ago

Yeah not to be an ass but that's pretty standard. And cool too I admit.

bogomipz4• 9 years ago

"40% of the traffic Remaining has to be given to RTB"

What is RTB?

lgas• 9 years ago

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• 9 years ago

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.

kartiksura6• 9 years ago

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• 9 years ago

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• 9 years ago

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.

kartiksura6• 9 years ago

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

Arcten• 9 years ago

Using the HyperLogLog algorithm to estimate the number of unique elements in huge datasets: http://blog.notdot.net/2012/09/Dam-Cool-Algorithms-Cardinali...

Very cool algorithm used in (among other things) redis.

Production algorithm tricks anyone?

Related Stories