Wednesday, May 30, 2012

How do I throttle my site"s API users?

The legitimate users of my site occasionally hammer the server with API requests that cause undesirable results. I want to institute a limit of no more than say one API call every 5 seconds or n calls per minute (haven't figured out the exact limit yet). I could obviously log every API call in a DB and do the calculation on every request to see if they're over the limit, but all this extra overhead on EVERY request would be defeating the purpose. What are other less resource-intensive methods I could use to institute a limit? I'm using PHP/Apache/Linux, for what it's worth.

Source: Tips4all


  1. Ok, there's no way to do what I asked without any writes to the server, but I can at least eliminate logging every single request. One way is by using the "leaky bucket" throttling method, where it only keeps track of the last request ($last_api_request) and a ratio of the number of requests/limit for the time frame ($minute_throttle). The leaky bucket never resets its counter (unlike the Twitter API's throttle which resets every hour), but if the bucket becomes full (user reached the limit), they must wait n seconds for the bucket to empty a little before they can make another request. In other words it's like a rolling limit: if there are previous requests within the time frame, they are slowly leaking out of the bucket; it only restricts you if you fill the bucket.

    This code snippet will calculate a new $minute_throttle value on every request. I specified the minute in $minute_throttle because you can add throttles for any time period, such as hourly, daily, etc... although more than one will quickly start to make it confusing for the users.

    $minute = 60;
    $minute_limit = 100; # users are limited to 100 requests/minute
    $last_api_request = $this->get_last_api_request(); # get from the DB; in epoch seconds
    $last_api_diff = time() - $last_api_request; # in seconds
    $minute_throttle = $this->get_throttle_minute(); # get from the DB
    if ( is_null( $minute_limit ) ) {
    $new_minute_throttle = 0;
    } else {
    $new_minute_throttle = $minute_throttle - $last_api_diff;
    $new_minute_throttle = $new_minute_throttle < 0 ? 0 : $new_minute_throttle;
    $new_minute_throttle += $minute / $minute_limit;
    $minute_hits_remaining = floor( ( $minute - $new_minute_throttle ) * $minute_limit / $minute );
    # can output this value with the request if desired:
    $minute_hits_remaining = $minute_hits_remaining >= 0 ? $minute_hits_remaining : 0;

    if ( $new_minute_throttle > $minute ) {
    $wait = ceil( $new_minute_throttle - $minute );
    usleep( 250000 );
    throw new My_Exception ( 'The one-minute API limit of ' . $minute_limit
    . ' requests has been exceeded. Please wait ' . $wait . ' seconds before attempting again.' )
    # Save the values back to the database.
    $this->save_last_api_request( time() );
    $this->save_throttle_minute( $new_minute_throttle );

  2. You say that "all thos extra overhead on EVERY request would be defeating the purpose", but I'm not sure that's correct. Isn't the purpose to prevent hammering of your server? This is probably the way I would implement it, as it really only requires a quick read/write. You could even farm out the API server checks to a different DB/disk if you were worried about the performance.

    However, if you want alternatives, you should check out mod_cband, a third-party apache module designed to assist in bandwidth throttling. Despite being primarily for bandwidth limiting, it can throttle based on requests-per-second as well. I've never used it, so I'm not sure what kind of results you'd get. There was another module called mod-throttle as well, but that project appears to be closed now, and was never released for anything above the Apache 1.3 series.

  3. Simplest solution would be to just give each API key a limited number of requests per 24 hours, and reset them at some known, fixed, time.

    If they exhaust their API requests (ie. the counter reaches zero, or the limit, depending on the direction you're counting), stop serving them data until you reset their counter.

    This way, it will be in their best interest to not hammer you with requests.