home

Measuring Time Spent Between Steps In A Funnel

May 01, 2012

Last time we talked about Map Reduce we were looking at building basic reports around funnels. Given a series of steps (like pages in a survey), how many people start and subsequently reach each page. Although simplistic, it showcases a real-world example of having the map phase emit multiple times.

What if we also wanted to show the average amount of time people spend on each step? First, our map phase gets more complicated. It goes from:

function() {
  this.steps.forEach(function(s) {
    emit(s.step, 1)
  });
}

To:

function() {
  var length = this.steps.length;
  for(var i = 0; i < length; ++i) {
    var step = this.steps[i];
    var next = this.steps[i+1];
    var time = next ? (next.at - step.at) / 1000 : 0
    emit(step.step, {count: 1,  time: time})
  }
}

Since we care to track more than just an incrementing count, we now emit a more complex value, which includes the our count (1) and the time elapsed (in seconds, hence the / 1000) until the next step. The last step will have a time of 0

.

Given the following data:

[{step: 1, at: 'Jan 1 2014 8:10:00'},
 {step: 2, at: 'Jan 1 2014 8:12:00'},
 {step: 3, at: 'Jan 1 2014 8:13:00'}]

We'd expect the following emits (key is the step):

{key: 1, value: {count: 1, time: 120}}
  {key: 2, value: {count: 1, time: 90}}
  {key: 3, value: {count: 1, time: 0}}

Before we can write our reduce function, there's something we absolutely have to understand about reduce: values associated with a key can be broken into chunks. In fact, this is how reduce is able to scale so well. What exactly does this mean? Well, if we go back to a simple count example where 'about.html' had 4 hits, you might expect reduce to be called with the following values parameter:

[{count: 1}, {count: 1}, {count: 1}, {count: 1}]

And it might, but it might also be called with (or some other variant):

//1st call
  [{count: 1}, {count: 1}]
  //2nd call
  [{count: 1}, {count: 1}]
  //3rd call
  [{count: 2}, {count: 2}]

There are a couple implications to this. First, whatever structure you emit as a value should be the same structure you return from reduce. If you emit('about.html', {count: 1}), then you should return {count: X} from reduce. It also means that the array of values you get as a parameter into reduce probably don't represent all values for that key. Finally, it means that reduce must be written in a way that it can be called multiple times without any side effects.

Why is all this so important for measuring the average time spent on each step? Well, to calculate the average we need to take the total time spent on a step and divide it by the number of people who reached that step. But, as is hopefully clear, we don't necessarily have all the values in a given iteration of reduce.

Consider these are all the values for the key representing the 3rd step of our survey:

[{time: 100},
 {time: 113},
 {time: 120}]

Now, if we knew for sure that we'd get all values at once (which we don't, but let's pretend), we could write a simple reduce function:

function(step, values) {
  if (values.length == 0) { return 0; }

  var total = 0;
  values.forEach(function(value) { total += value.time; });
  return total / values.length;
}

However, what if our engine decided to split this workload? Well, our first call to reduce, which received 100 and 113 would return 106.5 (213/2). Our 2nd call to reduce would get 106.5 and 120 returning a final (and incorrect) value of 113.25.

There are a couple ways to solve this, but the best, by far, is to only care about the total time spent and hits to a specific step. Staying with our above example, we'll emit the following:

[{count: 1, time: 100},
 {count: 1, time: 113},
 {count: 1, time: 120}]

And our final reduce will return: {count:3, time: 333}. Here's the actual code:

function(step, values) {
  var value = {count: 0, time: 0}
  values.forEach(function(v) {
    value.count += v.count;
    value.time += v.time
  });
  return value;
}

Finally, to get our average, we need to run a final step. In MongoDB we could use the finalize option, but we could just as easily do it in calling code:

function(key, value) {
  var average = value.count > 0 ? value.time / value.count : 0;
  return {count: value.count, time: average};
}

There you have it. The number of people and the average time spent on each step.

Normally I'd like to think that I make explanations easy to follow. In this case, things might seem more complicated than they actually are. The takeaway though is that while it's useful to think of the output from map as key => [value1, value2, value3, ....], the values can be broken into chunks and thus reduce can be called multiple times for a given key. This does have serious implication for how you write your reduce function.