Skip to main content

Divert all traffic to the winning variant in Mojito

When an experiment you run is wildly successful, you may want to send 100% of eligible traffic to the variant after completion. Usually we recommend hard-coding successful recipes, but when developer resources are scarce and the feature is delivering lots of value, you can use Mojito to divert traffic to the winner temporarily.

Publish a variant to all traffic

Consider the following circumstances:

  • You ran an experiment that increased sales by a large amount and it's time to switch the experiment off
  • The business requires the new feature to be implemented but can't spare the development resources (yet)
  • The business wants to run a follow-up experiment further refining the variants, and has not settled on an execution for production

These scenarios may call for temporary hosting via Mojito.

1. Your experiment whilst live​

Consider the winning recipe was the Green button for the following experiment.

It's live and distributing all traffic, evenly (33.3/33.3/33.3), amongst three recipes:


id: ex4
name: Button colours
state: live
sampleRate: 1
recipes:
  '0':
    name: Red
  '1':
    name: Green
    css: green.css
  '2':
    name: Blue
    css: blue.css
trigger: trigger.js

And after running the experiment, it's wildly successful and you're instructed to urgently roll out the feature...

2. Stop the experiment & divert all traffic to the winning variant​

Stopping and diverting traffic just needs the divertTo flag in the experiment.

Simply add the divertTo with the winning recipe id. Since Green won, we send all traffic to the Green variant (recipe ID 1), like so:


id: ex4
name: Button colours
state: live
sampleRate: 1
divertTo: '1'
recipes:
  '0':
    name: Red
  '1':
    name: Green
    css: green.css
  '2':
    name: Blue
    css: blue.css
trigger: trigger.js

Mojito's divertTo flag is a special flag for full featue roll outs. It enables:

  1. Bypassing a user's prior bucketing (users who would have seen Red or Blue buttons before, will now see Green when divert is enabled) - This is intended to maximise the value of your feature
  2. Prevents exposure tracking from firing during divert mode - So your experiment results won't be invalidated through SRM failure

This allows you to expose as many users as possible to your feature without impacting your experiment results.

Why not divert traffic by changing the recipe sample rates?​

You may feel another approah would be set the recipe sampleRate variables. But we don't recommend this:


id: ex4
name: Button colours
state: live
sampleRate: 1
recipes:
  '0':
    name: Red
    sampleRate: 0
  '1':
    name: Green
    css: green.css
    sampleRate: 1
  '2':
    name: Blue
    css: blue.css
    sampleRate: 0
trigger: trigger.js

Users who had previously been exposed to Red or Blue variants before you changed the sampleRate properties, will continue to see Red or Blue. Mojito persists the variants you were originally exposed to. This is by design, since most experiments assign treatments to user-level units.

Just changing the sample rates won't stop the tracking from firing. Therfore, through this approach, you may trigger an SRM failure in your reports.

Why you should prefer to hard-code your winning variants over using divert​

Keep in mind, that when diverting experiments to 100% of traffic, they may:

Ultimately, it's up to you to decide when and where to use this feature.