Sampling

Sampling is used to reduce the amount of tracked data being transported to different destinations.

This can help save ingestion and bandwidth costs.

Filtering vs sampling#

Filtering is the concept of allowing only certain events to be tracked and transported to different destinations if they match certain conditions.

Sampling, on the other hand, chooses a percentage of the tracked events to be transported. Meaning, not all users/devices will transport the tracked events, but only a subset of them.

Consistency#

It is recommended to use a consistent value to sample the events by (like userId attribute), because this will allow maintaining a consistent sampling strategy across different devices and sessions (think web, iOS, and Android apps) for the same user.

Defining sampling#

You can define sampling rules either at events or destinations level.

To keep things simple, we will start with destinations level below:

Simple sampling#

# ...
sample:
  by: attributes.userId
  percentage: 10 # 10% of unique users

This assumes you already have set a userId attribute in your application using Eventvisor SDKs.

Sample by custom range#

Instead of a percentage value, you can also be more specific by defining a custom range:

# ...
sample:
  by: attributes.userId
  range: [0, 50] # first 50% of users

Multiple properties#

Some applications are complex, and you may want to sample based on multiple properties together:

# ...
sample:
  by:
    - attributes.organizationId
    - attributes.userId
  percentage: 10

First available property#

You may also choose to sample based on user's ID if available, otherwise falling back gracefully to device ID.

In that case, you can use the first available value:

# ...
sample:
  by:
    or:
      - attributes.userId
      - attributes.deviceId
  percentage: 10

Conditional sampling#

You can also define multiple sampling rules based on different conditions. The first rule to match will be applied:

# ...
sample:
  - conditions:
      - attribute: platform
        operator: equals
        value: web
    by: attributes.deviceId
    percentage: 10
  # ...more rules here

Different sources#

We used attributes for the sampling strategies above, but you are not limited to them alone. You can use other sources as well:

# ...
sample:
  by:
    attribute: userId
  percentage: 10

Or use a lookup instead:

sample:
  by:
    lookup: localstorage.userId
  percentage: 10

Event level sampling#

Instead of destinations alone, you can also define sampling rules on a per-event basis:

# ...
sample:
  by: attributes.userId
  percentage: 10

Overriding destinations#

If for a certain event, you wish to define sampling rules targeting a specific destination only:

# ...
destinations:
  browser:
    sample:
      by: attributes.userId
      percentage: 10

Event-level sampling decides whether the event continues at all. Destination-level and destination-override sampling then apply independently to each route.

Percentages and ranges use values from 0 through 100 and support three decimal places. Ranges include the start and exclude the end. Sampling fails closed when none of the configured bucket sources resolves.

It is highly recommended that sampling is done at destination level, and event-level can be used for more complex use cases.