Sampling
Data Prepper provides the following sampling capabilities:
- Time sampling
- Percentage sampling
- Tail sampling
Time sampling
You can use the rate_limiter
action within the aggregate
processor to limit the number of events that can be processed per second. You can choose to either drop excess events or carry them forward to the next time period.
In the following example, only 100 events with a status code of 200
are sent to the sink per second from a given IP address. The when_exceeds
option is set to drop
, which means that all excess events from the configured time window will be dropped.
...
processor:
- aggregate:
identification_keys: ["clientip"]
action:
rate_limiter:
events_per_second: 100
when_exceeds: drop
when: "/status == 200"
...
If you instead set the when_exceeds
option to block
, the processor will block the pipeline until the time window has elapsed. Then it will process the blocked events.
Percentage sampling
Use the percent_sampler
action within the aggregate
processor to limit the number of events that are sent to a sink. All excess events will be dropped.
In the following example, only 20% of events with a status code of 200
are sent to the sink from a given IP address:
...
processor:
- aggregate:
identification_keys: ["clientip"]
duration :
action:
percent_sampler:
percent: 20
when: "/status == 200"
...
Tail sampling
Use the tail_sampler
action within the aggregate
processor to sample events based on a set of defined policies. This action waits for an aggregation to complete across different aggregation periods based on the configured wait period. When an aggregation is complete, and if it matches the specific error condition, it is sent to the sink. Otherwise, only a configured percentage of events is sent to the sink.
The following pipeline sends all OpenTelemetry traces with an error condition status of 2
to the sink. It only sends 20% of the traces that don’t match this error condition to the sink.
...
processor:
- aggregate:
identification_keys: ["traceId"]
action:
tail_sampler:
percent: 20
wait_period: "10s"
condition: "/status == 2"
...
If you set the error condition to false
or don’t include it, only the configured percentage of events is allowed to pass through, as determined by a probabilistic outcome.
Because it can be difficult to determine exactly when tail sampling should occur, you can use the wait_period
option to measure the idle time since the last event was received.