Feature: Support "collection_jitter" on a per-input basis

[jcmcken]

Proposal:

Each input should be allowed to specify its own collection_jitter.

Some input collections happen less frequently and with greater expense than others. Having a single, global collection_jitter makes it hard for expensive, long-running collections to co-exist with cheap, short-running collections, and still allow appropriate randomization to avoid collection stampedes while maintaining collection SLAs.

Current behavior:

The collection_jitter option is an agent configuration that applies equally to all inputs.

Desired behavior:

The collection_jitter should additionally be configurable per input. If one input occurs once per hour, it should be allowed to jitter within a large period (e.g. 30min) that might not be appropriate for an input that occurs once per minute.

Use case:

In my specific use-case, we have lots of agents distributed across many machines coordinating with a centralized service to create some statistics per machine. This collection happens once per hour. These agents need to be randomly distributed across the 1 hour period to lessen the load on that central service.

[cruscio]

+1

Specifying an input-specific collection_jitter next to the input-specific interval would go a long way to smoothing traffic spikes. My inbound network traffic graph on Influx data nodes looks like an ECG.

[solarce]

This would be very useful