2021-09-21-9292-throttle-transform.md

RFC 9292 - 2021-09-21 - throttle transform

This RFC proposes the addition of a new transform that provides a user the ability to control the throughput of specific event streams.

Context

• throttler transform
• Dead letter queues

Scope

In scope

• Dropping logs to rate limit an event stream
• Rate limit by number of events
• Exclude events with conditions
• Optionally specify buckets based on the value of a template string

Out of scope

• Apply backpressure as a form of rate limiting
• Rate limit by size

Pain

• Users lack the necessary tooling to control throughput which can cause excessive costs and negatively impact downstream services due to increased load
• Admins cannot set quotas for users/usergroups utilizing Vector

Proposal

User Experience

The throttle transform can be used to rate limit specific subsets of your event stream to limit load on downstream services or to enforce quotas on users. You can enforce rate limits on number of events, as well as excluding events based on a VRL condition to avoid dropping critical logs. Rate limits can be applied globally across all logs or by specifying a key to create buckets of events to rate limit more granularly.

The initial implementation sheds load by dropping any events above the configured threshold.

Implementation

The throttle transform will leverage the existing Governor crate.

Config:

pub struct ThrottleConfig {
    threshold: u32, // Throttle only on number of events
    window: f64,
    key: Option<String>, // Template string
    exclude: Option<AnyCondition>,
}

TaskTransform:

impl TaskTransform for Throttle {
    fn transform(
        self: Box<Self>,
        mut input_rx: Pin<Box<dyn Stream<Item = Event> + Send>>,
    ) -> Pin<Box<dyn Stream<Item = Event> + Send>>
    where
        Self: 'static,
    {
        let lim = RateLimiter::keyed(quota);

        let mut flush_keys = tokio::time::interval(Duration::from_secs(self.window * 2);
        let mut flush_stream = tokio::time::interval(Duration::from_millis(1000));

        Box::pin(
            stream! {
              loop {
                let mut output = Vec::new();
                let done = tokio::select! {
                    _ = flush_stream.tick() => {
                        false
                    }
                    _ = flush_keys.tick() => {
                      lim.retain_recent();
                          false
                    }
                    maybe_event = input_rx.next() => {
                        match maybe_event {
                            None => true,
                            Some(event) => {
                    if let Some(condition) = self.exclude_as_ref() {
                    if condition.check(&event) {
                        output.push(event);
                        false
                    }
                }

                let value = self
                    .key_field
                    .as_ref()
                    .and_then(|key_field| event.get(key_field))
                    .map(|v| v.to_string_lossy());

                                match lim.check_key_n(value, size) {
                                    Ok(()) => {
                                        output.push(event);
                                        false
                                    }
                                    _ => {
                                        emit!(EventRateLimited);
                                        false
                                    }
                                }
                            }
                        }
                    }
                };
                yield stream::iter(output.into_iter());
                if done { break }
              }
            }
            .flatten(),
        )
    }
}

Rationale

• Controlling throughput provides better reliability for downstream services which can be critical in observability platforms
• Setting granular quotas is invaluable as an administrator running Vector for a number of users to ensure fair use of the pipeline

Drawbacks

• If we allow rate limiting by Bytes in this transform it will be the unserialized form of the event, which will differ from what downstream sinks will actually receive

Prior Art

• Governor
• Throttle - FluentBit
• grouper::bucker - Tremor

Alternatives

• Extend sample transform to allow for a window configuration
• Build rate limiting controls as a generic sink feature, rather than a separate transform

Outstanding Questions

• Rate limiting seems like it could be generically a sink concern and implemented as a composable part of our sink pattern. This could give more "accurate" serialized sizes and possibly be easier to manage for administrators (depending on needs). If rate limiting is also a sink concern should it only be implemented there or also available as a transform? We may additionally add this to sinks, but this transform has value on its own.

Plan Of Attack

• feat(new transform): Initial throttle transform spike
• Add documentations

Future Improvements

• Configure the transform with additional behaviors
• Allow rate limiting against size
• Batching multiple events through the rate limiter
• Optionally persist rate limiter state across restarts