feat: add capability to purge old histogram data #460
Conversation
- add purge_timeout option to PrometheusBuilder - run a purger that purges based on the purge_timeout
| if let Ok(handle) = runtime::Handle::try_current() { | ||
| handle.spawn(purger); | ||
| } else { | ||
| let thread_name = "metrics-exporter-prometheus-purger"; | ||
|
|
||
| let runtime = runtime::Builder::new_current_thread() | ||
| .enable_all() | ||
| .build() | ||
| .map_err(|e| BuildError::FailedToCreateRuntime(e.to_string()))?; | ||
|
|
||
| thread::Builder::new() | ||
| .name(thread_name.to_owned()) | ||
| .spawn(move || runtime.block_on(purger)) | ||
| .map_err(|e| BuildError::FailedToCreateRuntime(e.to_string()))?; | ||
| }; | ||
| } |
There was a problem hiding this comment.
No need for the conditional logic here: just spawn the future directly.
(We already document that this method must be called from within a Tokio runtime or else it will panic.)
|
|
||
| let exporter_config = self.exporter_config.clone(); | ||
| let recorder = self.build_recorder(); | ||
| let handle = recorder.handle(); | ||
|
|
||
| // use the handle to recorder | ||
| // #[cfg(not(feature = "push-gateway"))] |
|
|
||
| /// Purges registry's histogram data by draining it into the distribution. This should be | ||
| /// called periodically to prevent the accumulation of histogram samples. | ||
| pub fn purge(&self) { |
There was a problem hiding this comment.
I think we should just change the wording overall from "purge" to "upkeep", and make this run_upkeep.
Purging to me sounds like getting rid of, when realistically we're just doing periodic cleanup work.
| /// Sets the purge timeout for metrics. | ||
| /// | ||
| /// If a purge timeout is set, the purger will call `.render()` on the registry, causing | ||
| /// the values from histograms to be drained out. This ensures that stale histogram values | ||
| /// do not persist indefinitely. |
There was a problem hiding this comment.
Similar note here about just changing the wording overall from "purge"/"purging" to "upkeep".
I would also remove the specific bit about calling .render(), since it doesn't actually do that anymore. Just be generic, something like:
/// Sets the upkeep interval.
///
/// The upkeep task handles periodic maintenance operations, such as draining histogram data,
/// to ensure that all recorded data is up-to-date and prevent unbounded memory growth.
| @@ -128,6 +129,7 @@ impl PrometheusBuilder { | |||
| buckets: None, | |||
| bucket_overrides: None, | |||
| idle_timeout: None, | |||
| purge_timeout: None, | |||
There was a problem hiding this comment.
I think we should actually enable this by default.
Thinking more about it, it's a quality of life improvement to avoid unbounded memory growth for people with a high rate of histogram metrics, or who scrape/push their Prometheus exporter infrequently.
Probably 5 seconds as the default?
There was a problem hiding this comment.
Makes sense, 5 seconds is a good default. Also I can't think of a use case where somebody would not want to run upkeep so making this non-optional would be sensible
tobz
added
C-exporter
|
Released in Thanks again for your contribution! |
tobz
removed
the
S-awaiting-release
Copy of #451
What
Implements third way as prescribed here to purge old histogram data:
Fixes #245