[pdm-devel] [PATCH proxmox-datacenter-manager 10/25] metric collection: collect overdue metrics on startup/timer change
Wolfgang Bumiller
w.bumiller at proxmox.com
Thu Feb 13 09:55:28 CET 2025
On Tue, Feb 11, 2025 at 01:05:26PM +0100, Lukas Wagner wrote:
> Due to the fact that the timer fires at aligned points in time and might
> now fire right away after being set up, it could happen that we get gaps
> in the data if we change the timer interval or at daemon startup.
>
> To mitigate this, on daemon startup and also if the collection interval
> changes, we
> - check if the time until the next scheduled regular collection
> plus the time *since* the last successful collection exceeds
> the configured collection interval
> - if yes, we collect immediately
> - if no, we do nothing and let the remote be collected at the
> next timer tick
>
> Signed-off-by: Lukas Wagner <l.wagner at proxmox.com>
> ---
> .../src/metric_collection/collection_task.rs | 72 +++++++++++++++++--
> 1 file changed, 65 insertions(+), 7 deletions(-)
>
> diff --git a/server/src/metric_collection/collection_task.rs b/server/src/metric_collection/collection_task.rs
> index b4e3207..f0742ea 100644
> --- a/server/src/metric_collection/collection_task.rs
> +++ b/server/src/metric_collection/collection_task.rs
> @@ -1,4 +1,7 @@
> -use std::{sync::Arc, time::Duration};
> +use std::{
> + sync::Arc,
> + time::{Duration, Instant},
> +};
>
> use anyhow::Error;
> use rand::Rng;
> @@ -69,13 +72,18 @@ impl MetricCollectionTask {
> /// This function does never return.
> #[tracing::instrument(skip_all, name = "metric_collection_task")]
> pub(super) async fn run(&mut self) {
> - let mut timer = Self::setup_timer(self.settings.collection_interval_or_default());
> + let (mut timer, mut next_run) =
> + Self::setup_timer(self.settings.collection_interval_or_default());
>
> log::debug!(
> "metric collection starting up. collection interval set to {} seconds",
> self.settings.collection_interval_or_default()
> );
>
> + // Check and fetch any remote which would be overdue by the time the
> + // timer first fires.
> + self.fetch_overdue_and_save_state(next_run).await;
> +
> loop {
> let old_settings = self.settings.clone();
> tokio::select! {
> @@ -124,7 +132,12 @@ impl MetricCollectionTask {
> "metric collection interval changed to {} seconds, reloading timer",
> interval
> );
> - timer = Self::setup_timer(interval);
> + (timer, next_run) = Self::setup_timer(interval);
> + // If change (and therefore reset) our timer right before it fires,
> + // we could potentially miss one collection event.
Couldn't we instead just pass `next_run` through to `setup_timer` and
call `reset_at(next_run)` on it? (`first_run` would only be used in the
initial setup, so `next_run` could either be an `Option`, or the setup
code does the `next_aligned_instant` call...
This should be much less code by making the new
`fetch_overdue{,_and_save_sate}()` functions unnecessary, or am I
missing something?
> + // Therefore fetch all remotes which would be due for metric collection before
> + // the new timer fires.
> + self.fetch_overdue_and_save_state(next_run).await;
> }
> }
> }
> @@ -208,12 +221,12 @@ impl MetricCollectionTask {
> /// Set up a [`tokio::time::Interval`] instance with the provided interval.
> /// The timer will be aligned, e.g. an interval of `60` will let the timer
> /// fire at minute boundaries.
> - fn setup_timer(interval: u64) -> Interval {
> + fn setup_timer(interval: u64) -> (Interval, Instant) {
> let mut timer = tokio::time::interval(Duration::from_secs(interval));
> - let first_run = task_utils::next_aligned_instant(interval).into();
> - timer.reset_at(first_run);
> + let first_run = task_utils::next_aligned_instant(interval);
> + timer.reset_at(first_run.into());
>
> - timer
> + (timer, first_run)
> }
>
> /// Convenience helper to load `remote.cfg`, logging the error
> @@ -292,6 +305,51 @@ impl MetricCollectionTask {
> }
> }
>
> + /// Fetch metric data from remotes which are overdue for collection and save
> + /// collection state.
> + async fn fetch_overdue_and_save_state(&mut self, next_run: Instant) {
> + if let Some(remotes) = Self::load_remote_config() {
> + self.fetch_overdue(&remotes, next_run).await;
> + if let Err(e) = self.state.save() {
> + log::error!("could not update metric collection state: {e}");
> + }
> + }
> + }
> +
> + /// Fetch metric data from remotes which are overdue for collection.
> + ///
> + /// Use this on startup of the metric collection loop as well as
> + /// when the collection interval changes.
> + ///
> + /// Does nothing if the remote config could not be read, in this case an
> + /// error is logged.
> + async fn fetch_overdue(&mut self, remotes: &SectionConfigData<Remote>, next_run: Instant) {
> + let left_until_scheduled = next_run - Instant::now();
> + let now = proxmox_time::epoch_i64();
> +
> + let mut overdue = Vec::new();
> +
> + for remote in &remotes.order {
> + let last_collection = self
> + .state
> + .get_status(remote)
> + .and_then(|s| s.last_collection)
> + .unwrap_or(0);
> +
> + let diff = now - last_collection;
> +
> + if diff + left_until_scheduled.as_secs() as i64
> + > self.settings.collection_interval_or_default() as i64
> + {
> + log::debug!(
> + "starting metric collection for remote '{remote}' - triggered because collection is overdue"
> + );
> + overdue.push(remote.clone());
> + }
> + }
> + self.fetch_remotes(remotes, &overdue).await;
> + }
> +
> /// Fetch a single remote.
> #[tracing::instrument(skip_all, fields(remote = remote.id), name = "metric_collection_task")]
> async fn fetch_single_remote(
> --
> 2.39.5
>
>
>
> _______________________________________________
> pdm-devel mailing list
> pdm-devel at lists.proxmox.com
> https://lists.proxmox.com/cgi-bin/mailman/listinfo/pdm-devel
>
>
More information about the pdm-devel
mailing list