Files
garage/src/util/background/worker.rs
T
Gauthier Zirnhelt 2798667345 Fix the LifecycleWorker being uncooperative (#1396)
## Summary

This PR ensures that the `LifecycleWorker` yields at least once to the Tokio scheduler in between each batch of 100 objects.

## Problem being solved

I'm administrating a Garage cluster which has been experiencing timeouts on all endpoints while the lifecycle worker is running at midnight UTC : `Ping timeout` error messages and even requests eventually failing due to `Could not reach quorum ...`.

I have found that this happens while the lifecycle worker is working on a big bucket (containing millions of objects) with a lifecycle rule that applies to very few objects.
The `process_object()` function does not hit any `await`:
- `last_bucket` is always the same, so the `bucket_table` is not read asynchronously
- no transaction is made on the `object_table` because my lifecycle rule (almost) never applies to any object

The first commit in this PR adds an executable which reproduces the problem that I've been experiencing in a self-contained way : the lifecycle worker starves the Tokio scheduler so much that no other task is able to run (or very rarely).
To run it : `cargo run -p garage_model --bin lifecycle-starvation-test`.
This commit can be dropped post-review, as it's only useful to demonstrate the starvation.

The error messages completely stopped after adding the extra yield to the nodes of my cluster.
The duration of the lifecycle worker task does not appear to have changed at all from what I can see (looking at the timestamps produced either by the self-contained binary or by each of my nodes with the `Lifecycle worker finished` message).

## Note

An other potential fix would have been to force the `WorkerProcessor` to yield before re-enqueuing a busy task, but this would have affected all Garage workers even though it's only the `LifecycleWorker` being uncooperative.

Reviewed-on: https://git.deuxfleurs.fr/Deuxfleurs/garage/pulls/1396
Reviewed-by: Alex <lx@deuxfleurs.fr>
Co-authored-by: Gauthier Zirnhelt <gauthier.zirnhelt@insimo.fr>
Co-committed-by: Gauthier Zirnhelt <gauthier.zirnhelt@insimo.fr>
2026-04-15 12:13:18 +02:00

229 lines
6.3 KiB
Rust

use std::collections::HashMap;
use std::sync::Arc;
use std::time::Duration;
use async_trait::async_trait;
use futures::future::*;
use futures::stream::FuturesUnordered;
use futures::StreamExt;
use tokio::select;
use tokio::sync::{mpsc, watch};
use crate::background::{WorkerInfo, WorkerStatus};
use crate::error::Error;
use crate::time::now_msec;
// All workers that haven't exited for this time after an exit signal was received
// will be interrupted in the middle of whatever they are doing.
const EXIT_DEADLINE: Duration = Duration::from_secs(8);
#[derive(PartialEq, Copy, Clone, Debug)]
pub enum WorkerState {
Busy,
Throttled(f32),
Idle,
Done,
}
#[async_trait]
pub trait Worker: Send {
fn name(&self) -> String;
fn status(&self) -> WorkerStatus {
Default::default()
}
/// Work: do a basic unit of work, if one is available (otherwise, should return
/// `WorkerState::Idle` immediately). We will do our best to not interrupt this future in the
/// middle of processing, it will only be interrupted at the last minute when Garage is trying
/// to exit and this hasn't returned yet. This function may return an error to indicate that
/// its unit of work could not be processed due to an error: the error will be logged and
/// .`work()` will be called again after a short delay.
async fn work(&mut self, must_exit: &mut watch::Receiver<bool>) -> Result<WorkerState, Error>;
/// Wait for work: await for some task to become available. This future can be interrupted in
/// the middle for any reason, for example if an interrupt signal was received.
async fn wait_for_work(&mut self) -> WorkerState;
}
pub(crate) struct WorkerProcessor {
stop_signal: watch::Receiver<bool>,
worker_chan: mpsc::UnboundedReceiver<Box<dyn Worker>>,
worker_info: Arc<std::sync::Mutex<HashMap<usize, WorkerInfo>>>,
}
impl WorkerProcessor {
pub(crate) fn new(
worker_chan: mpsc::UnboundedReceiver<Box<dyn Worker>>,
stop_signal: watch::Receiver<bool>,
worker_info: Arc<std::sync::Mutex<HashMap<usize, WorkerInfo>>>,
) -> Self {
Self {
stop_signal,
worker_chan,
worker_info,
}
}
pub(crate) async fn run(&mut self) {
let mut workers = FuturesUnordered::new();
let mut next_task_id = 1;
while !*self.stop_signal.borrow() {
let await_next_worker = async {
if workers.is_empty() {
futures::future::pending().await
} else {
workers.next().await
}
};
select! {
new_worker_opt = self.worker_chan.recv() => {
if let Some(new_worker) = new_worker_opt {
let task_id = next_task_id;
next_task_id += 1;
let stop_signal = self.stop_signal.clone();
let mut worker = WorkerHandler {
task_id,
stop_signal,
worker: new_worker,
state: WorkerState::Busy,
errors: 0,
consecutive_errors: 0,
last_error: None,
};
workers.push(async move {
worker.step().await;
worker
}.boxed());
}
}
worker = await_next_worker => {
if let Some(mut worker) = worker {
trace!("{} (TID {}): {:?}", worker.worker.name(), worker.task_id, worker.state);
// Save worker info
{
let mut wi = self.worker_info.lock().unwrap();
match wi.get_mut(&worker.task_id) {
Some(i) => {
i.state = worker.state;
i.status = worker.worker.status();
i.errors = worker.errors;
i.consecutive_errors = worker.consecutive_errors;
if worker.last_error.is_some() {
i.last_error = worker.last_error.take();
}
}
None => {
wi.insert(worker.task_id, WorkerInfo {
name: worker.worker.name(),
state: worker.state,
status: worker.worker.status(),
errors: worker.errors,
consecutive_errors: worker.consecutive_errors,
last_error: worker.last_error.take(),
});
}
}
}
if worker.state == WorkerState::Done {
info!("Worker {} (TID {}) exited", worker.worker.name(), worker.task_id);
} else {
// Yield to the Tokio scheduler between consecutive Busy steps so
// that a worker which never suspends on its own cannot starve other tasks.
if worker.state == WorkerState::Busy {
tokio::task::yield_now().await;
}
workers.push(async move {
worker.step().await;
worker
}.boxed());
}
}
}
_ = self.stop_signal.changed() => (),
}
}
// We are exiting, drain everything
let drain_everything = async move {
while let Some(worker) = workers.next().await {
info!(
"Worker {} (TID {}) exited (last state: {:?})",
worker.worker.name(),
worker.task_id,
worker.state
);
}
};
select! {
_ = drain_everything => {
info!("All workers exited peacefully \\o/");
}
_ = tokio::time::sleep(EXIT_DEADLINE) => {
error!("Some workers could not exit in time, we are cancelling some things in the middle");
}
}
}
}
struct WorkerHandler {
task_id: usize,
stop_signal: watch::Receiver<bool>,
worker: Box<dyn Worker>,
state: WorkerState,
errors: usize,
consecutive_errors: usize,
last_error: Option<(String, u64)>,
}
impl WorkerHandler {
async fn step(&mut self) {
match self.state {
WorkerState::Busy => match self.worker.work(&mut self.stop_signal).await {
Ok(s) => {
self.state = s;
self.consecutive_errors = 0;
}
Err(e) => {
error!(
"Error in worker {} (TID {}): {}",
self.worker.name(),
self.task_id,
e
);
self.errors += 1;
self.consecutive_errors += 1;
self.last_error = Some((format!("{}", e), now_msec()));
// Sleep a bit so that error won't repeat immediately, exponential backoff
// strategy (min 1sec, max ~60sec)
self.state = WorkerState::Throttled(
(1.5f32).powf(std::cmp::min(10, self.consecutive_errors - 1) as f32),
);
}
},
WorkerState::Throttled(delay) => {
// Sleep for given delay and go back to busy state
select! {
_ = tokio::time::sleep(Duration::from_secs_f32(delay)) => {
self.state = WorkerState::Busy;
}
_ = self.stop_signal.changed() => (),
}
}
WorkerState::Idle => {
select! {
new_st = self.worker.wait_for_work() => {
self.state = new_st;
}
_ = self.stop_signal.changed() => (),
}
}
WorkerState::Done => unreachable!(),
}
}
}