pageserver: apply 5 fails -> 1hr wait circuit breaker to compaction

pageserver/src/metrics.rs

@@ -569,6 +569,22 @@ static VALID_LSN_LEASE_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+pub(crate) static CIRCUIT_BREAKERS_BROKEN: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pageserver_circuit_breaker_broken",
+        "How many times a circuit breaker has broken"
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static CIRCUIT_BREAKERS_UNBROKEN: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pageserver_circuit_breaker_unbroken",
+        "How many times a circuit breaker has been un-broken (recovered)"
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) mod initial_logical_size {
     use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
     use once_cell::sync::Lazy;

pageserver/src/tenant.rs

@@ -39,6 +39,7 @@ use tokio::task::JoinSet;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::backoff;
+use utils::circuit_breaker::CircuitBreaker;
 use utils::completion;
 use utils::crashsafe::path_with_suffix_extension;
 use utils::failpoint_support;
@@ -76,7 +77,8 @@ use crate::is_uninit_mark;
 use crate::l0_flush::L0FlushGlobalState;
 use crate::metrics::TENANT;
 use crate::metrics::{
-    remove_tenant_metrics, BROKEN_TENANTS_SET, TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
+    remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
+    TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
 };
 use crate::repository::GcResult;
 use crate::task_mgr;
@@ -276,6 +278,10 @@ pub struct Tenant {
 
     eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
 
+    /// Track repeated failures to compact, so that we can back off.
+    /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
+    compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
+
     /// If the tenant is in Activating state, notify this to encourage it
     /// to proceed to Active as soon as possible, rather than waiting for lazy
     /// background warmup.
@@ -1641,13 +1647,31 @@ impl Tenant {
             timelines_to_compact
         };
 
+        // Before doing any I/O work, check our circuit breaker
+        if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
+            info!("Skipping compaction due to previous failures");
+            return Ok(());
+        }
+
         for (timeline_id, timeline) in &timelines_to_compact {
             timeline
                 .compact(cancel, EnumSet::empty(), ctx)
                 .instrument(info_span!("compact_timeline", %timeline_id))
-                .await?;
+                .await
+                .map_err(|e| {
+                    self.compaction_circuit_breaker
+                        .lock()
+                        .unwrap()
+                        .fail(&CIRCUIT_BREAKERS_BROKEN, &e);
+                    e
+                })?;
         }
 
+        self.compaction_circuit_breaker
+            .lock()
+            .unwrap()
+            .success(&CIRCUIT_BREAKERS_UNBROKEN);
+
         Ok(())
     }
 
@@ -2570,6 +2594,14 @@ impl Tenant {
             cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
             cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
             eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
+            compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
+                format!("compaction-{tenant_shard_id}"),
+                5,
+                // Compaction can be a very expensive operation, and might leak disk space.  It also ought
+                // to be infallible, as long as remote storage is available.  So if it repeatedly fails,
+                // use an extremely long backoff.
+                Some(Duration::from_secs(3600 * 24)),
+            )),
             activate_now_sem: tokio::sync::Semaphore::new(0),
             cancel: CancellationToken::default(),
             gate: Gate::default(),