pageserver: exclude un-read layers from short residence statistic (#8396

)

## Problem

The `evictions_with_low_residence_duration` is used as an indicator of
cache thrashing. However, there are situations where it is quite
legitimate to only have a short residence during compaction, where a
delta is downloaded, used to generate an image layer, and then
discarded. This can lead to false positive alerts.

## Summary of changes

- Only track low residence duration for layers that have been accessed
at least once (compaction doesn't count as an access). This will give us
a metric that indicates thrashing on layers that the _user_ is using,
rather than those we're downloading for housekeeping purposes.

Once we add "layer visibility" as an explicit property of layers, this
can also be used as a cleaner condition (residence of non-visible layers
should never be alertable)

pageserver/src/tenant/storage_layer.rs

@@ -676,6 +676,26 @@ impl LayerAccessStats {
             },
         }
     }
+
+    /// Whether this layer has been accessed (excluding in [`AccessStatsBehavior::Skip`]).
+    ///
+    /// This indicates whether the layer has been used for some purpose that would motivate
+    /// us to keep it on disk, such as for serving a getpage request.
+    fn accessed(&self) -> bool {
+        let locked = self.0.lock().unwrap();
+        let inner = &locked.for_eviction_policy;
+
+        // Consider it accessed if the most recent access is more recent than
+        // the most recent change in residence status.
+        match (
+            inner.last_accesses.recent(),
+            inner.last_residence_changes.recent(),
+        ) {
+            (None, _) => false,
+            (Some(_), None) => true,
+            (Some(a), Some(r)) => a.when >= r.timestamp,
+        }
+    }
 }
 
 /// Get a layer descriptor from a layer.

pageserver/src/tenant/storage_layer/layer.rs

@@ -1469,14 +1469,22 @@ impl LayerInner {
                 let duration = SystemTime::now().duration_since(local_layer_mtime);
                 match duration {
                     Ok(elapsed) => {
-                        timeline
-                            .metrics
-                            .evictions_with_low_residence_duration
-                            .read()
-                            .unwrap()
-                            .observe(elapsed);
+                        let accessed = self.access_stats.accessed();
+                        if accessed {
+                            // Only layers used for reads contribute to our "low residence" metric that is used
+                            // to detect thrashing.  Layers promoted for other reasons (e.g. compaction) are allowed
+                            // to be rapidly evicted without contributing to this metric.
+                            timeline
+                                .metrics
+                                .evictions_with_low_residence_duration
+                                .read()
+                                .unwrap()
+                                .observe(elapsed);
+                        }
+
                         tracing::info!(
                             residence_millis = elapsed.as_millis(),
+                            accessed,
                             "evicted layer after known residence period"
                         );
                     }

test_runner/regress/test_tenant_conf.py

@@ -8,6 +8,7 @@
 from fixtures.pageserver.utils import assert_tenant_state, wait_for_upload
 from fixtures.remote_storage import LocalFsStorage, RemoteStorageKind
 from fixtures.utils import wait_until
+from fixtures.workload import Workload
 
 
 def test_tenant_config(neon_env_builder: NeonEnvBuilder):
@@ -265,6 +266,13 @@ def test_live_reconfig_get_evictions_low_residence_duration_metric_threshold(
     (tenant_id, timeline_id) = env.initial_tenant, env.initial_timeline
     ps_http = env.pageserver.http_client()
 
+    # When we evict/download layers, we will use this Workload to generate getpage requests
+    # that touch some layers, as otherwise the pageserver doesn't report totally unused layers
+    # as problems when they have short residence duration.
+    workload = Workload(env, tenant_id, timeline_id)
+    workload.init()
+    workload.write_rows(100)
+
     def get_metric():
         metrics = ps_http.get_metrics()
         metric = metrics.query_one(
@@ -285,6 +293,7 @@ def get_metric():
     assert default_value == "1day"
 
     ps_http.download_all_layers(tenant_id, timeline_id)
+    workload.validate()
     ps_http.evict_all_layers(tenant_id, timeline_id)
     metric = get_metric()
     assert int(metric.value) > 0, "metric is updated"
@@ -305,6 +314,7 @@ def get_metric():
     assert int(metric.value) == 0
 
     ps_http.download_all_layers(tenant_id, timeline_id)
+    workload.validate()
     ps_http.evict_all_layers(tenant_id, timeline_id)
     metric = get_metric()
     assert int(metric.labels["low_threshold_secs"]) == 2 * 24 * 60 * 60
@@ -318,6 +328,7 @@ def get_metric():
     assert int(metric.value) == 0, "value resets if label changes"
 
     ps_http.download_all_layers(tenant_id, timeline_id)
+    workload.validate()
     ps_http.evict_all_layers(tenant_id, timeline_id)
     metric = get_metric()
     assert int(metric.labels["low_threshold_secs"]) == 2 * 60 * 60