feat(pageserver): create image layers at L0-L1 boundary (neondatabase#12023)

skyzh · web-flow · commit 67ddf1de28e5 · 2025-05-28T07:00:52.000Z
## Problem Previous attempt neondatabase#10548 caused some issues in staging and we reverted it. This is a re-attempt to address neondatabase#11063. Currently we create image layers at latest record LSN. We would create "future image layers" (i.e., image layers with LSN larger than disk consistent LSN) that need special handling at startup. We also waste a lot of read operations to reconstruct from L0 layers while we could have compacted all of the L0 layers and operate on a flat level of historic layers. ## Summary of changes * Run repartition at L0-L1 boundary. * Roll out with feature flags. * Piggyback a change that downgrades "image layer creating below gc_cutoff" to debug level. --------- Signed-off-by: Alex Chi Z <chi@neon.tech>
diff --git a/pageserver/src/tenant.rs b/pageserver/src/tenant.rs
@@ -5315,6 +5315,7 @@ impl TenantShard {
             l0_compaction_trigger: self.l0_compaction_trigger.clone(),
             l0_flush_global_state: self.l0_flush_global_state.clone(),
             basebackup_prepare_sender: self.basebackup_prepare_sender.clone(),
+            feature_resolver: self.feature_resolver.clone(),
         }
     }
 
@@ -8359,10 +8360,24 @@ mod tests {
             }
 
             tline.freeze_and_flush().await?;
+            // Force layers to L1
+            tline
+                .compact(
+                    &cancel,
+                    {
+                        let mut flags = EnumSet::new();
+                        flags.insert(CompactFlags::ForceL0Compaction);
+                        flags
+                    },
+                    &ctx,
+                )
+                .await?;
 
             if iter % 5 == 0 {
+                let scan_lsn = Lsn(lsn.0 + 1);
+                info!("scanning at {}", scan_lsn);
                 let (_, before_delta_file_accessed) =
-                    scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
+                    scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
                         .await?;
                 tline
                     .compact(
@@ -8371,13 +8386,14 @@ mod tests {
                             let mut flags = EnumSet::new();
                             flags.insert(CompactFlags::ForceImageLayerCreation);
                             flags.insert(CompactFlags::ForceRepartition);
+                            flags.insert(CompactFlags::ForceL0Compaction);
                             flags
                         },
                         &ctx,
                     )
                     .await?;
                 let (_, after_delta_file_accessed) =
-                    scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
+                    scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
                         .await?;
                 assert!(
                     after_delta_file_accessed < before_delta_file_accessed,
@@ -8818,6 +8834,8 @@ mod tests {
 
         let cancel = CancellationToken::new();
 
+        // Image layer creation happens on the disk_consistent_lsn so we need to force set it now.
+        tline.force_set_disk_consistent_lsn(Lsn(0x40));
         tline
             .compact(
                 &cancel,
@@ -8831,8 +8849,7 @@ mod tests {
             )
             .await
             .unwrap();
-
-        // Image layers are created at last_record_lsn
+        // Image layers are created at repartition LSN
         let images = tline
             .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
             .await
diff --git a/pageserver/src/tenant/timeline.rs b/pageserver/src/tenant/timeline.rs
@@ -103,6 +103,7 @@ use crate::context::{
     DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
 };
 use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
+use crate::feature_resolver::FeatureResolver;
 use crate::keyspace::{KeyPartitioning, KeySpace};
 use crate::l0_flush::{self, L0FlushGlobalState};
 use crate::metrics::{
@@ -198,6 +199,7 @@ pub struct TimelineResources {
     pub l0_compaction_trigger: Arc<Notify>,
     pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
     pub basebackup_prepare_sender: BasebackupPrepareSender,
+    pub feature_resolver: FeatureResolver,
 }
 
 pub struct Timeline {
@@ -444,6 +446,8 @@ pub struct Timeline {
 
     /// A channel to send async requests to prepare a basebackup for the basebackup cache.
     basebackup_prepare_sender: BasebackupPrepareSender,
+
+    feature_resolver: FeatureResolver,
 }
 
 pub(crate) enum PreviousHeatmap {
@@ -3072,6 +3076,8 @@ impl Timeline {
                 wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
 
                 basebackup_prepare_sender: resources.basebackup_prepare_sender,
+
+                feature_resolver: resources.feature_resolver,
             };
 
             result.repartition_threshold =
@@ -4906,6 +4912,7 @@ impl Timeline {
                     LastImageLayerCreationStatus::Initial,
                     false, // don't yield for L0, we're flushing L0
                 )
+                .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
                 .await?;
             debug_assert!(
                 matches!(is_complete, LastImageLayerCreationStatus::Complete),
@@ -5462,7 +5469,8 @@ impl Timeline {
 
     /// Returns the image layers generated and an enum indicating whether the process is fully completed.
     /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
-    #[tracing::instrument(skip_all, fields(%lsn, %mode))]
+    ///
+    /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
     async fn create_image_layers(
         self: &Arc<Timeline>,
         partitioning: &KeyPartitioning,
diff --git a/pageserver/src/tenant/timeline/compaction.rs b/pageserver/src/tenant/timeline/compaction.rs
@@ -1278,11 +1278,55 @@ impl Timeline {
         }
 
         let gc_cutoff = *self.applied_gc_cutoff_lsn.read();
+        let l0_l1_boundary_lsn = {
+            // We do the repartition on the L0-L1 boundary. All data below the boundary
+            // are compacted by L0 with low read amplification, thus making the `repartition`
+            // function run fast.
+            let guard = self.layers.read().await;
+            guard
+                .all_persistent_layers()
+                .iter()
+                .map(|x| {
+                    // Use the end LSN of delta layers OR the start LSN of image layers.
+                    if x.is_delta {
+                        x.lsn_range.end
+                    } else {
+                        x.lsn_range.start
+                    }
+                })
+                .max()
+        };
+
+        let (partition_mode, partition_lsn) = if cfg!(test)
+            || cfg!(feature = "testing")
+            || self
+                .feature_resolver
+                .evaluate_boolean("image-compaction-boundary", self.tenant_shard_id.tenant_id)
+                .is_ok()
+        {
+            let last_repartition_lsn = self.partitioning.read().1;
+            let lsn = match l0_l1_boundary_lsn {
+                Some(boundary) => gc_cutoff
+                    .max(boundary)
+                    .max(last_repartition_lsn)
+                    .max(self.initdb_lsn)
+                    .max(self.ancestor_lsn),
+                None => self.get_last_record_lsn(),
+            };
+            if lsn <= self.initdb_lsn || lsn <= self.ancestor_lsn {
+                // Do not attempt to create image layers below the initdb or ancestor LSN -- no data below it
+                ("l0_l1_boundary", self.get_last_record_lsn())
+            } else {
+                ("l0_l1_boundary", lsn)
+            }
+        } else {
+            ("latest_record", self.get_last_record_lsn())
+        };
 
         // 2. Repartition and create image layers if necessary
         match self
             .repartition(
-                self.get_last_record_lsn(),
+                partition_lsn,
                 self.get_compaction_target_size(),
                 options.flags,
                 ctx,
@@ -1301,25 +1345,27 @@ impl Timeline {
                     .extend(sparse_partitioning.into_dense().parts);
 
                 // 3. Create new image layers for partitions that have been modified "enough".
+                let mode = if options
+                    .flags
+                    .contains(CompactFlags::ForceImageLayerCreation)
+                {
+                    ImageLayerCreationMode::Force
+                } else {
+                    ImageLayerCreationMode::Try
+                };
                 let (image_layers, outcome) = self
                     .create_image_layers(
                         &partitioning,
                         lsn,
-                        if options
-                            .flags
-                            .contains(CompactFlags::ForceImageLayerCreation)
-                        {
-                            ImageLayerCreationMode::Force
-                        } else {
-                            ImageLayerCreationMode::Try
-                        },
+                        mode,
                         &image_ctx,
                         self.last_image_layer_creation_status
                             .load()
                             .as_ref()
                             .clone(),
                         options.flags.contains(CompactFlags::YieldForL0),
                     )
+                    .instrument(info_span!("create_image_layers", mode = %mode, partition_mode = %partition_mode, lsn = %lsn))
                     .await
                     .inspect_err(|err| {
                         if let CreateImageLayersError::GetVectoredError(
@@ -1344,7 +1390,8 @@ impl Timeline {
             }
 
             Ok(_) => {
-                info!("skipping repartitioning due to image compaction LSN being below GC cutoff");
+                // This happens very frequently so we don't want to log it.
+                debug!("skipping repartitioning due to image compaction LSN being below GC cutoff");
             }
 
             // Suppress errors when cancelled.
diff --git a/test_runner/regress/test_layers_from_future.py b/test_runner/regress/test_layers_from_future.py
@@ -20,6 +20,9 @@
 from fixtures.utils import query_scalar, wait_until
 
 
+@pytest.mark.skip(
+    reason="We won't create future layers any more after https://github.com/neondatabase/neon/pull/10548"
+)
 @pytest.mark.parametrize(
     "attach_mode",
     ["default_generation", "same_generation"],
