feat(pageserver): add compaction_upper_limit config (#10550)

## Problem

Follow-up of the incident, we should not use the same bound on
lower/upper limit of compaction files. This patch adds an upper bound
limit, which is set to 50 for now.

## Summary of changes

Add `compaction_upper_limit`.

---------

Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Christian Schwarz <christian@neon.tech>

control_plane/src/pageserver.rs

@@ -347,6 +347,11 @@ impl PageServerNode {
                 .map(|x| x.parse::<usize>())
                 .transpose()
                 .context("Failed to parse 'compaction_threshold' as an integer")?,
+            compaction_upper_limit: settings
+                .remove("compaction_upper_limit")
+                .map(|x| x.parse::<usize>())
+                .transpose()
+                .context("Failed to parse 'compaction_upper_limit' as an integer")?,
             compaction_algorithm: settings
                 .remove("compaction_algorithm")
                 .map(serde_json::from_str)

libs/pageserver_api/src/config.rs

@@ -256,6 +256,11 @@ pub struct TenantConfigToml {
     pub compaction_period: Duration,
     /// Level0 delta layer threshold for compaction.
     pub compaction_threshold: usize,
+    /// Controls the amount of L0 included in a single compaction iteration.
+    /// The unit is `checkpoint_distance`, i.e., a size.
+    /// We add L0s to the set of layers to compact until their cumulative
+    /// size exceeds `compaction_upper_limit * checkpoint_distance`.
+    pub compaction_upper_limit: usize,
     pub compaction_algorithm: crate::models::CompactionAlgorithmSettings,
     /// Level0 delta layer threshold at which to delay layer flushes for compaction backpressure,
     /// such that they take 2x as long, and start waiting for layer flushes during ephemeral layer
@@ -523,6 +528,12 @@ pub mod tenant_conf_defaults {
 
     pub const DEFAULT_COMPACTION_PERIOD: &str = "20 s";
     pub const DEFAULT_COMPACTION_THRESHOLD: usize = 10;
+
+    // This value needs to be tuned to avoid OOM. We have 3/4 of the total CPU threads to do background works, that's 16*3/4=9 on
+    // most of our pageservers. Compaction ~50 layers requires about 2GB memory (could be reduced later by optimizing L0 hole
+    // calculation to avoid loading all keys into the memory). So with this config, we can get a maximum peak compaction usage of 18GB.
+    pub const DEFAULT_COMPACTION_UPPER_LIMIT: usize = 50;
+
     pub const DEFAULT_COMPACTION_ALGORITHM: crate::models::CompactionAlgorithm =
         crate::models::CompactionAlgorithm::Legacy;
 
@@ -563,6 +574,7 @@ impl Default for TenantConfigToml {
             compaction_period: humantime::parse_duration(DEFAULT_COMPACTION_PERIOD)
                 .expect("cannot parse default compaction period"),
             compaction_threshold: DEFAULT_COMPACTION_THRESHOLD,
+            compaction_upper_limit: DEFAULT_COMPACTION_UPPER_LIMIT,
             compaction_algorithm: crate::models::CompactionAlgorithmSettings {
                 kind: DEFAULT_COMPACTION_ALGORITHM,
             },

libs/pageserver_api/src/models.rs

@@ -458,6 +458,8 @@ pub struct TenantConfigPatch {
     pub compaction_period: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub compaction_threshold: FieldPatch<usize>,
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub compaction_upper_limit: FieldPatch<usize>,
     // defer parsing compaction_algorithm, like eviction_policy
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub compaction_algorithm: FieldPatch<CompactionAlgorithmSettings>,
@@ -522,6 +524,7 @@ pub struct TenantConfig {
     pub compaction_target_size: Option<u64>,
     pub compaction_period: Option<String>,
     pub compaction_threshold: Option<usize>,
+    pub compaction_upper_limit: Option<usize>,
     // defer parsing compaction_algorithm, like eviction_policy
     pub compaction_algorithm: Option<CompactionAlgorithmSettings>,
     pub l0_flush_delay_threshold: Option<usize>,
@@ -559,6 +562,7 @@ impl TenantConfig {
             mut compaction_target_size,
             mut compaction_period,
             mut compaction_threshold,
+            mut compaction_upper_limit,
             mut compaction_algorithm,
             mut l0_flush_delay_threshold,
             mut l0_flush_stall_threshold,
@@ -594,6 +598,9 @@ impl TenantConfig {
             .apply(&mut compaction_target_size);
         patch.compaction_period.apply(&mut compaction_period);
         patch.compaction_threshold.apply(&mut compaction_threshold);
+        patch
+            .compaction_upper_limit
+            .apply(&mut compaction_upper_limit);
         patch.compaction_algorithm.apply(&mut compaction_algorithm);
         patch
             .l0_flush_delay_threshold
@@ -653,6 +660,7 @@ impl TenantConfig {
             compaction_target_size,
             compaction_period,
             compaction_threshold,
+            compaction_upper_limit,
             compaction_algorithm,
             l0_flush_delay_threshold,
             l0_flush_stall_threshold,

pageserver/src/http/openapi_spec.yml

@@ -984,6 +984,8 @@ components:
           type: string
         compaction_threshold:
           type: string
+        compaction_upper_limit:
+          type: string
         image_creation_threshold:
           type: integer
         walreceiver_connect_timeout:

pageserver/src/tenant.rs

@@ -3816,6 +3816,13 @@ impl Tenant {
             .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
     }
 
+    pub fn get_compaction_upper_limit(&self) -> usize {
+        let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
+        tenant_conf
+            .compaction_upper_limit
+            .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
+    }
+
     pub fn get_gc_horizon(&self) -> u64 {
         let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
         tenant_conf
@@ -5469,6 +5476,7 @@ pub(crate) mod harness {
                 compaction_target_size: Some(tenant_conf.compaction_target_size),
                 compaction_period: Some(tenant_conf.compaction_period),
                 compaction_threshold: Some(tenant_conf.compaction_threshold),
+                compaction_upper_limit: Some(tenant_conf.compaction_upper_limit),
                 compaction_algorithm: Some(tenant_conf.compaction_algorithm),
                 l0_flush_delay_threshold: tenant_conf.l0_flush_delay_threshold,
                 l0_flush_stall_threshold: tenant_conf.l0_flush_stall_threshold,

pageserver/src/tenant/config.rs

@@ -277,6 +277,10 @@ pub struct TenantConfOpt {
     #[serde(default)]
     pub compaction_threshold: Option<usize>,
 
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(default)]
+    pub compaction_upper_limit: Option<usize>,
+
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
     pub compaction_algorithm: Option<CompactionAlgorithmSettings>,
@@ -401,6 +405,9 @@ impl TenantConfOpt {
             compaction_threshold: self
                 .compaction_threshold
                 .unwrap_or(global_conf.compaction_threshold),
+            compaction_upper_limit: self
+                .compaction_upper_limit
+                .unwrap_or(global_conf.compaction_upper_limit),
             compaction_algorithm: self
                 .compaction_algorithm
                 .as_ref()
@@ -478,6 +485,7 @@ impl TenantConfOpt {
             mut compaction_target_size,
             mut compaction_period,
             mut compaction_threshold,
+            mut compaction_upper_limit,
             mut compaction_algorithm,
             mut l0_flush_delay_threshold,
             mut l0_flush_stall_threshold,
@@ -519,6 +527,9 @@ impl TenantConfOpt {
             .map(|v| humantime::parse_duration(&v))?
             .apply(&mut compaction_period);
         patch.compaction_threshold.apply(&mut compaction_threshold);
+        patch
+            .compaction_upper_limit
+            .apply(&mut compaction_upper_limit);
         patch.compaction_algorithm.apply(&mut compaction_algorithm);
         patch
             .l0_flush_delay_threshold
@@ -596,6 +607,7 @@ impl TenantConfOpt {
             compaction_target_size,
             compaction_period,
             compaction_threshold,
+            compaction_upper_limit,
             compaction_algorithm,
             l0_flush_delay_threshold,
             l0_flush_stall_threshold,
@@ -657,6 +669,7 @@ impl From<TenantConfOpt> for models::TenantConfig {
             compaction_target_size: value.compaction_target_size,
             compaction_period: value.compaction_period.map(humantime),
             compaction_threshold: value.compaction_threshold,
+            compaction_upper_limit: value.compaction_upper_limit,
             l0_flush_delay_threshold: value.l0_flush_delay_threshold,
             l0_flush_stall_threshold: value.l0_flush_stall_threshold,
             l0_flush_wait_upload: value.l0_flush_wait_upload,

pageserver/src/tenant/timeline.rs

@@ -2181,6 +2181,14 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
     }
 
+    fn get_compaction_upper_limit(&self) -> usize {
+        let tenant_conf = self.tenant_conf.load();
+        tenant_conf
+            .tenant_conf
+            .compaction_upper_limit
+            .unwrap_or(self.conf.default_tenant_conf.compaction_upper_limit)
+    }
+
     fn get_l0_flush_delay_threshold(&self) -> Option<usize> {
         // Disable L0 flushes by default. This and compaction needs further tuning.
         const DEFAULT_L0_FLUSH_DELAY_FACTOR: usize = 0; // TODO: default to e.g. 3

pageserver/src/tenant/timeline/compaction.rs

@@ -47,9 +47,7 @@ use crate::tenant::timeline::{ImageLayerCreationOutcome, IoConcurrency};
 use crate::tenant::timeline::{Layer, ResidentLayer};
 use crate::tenant::{gc_block, DeltaLayer, MaybeOffloaded};
 use crate::virtual_file::{MaybeFatalIo, VirtualFile};
-use pageserver_api::config::tenant_conf_defaults::{
-    DEFAULT_CHECKPOINT_DISTANCE, DEFAULT_COMPACTION_THRESHOLD,
-};
+use pageserver_api::config::tenant_conf_defaults::DEFAULT_CHECKPOINT_DISTANCE;
 
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::KeySpace;
@@ -1117,14 +1115,7 @@ impl Timeline {
         // Under normal circumstances, we will accumulate up to compaction_interval L0s of size
         // checkpoint_distance each.  To avoid edge cases using extra system resources, bound our
         // work in this function to only operate on this much delta data at once.
-        //
-        // Take the max of the configured value & the default, so that tests that configure tiny values
-        // can still use a sensible amount of memory, but if a deployed system configures bigger values we
-        // still let them compact a full stack of L0s in one go.
-        let delta_size_limit = std::cmp::max(
-            self.get_compaction_threshold(),
-            DEFAULT_COMPACTION_THRESHOLD,
-        ) as u64
+        let delta_size_limit = self.get_compaction_upper_limit() as u64
             * std::cmp::max(self.get_checkpoint_distance(), DEFAULT_CHECKPOINT_DISTANCE);
 
         let mut fully_compacted = true;

test_runner/regress/test_attach_tenant_config.py

@@ -139,6 +139,7 @@ def test_fully_custom_config(positive_env: NeonEnv):
     fully_custom_config = {
         "compaction_period": "1h",
         "compaction_threshold": 13,
+        "compaction_upper_limit": 100,
         "l0_flush_delay_threshold": 25,
         "l0_flush_stall_threshold": 42,
         "l0_flush_wait_upload": True,