Skip sync safekeepers if possible

compute_tools/src/compute.rs

@@ -338,9 +338,13 @@ impl ComputeNode {
         let lsn = match spec.mode {
             ComputeMode::Primary => {
                 info!("starting safekeepers syncing");
-                let lsn = self
-                    .sync_safekeepers(pspec.storage_auth_token.clone())
-                    .with_context(|| "failed to sync safekeepers")?;
+                let lsn = if let Some(synced_lsn) = spec.skip_sync_safekeepers {
+                    info!("no need to sync");
+                    synced_lsn
+                } else {
+                    self.sync_safekeepers(pspec.storage_auth_token.clone())
+                        .with_context(|| "failed to sync safekeepers")?
+                };
                 info!("safekeepers synced at LSN {}", lsn);
                 lsn
             }

control_plane/src/endpoint.rs

@@ -68,6 +68,7 @@ pub struct EndpointConf {
     http_port: u16,
     pg_version: u32,
     skip_pg_catalog_updates: bool,
+    skip_sync_safekeepers: Option<utils::lsn::Lsn>,
 }
 
 //
@@ -137,6 +138,7 @@ impl ComputeControlPlane {
             tenant_id,
             pg_version,
             skip_pg_catalog_updates: false,
+            skip_sync_safekeepers: None,
         });
 
         ep.create_endpoint_dir()?;
@@ -151,6 +153,7 @@ impl ComputeControlPlane {
                 pg_port,
                 pg_version,
                 skip_pg_catalog_updates: false,
+                skip_sync_safekeepers: None,
             })?,
         )?;
         std::fs::write(
@@ -189,6 +192,7 @@ pub struct Endpoint {
 
     // Optimizations
     skip_pg_catalog_updates: bool,
+    skip_sync_safekeepers: Option<utils::lsn::Lsn>,
 }
 
 impl Endpoint {
@@ -223,6 +227,7 @@ impl Endpoint {
             tenant_id: conf.tenant_id,
             pg_version: conf.pg_version,
             skip_pg_catalog_updates: conf.skip_pg_catalog_updates,
+            skip_sync_safekeepers: conf.skip_sync_safekeepers,
         })
     }
 
@@ -457,6 +462,7 @@ impl Endpoint {
 
         // Create spec file
         let spec = ComputeSpec {
+            skip_sync_safekeepers: self.skip_sync_safekeepers,
             skip_pg_catalog_updates: self.skip_pg_catalog_updates,
             format_version: 1.0,
             operation_uuid: None,

libs/compute_api/src/spec.rs

@@ -33,6 +33,15 @@ pub struct ComputeSpec {
     #[serde(default)] // Default false
     pub skip_pg_catalog_updates: bool,
 
+    /// An optinal hint that can be passed to speed up startup time if we know
+    /// that safekeepers have already been synced at the given LSN.
+    ///
+    /// NOTE: If there's any possibility that the safekeepers could have advanced
+    ///       (e.g. if we started compute, and it crashed) we should stay on the
+    ///       safe side and provide None.
+    #[serde(default)]
+    pub skip_sync_safekeepers: Option<Lsn>,
+
     // Information needed to connect to the storage layer.
     //
     // `tenant_id`, `timeline_id` and `pageserver_connstring` are always needed.

test_runner/performance/test_startup.py

@@ -30,7 +30,18 @@ def test_startup_simple(neon_env_builder: NeonEnvBuilder, zenbenchmark: NeonBenc
     neon_env_builder.num_safekeepers = 3
     env = neon_env_builder.init_start()
 
-    env.neon_cli.create_branch("test_startup")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_startup")
+
+    def get_synced_lsn():
+        """Assert safekeepers are synced and get the LSN."""
+        commit_lsns = [
+            sk.http_client().timeline_status(tenant_id, timeline_id).commit_lsn
+            for sk in env.safekeepers
+        ]
+        assert len(commit_lsns) == 3
+        assert len(set(commit_lsns)) == 1
+        return commit_lsns[0]
 
     endpoint = None
 
@@ -63,7 +74,8 @@ def test_startup_simple(neon_env_builder: NeonEnvBuilder, zenbenchmark: NeonBenc
         endpoint.stop()
 
         # Imitate optimizations that console would do for the second start
-        endpoint.respec(skip_pg_catalog_updates=True)
+        lsn = get_synced_lsn()
+        endpoint.respec(skip_pg_catalog_updates=True, skip_sync_safekeepers=lsn.lsn_int)
 
 
 # This test sometimes runs for longer than the global 5 minute timeout.