pass multiple safekeeper hostaddrs to primary_conninfo

Usage example: neon_local pg start main_roo --replicates main

control_plane/src/bin/neon_local.rs

@@ -8,6 +8,7 @@
 use anyhow::{anyhow, bail, Context, Result};
 use clap::{value_parser, Arg, ArgAction, ArgMatches, Command};
 use control_plane::compute::ComputeControlPlane;
+use control_plane::compute::Replication;
 use control_plane::local_env::LocalEnv;
 use control_plane::pageserver::PageServerNode;
 use control_plane::safekeeper::SafekeeperNode;
@@ -472,7 +473,14 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
             println!("Creating node for imported timeline ...");
             env.register_branch_mapping(name.to_string(), tenant_id, timeline_id)?;
 
-            cplane.new_node(tenant_id, name, timeline_id, None, None, pg_version)?;
+            cplane.new_node(
+                tenant_id,
+                name,
+                timeline_id,
+                None,
+                pg_version,
+                Replication::Primary,
+            )?;
             println!("Done");
         }
         Some(("branch", branch_match)) => {
@@ -558,20 +566,19 @@ fn handle_pg(pg_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
                 .iter()
                 .filter(|((node_tenant_id, _), _)| node_tenant_id == &tenant_id)
             {
-                let lsn_str = match node.lsn {
-                    None => {
-                        // -> primary node
+                let lsn_str = match node.replication {
+                    Replication::Static(lsn) => {
+                        // -> read-only node
+                        // Use the node's LSN.
+                        lsn.to_string()
+                    }
+                    _ => {
                         // Use the LSN at the end of the timeline.
                         timeline_infos
                             .get(&node.timeline_id)
                             .map(|bi| bi.last_record_lsn.to_string())
                             .unwrap_or_else(|| "?".to_string())
                     }
-                    Some(lsn) => {
-                        // -> read-only node
-                        // Use the node's LSN.
-                        lsn.to_string()
-                    }
                 };
 
                 let branch_name = timeline_name_mappings
@@ -617,7 +624,26 @@ fn handle_pg(pg_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
                 .copied()
                 .context("Failed to parse postgres version from the argument string")?;
 
-            cplane.new_node(tenant_id, &node_name, timeline_id, lsn, port, pg_version)?;
+            let hot_standby = sub_args
+                .get_one::<bool>("hot-standby")
+                .copied()
+                .unwrap_or(false);
+
+            let replication = match (lsn, hot_standby) {
+                (Some(lsn), false) => Replication::Static(lsn),
+                (None, true) => Replication::Replica,
+                (None, false) => Replication::Primary,
+                (Some(_), true) => anyhow::bail!("cannot specify both lsn and hot-standby"),
+            };
+
+            cplane.new_node(
+                tenant_id,
+                &node_name,
+                timeline_id,
+                port,
+                pg_version,
+                replication,
+            )?;
         }
         "start" => {
             let port: Option<u16> = sub_args.get_one::<u16>("port").copied();
@@ -635,7 +661,21 @@ fn handle_pg(pg_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
                 None
             };
 
+            let hot_standby = sub_args
+                .get_one::<bool>("hot-standby")
+                .copied()
+                .unwrap_or(false);
+
             if let Some(node) = node {
+                match (&node.replication, hot_standby) {
+                    (Replication::Static(_), true) => {
+                        bail!("Cannot start a node in hot standby mode when it is already configured as a static replica")
+                    }
+                    (Replication::Primary, true) => {
+                        bail!("Cannot start a node as a hot standby replica, it is already configured as primary node")
+                    }
+                    _ => {}
+                }
                 println!("Starting existing postgres {node_name}...");
                 node.start(&auth_token)?;
             } else {
@@ -657,6 +697,14 @@ fn handle_pg(pg_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
                     .get_one::<u32>("pg-version")
                     .copied()
                     .context("Failed to `pg-version` from the argument string")?;
+
+                let replication = match (lsn, hot_standby) {
+                    (Some(lsn), false) => Replication::Static(lsn),
+                    (None, true) => Replication::Replica,
+                    (None, false) => Replication::Primary,
+                    (Some(_), true) => anyhow::bail!("cannot specify both lsn and hot-standby"),
+                };
+
                 // when used with custom port this results in non obvious behaviour
                 // port is remembered from first start command, i e
                 // start --port X
@@ -664,8 +712,14 @@ fn handle_pg(pg_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
                 // start <-- will also use port X even without explicit port argument
                 println!("Starting new postgres (v{pg_version}) {node_name} on timeline {timeline_id} ...");
 
-                let node =
-                    cplane.new_node(tenant_id, node_name, timeline_id, lsn, port, pg_version)?;
+                let node = cplane.new_node(
+                    tenant_id,
+                    node_name,
+                    timeline_id,
+                    port,
+                    pg_version,
+                    replication,
+                )?;
                 node.start(&auth_token)?;
             }
         }
@@ -918,6 +972,12 @@ fn cli() -> Command {
         .help("Specify Lsn on the timeline to start from. By default, end of the timeline would be used.")
         .required(false);
 
+    let hot_standby_arg = Arg::new("hot-standby")
+        .value_parser(value_parser!(bool))
+        .long("hot-standby")
+        .help("If set, the node will be a hot replica on the specified timeline")
+        .required(false);
+
     Command::new("Neon CLI")
         .arg_required_else_help(true)
         .version(GIT_VERSION)
@@ -1042,6 +1102,7 @@ fn cli() -> Command {
                             .long("config-only")
                             .required(false))
                     .arg(pg_version_arg.clone())
+                    .arg(hot_standby_arg.clone())
                 )
                 .subcommand(Command::new("start")
                     .about("Start a postgres compute node.\n This command actually creates new node from scratch, but preserves existing config files")
@@ -1052,6 +1113,7 @@ fn cli() -> Command {
                     .arg(lsn_arg)
                     .arg(port_arg)
                     .arg(pg_version_arg)
+                    .arg(hot_standby_arg)
                 )
                 .subcommand(
                     Command::new("stop")

control_plane/src/compute.rs

@@ -78,11 +78,12 @@ impl ComputeControlPlane {
         tenant_id: TenantId,
         name: &str,
         timeline_id: TimelineId,
-        lsn: Option<Lsn>,
         port: Option<u16>,
         pg_version: u32,
+        replication: Replication,
     ) -> Result<Arc<PostgresNode>> {
         let port = port.unwrap_or_else(|| self.get_port());
+
         let node = Arc::new(PostgresNode {
             name: name.to_owned(),
             address: SocketAddr::new("127.0.0.1".parse().unwrap(), port),
@@ -90,7 +91,7 @@ impl ComputeControlPlane {
             pageserver: Arc::clone(&self.pageserver),
             is_test: false,
             timeline_id,
-            lsn,
+            replication,
             tenant_id,
             uses_wal_proposer: false,
             pg_version,
@@ -108,6 +109,16 @@ impl ComputeControlPlane {
 
 ///////////////////////////////////////////////////////////////////////////////
 
+#[derive(Debug, Clone, Eq, PartialEq)]
+pub enum Replication {
+    // Regular read-write node
+    Primary,
+    // if recovery_target_lsn is provided
+    Static(Lsn),
+    // Hot standby running on a timleine
+    Replica,
+}
+
 #[derive(Debug)]
 pub struct PostgresNode {
     pub address: SocketAddr,
@@ -116,7 +127,7 @@ pub struct PostgresNode {
     pageserver: Arc<PageServerNode>,
     is_test: bool,
     pub timeline_id: TimelineId,
-    pub lsn: Option<Lsn>, // if it's a read-only node. None for primary
+    pub replication: Replication,
     pub tenant_id: TenantId,
     uses_wal_proposer: bool,
     pg_version: u32,
@@ -162,9 +173,17 @@ impl PostgresNode {
             fs::read_to_string(pg_version_path).unwrap_or_else(|_| DEFAULT_PG_VERSION.to_string());
         let pg_version = u32::from_str(&pg_version_str)?;
 
-        // parse recovery_target_lsn, if any
-        let recovery_target_lsn: Option<Lsn> =
-            conf.parse_field_optional("recovery_target_lsn", &context)?;
+        // parse recovery_target_lsn and primary_conninfo into Recovery Target, if any
+        let replication = if let Some(lsn_str) = conf.get("recovery_target_lsn") {
+            Replication::Static(Lsn::from_str(lsn_str)?)
+        } else if let Some(slot_name) = conf.get("primary_slot_name") {
+            let slot_name = slot_name.to_string();
+            let prefix = format!("repl_{}_{}_", timeline_id, name);
+            assert!(slot_name.starts_with(&prefix));
+            Replication::Replica
+        } else {
+            Replication::Primary
+        };
 
         // ok now
         Ok(PostgresNode {
@@ -174,7 +193,7 @@ impl PostgresNode {
             pageserver: Arc::clone(pageserver),
             is_test: false,
             timeline_id,
-            lsn: recovery_target_lsn,
+            replication,
             tenant_id,
             uses_wal_proposer,
             pg_version,
@@ -327,50 +346,83 @@ impl PostgresNode {
         }
         conf.append("neon.tenant_id", &self.tenant_id.to_string());
         conf.append("neon.timeline_id", &self.timeline_id.to_string());
-        if let Some(lsn) = self.lsn {
-            conf.append("recovery_target_lsn", &lsn.to_string());
-        }
 
         conf.append_line("");
-        // Configure backpressure
-        // - Replication write lag depends on how fast the walreceiver can process incoming WAL.
-        //   This lag determines latency of get_page_at_lsn. Speed of applying WAL is about 10MB/sec,
-        //   so to avoid expiration of 1 minute timeout, this lag should not be larger than 600MB.
-        //   Actually latency should be much smaller (better if < 1sec). But we assume that recently
-        //   updates pages are not requested from pageserver.
-        // - Replication flush lag depends on speed of persisting data by checkpointer (creation of
-        //   delta/image layers) and advancing disk_consistent_lsn. Safekeepers are able to
-        //   remove/archive WAL only beyond disk_consistent_lsn. Too large a lag can cause long
-        //   recovery time (in case of pageserver crash) and disk space overflow at safekeepers.
-        // - Replication apply lag depends on speed of uploading changes to S3 by uploader thread.
-        //   To be able to restore database in case of pageserver node crash, safekeeper should not
-        //   remove WAL beyond this point. Too large lag can cause space exhaustion in safekeepers
-        //   (if they are not able to upload WAL to S3).
-        conf.append("max_replication_write_lag", "15MB");
-        conf.append("max_replication_flush_lag", "10GB");
+        // Replication-related configurations, such as WAL sending
+        match &self.replication {
+            Replication::Primary => {
+                // Configure backpressure
+                // - Replication write lag depends on how fast the walreceiver can process incoming WAL.
+                //   This lag determines latency of get_page_at_lsn. Speed of applying WAL is about 10MB/sec,
+                //   so to avoid expiration of 1 minute timeout, this lag should not be larger than 600MB.
+                //   Actually latency should be much smaller (better if < 1sec). But we assume that recently
+                //   updates pages are not requested from pageserver.
+                // - Replication flush lag depends on speed of persisting data by checkpointer (creation of
+                //   delta/image layers) and advancing disk_consistent_lsn. Safekeepers are able to
+                //   remove/archive WAL only beyond disk_consistent_lsn. Too large a lag can cause long
+                //   recovery time (in case of pageserver crash) and disk space overflow at safekeepers.
+                // - Replication apply lag depends on speed of uploading changes to S3 by uploader thread.
+                //   To be able to restore database in case of pageserver node crash, safekeeper should not
+                //   remove WAL beyond this point. Too large lag can cause space exhaustion in safekeepers
+                //   (if they are not able to upload WAL to S3).
+                conf.append("max_replication_write_lag", "15MB");
+                conf.append("max_replication_flush_lag", "10GB");
 
-        if !self.env.safekeepers.is_empty() {
-            // Configure the node to connect to the safekeepers
-            conf.append("synchronous_standby_names", "walproposer");
+                if !self.env.safekeepers.is_empty() {
+                    // Configure the node to connect to the safekeepers
+                    conf.append("synchronous_standby_names", "walproposer");
 
-            let safekeepers = self
-                .env
-                .safekeepers
-                .iter()
-                .map(|sk| format!("localhost:{}", sk.pg_port))
-                .collect::<Vec<String>>()
-                .join(",");
-            conf.append("neon.safekeepers", &safekeepers);
-        } else {
-            // We only use setup without safekeepers for tests,
-            // and don't care about data durability on pageserver,
-            // so set more relaxed synchronous_commit.
-            conf.append("synchronous_commit", "remote_write");
+                    let safekeepers = self
+                        .env
+                        .safekeepers
+                        .iter()
+                        .map(|sk| format!("localhost:{}", sk.pg_port))
+                        .collect::<Vec<String>>()
+                        .join(",");
+                    conf.append("neon.safekeepers", &safekeepers);
+                } else {
+                    // We only use setup without safekeepers for tests,
+                    // and don't care about data durability on pageserver,
+                    // so set more relaxed synchronous_commit.
+                    conf.append("synchronous_commit", "remote_write");
 
-            // Configure the node to stream WAL directly to the pageserver
-            // This isn't really a supported configuration, but can be useful for
-            // testing.
-            conf.append("synchronous_standby_names", "pageserver");
+                    // Configure the node to stream WAL directly to the pageserver
+                    // This isn't really a supported configuration, but can be useful for
+                    // testing.
+                    conf.append("synchronous_standby_names", "pageserver");
+                }
+            }
+            Replication::Static(lsn) => {
+                conf.append("recovery_target_lsn", &lsn.to_string());
+            }
+            Replication::Replica => {
+                assert!(!self.env.safekeepers.is_empty());
+
+                // TODO: use future host field from safekeeper spec
+                // Pass the list of safekeepers to the replica so that it can connect to any of them,
+                // whichever is alailiable.
+                let sk_ports = self
+                    .env
+                    .safekeepers
+                    .iter()
+                    .map(|x| x.pg_port.to_string())
+                    .collect::<Vec<_>>()
+                    .join(",");
+                let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
+
+                let connstr = format!(
+                    "host={} port={} options='-c timeline_id={} tenant_id={}' application_name=replica replication=true",
+                    sk_hosts,
+                    sk_ports,
+                    &self.timeline_id.to_string(),
+                    &self.tenant_id.to_string(),
+                );
+
+                let slot_name = format!("repl_{}_", self.timeline_id);
+                conf.append("primary_conninfo", connstr.as_str());
+                conf.append("primary_slot_name", slot_name.as_str());
+                conf.append("hot_standby", "on");
+            }
         }
 
         let mut file = File::create(self.pgdata().join("postgresql.conf"))?;
@@ -383,21 +435,27 @@ impl PostgresNode {
     }
 
     fn load_basebackup(&self, auth_token: &Option<String>) -> Result<()> {
-        let backup_lsn = if let Some(lsn) = self.lsn {
-            Some(lsn)
-        } else if self.uses_wal_proposer {
-            // LSN 0 means that it is bootstrap and we need to download just
-            // latest data from the pageserver. That is a bit clumsy but whole bootstrap
-            // procedure evolves quite actively right now, so let's think about it again
-            // when things would be more stable (TODO).
-            let lsn = self.sync_safekeepers(auth_token, self.pg_version)?;
-            if lsn == Lsn(0) {
-                None
-            } else {
-                Some(lsn)
+        let backup_lsn = match &self.replication {
+            Replication::Primary => {
+                if self.uses_wal_proposer {
+                    // LSN 0 means that it is bootstrap and we need to download just
+                    // latest data from the pageserver. That is a bit clumsy but whole bootstrap
+                    // procedure evolves quite actively right now, so let's think about it again
+                    // when things would be more stable (TODO).
+                    let lsn = self.sync_safekeepers(auth_token, self.pg_version)?;
+                    if lsn == Lsn(0) {
+                        None
+                    } else {
+                        Some(lsn)
+                    }
+                } else {
+                    None
+                }
+            }
+            Replication::Static(lsn) => Some(*lsn),
+            Replication::Replica => {
+                None // Take the latest snapshot available to start with
             }
-        } else {
-            None
         };
 
         self.do_basebackup(backup_lsn)?;
@@ -487,7 +545,7 @@ impl PostgresNode {
         // 3. Load basebackup
         self.load_basebackup(auth_token)?;
 
-        if self.lsn.is_some() {
+        if self.replication != Replication::Primary {
             File::create(self.pgdata().join("standby.signal"))?;
         }
 

control_plane/src/postgresql_conf.rs

@@ -13,7 +13,7 @@ use std::io::BufRead;
 use std::str::FromStr;
 
 /// In-memory representation of a postgresql.conf file
-#[derive(Default)]
+#[derive(Default, Debug)]
 pub struct PostgresConf {
     lines: Vec<String>,
     hash: HashMap<String, String>,