feat: execute auto-scaling in batches (#15420)

src/common/src/config.rs

@@ -237,6 +237,18 @@ pub struct MetaConfig {
     #[serde(default)]
     pub disable_automatic_parallelism_control: bool,
 
+    /// The number of streaming jobs per scaling operation.
+    #[serde(default = "default::meta::parallelism_control_batch_size")]
+    pub parallelism_control_batch_size: usize,
+
+    /// The period of parallelism control trigger.
+    #[serde(default = "default::meta::parallelism_control_trigger_period_sec")]
+    pub parallelism_control_trigger_period_sec: u64,
+
+    /// The first delay of parallelism control.
+    #[serde(default = "default::meta::parallelism_control_trigger_first_delay_sec")]
+    pub parallelism_control_trigger_first_delay_sec: u64,
+
     #[serde(default = "default::meta::meta_leader_lease_secs")]
     pub meta_leader_lease_secs: u64,
 
@@ -1121,6 +1133,18 @@ pub mod default {
         pub fn event_log_channel_max_size() -> u32 {
             10
         }
+
+        pub fn parallelism_control_batch_size() -> usize {
+            10
+        }
+
+        pub fn parallelism_control_trigger_period_sec() -> u64 {
+            10
+        }
+
+        pub fn parallelism_control_trigger_first_delay_sec() -> u64 {
+            30
+        }
     }
 
     pub mod server {

src/config/docs.md

@@ -41,6 +41,9 @@ This page is automatically generated by `./risedev generate-example-config`
 | min_table_split_write_throughput |  If the size of one table is smaller than `min_table_split_write_throughput`, we would not  split it to an single group. | 4194304 |
 | move_table_size_limit |  | 10737418240 |
 | node_num_monitor_interval_sec |  | 10 |
+| parallelism_control_batch_size |  The number of streaming jobs per scaling operation. | 10 |
+| parallelism_control_trigger_first_delay_sec |  The first delay of parallelism control. | 30 |
+| parallelism_control_trigger_period_sec |  The period of parallelism control trigger. | 10 |
 | partition_vnode_count |  | 16 |
 | periodic_compaction_interval_sec |  Schedule compaction for all compaction groups with this interval. | 60 |
 | periodic_space_reclaim_compaction_interval_sec |  Schedule space_reclaim compaction for all compaction groups with this interval. | 3600 |

src/config/example.toml

@@ -26,6 +26,9 @@ min_delta_log_num_for_hummock_version_checkpoint = 10
 max_heartbeat_interval_secs = 300
 disable_recovery = false
 disable_automatic_parallelism_control = false
+parallelism_control_batch_size = 10
+parallelism_control_trigger_period_sec = 10
+parallelism_control_trigger_first_delay_sec = 30
 meta_leader_lease_secs = 30
 default_parallelism = "Full"
 enable_compaction_deterministic = false

src/meta/node/src/lib.rs

@@ -285,6 +285,13 @@ pub fn start(opts: MetaNodeOpts) -> Pin<Box<dyn Future<Output = ()> + Send>> {
                 disable_automatic_parallelism_control: config
                     .meta
                     .disable_automatic_parallelism_control,
+                parallelism_control_batch_size: config.meta.parallelism_control_batch_size,
+                parallelism_control_trigger_period_sec: config
+                    .meta
+                    .parallelism_control_trigger_period_sec,
+                parallelism_control_trigger_first_delay_sec: config
+                    .meta
+                    .parallelism_control_trigger_first_delay_sec,
                 in_flight_barrier_nums,
                 max_idle_ms,
                 compaction_deterministic_test: config.meta.enable_compaction_deterministic,

src/meta/src/barrier/command.rs

@@ -615,7 +615,7 @@ impl CommandContext {
                         actor_splits,
                         actor_new_dispatchers,
                     });
-                    tracing::debug!("update mutation: {mutation:#?}");
+                    tracing::debug!("update mutation: {mutation:?}");
                     Some(mutation)
                 }
             };

src/meta/src/barrier/recovery.rs

@@ -721,7 +721,7 @@ impl GlobalBarrierManagerContext {
             return Err(e);
         }
 
-        debug!("scaling-in actors succeed.");
+        debug!("scaling actors succeed.");
         Ok(())
     }
 
@@ -855,7 +855,7 @@ impl GlobalBarrierManagerContext {
             );
         }
 
-        debug!("scaling-in actors succeed.");
+        debug!("scaling actors succeed.");
         Ok(())
     }
 

src/meta/src/manager/env.rs

@@ -93,6 +93,12 @@ pub struct MetaOpts {
     pub enable_recovery: bool,
     /// Whether to disable the auto-scaling feature.
     pub disable_automatic_parallelism_control: bool,
+    /// The number of streaming jobs per scaling operation.
+    pub parallelism_control_batch_size: usize,
+    /// The period of parallelism control trigger.
+    pub parallelism_control_trigger_period_sec: u64,
+    /// The first delay of parallelism control.
+    pub parallelism_control_trigger_first_delay_sec: u64,
     /// The maximum number of barriers in-flight in the compute nodes.
     pub in_flight_barrier_nums: usize,
     /// After specified seconds of idle (no mview or flush), the process will be exited.
@@ -221,6 +227,9 @@ impl MetaOpts {
         Self {
             enable_recovery,
             disable_automatic_parallelism_control: false,
+            parallelism_control_batch_size: 1,
+            parallelism_control_trigger_period_sec: 10,
+            parallelism_control_trigger_first_delay_sec: 30,
             in_flight_barrier_nums: 40,
             max_idle_ms: 0,
             compaction_deterministic_test: false,

src/meta/src/stream/scale.rs

@@ -31,7 +31,9 @@ use risingwave_common::catalog::TableId;
 use risingwave_common::hash::{ActorMapping, ParallelUnitId, VirtualNode};
 use risingwave_common::util::iter_util::ZipEqDebug;
 use risingwave_meta_model_v2::StreamingParallelism;
-use risingwave_pb::common::{ActorInfo, Buffer, ParallelUnit, ParallelUnitMapping, WorkerNode};
+use risingwave_pb::common::{
+    ActorInfo, Buffer, ParallelUnit, ParallelUnitMapping, WorkerNode, WorkerType,
+};
 use risingwave_pb::meta::get_reschedule_plan_request::{Policy, StableResizePolicy};
 use risingwave_pb::meta::subscribe_response::{Info, Operation};
 use risingwave_pb::meta::table_fragments::actor_status::ActorState;
@@ -48,7 +50,7 @@ use thiserror_ext::AsReport;
 use tokio::sync::oneshot::Receiver;
 use tokio::sync::{oneshot, RwLock, RwLockReadGuard, RwLockWriteGuard};
 use tokio::task::JoinHandle;
-use tokio::time::MissedTickBehavior;
+use tokio::time::{Instant, MissedTickBehavior};
 
 use crate::barrier::{Command, Reschedule, StreamRpcManager};
 use crate::manager::{IdCategory, LocalNotification, MetaSrvEnv, MetadataManager, WorkerId};
@@ -2676,12 +2678,14 @@ impl GlobalStreamManager {
         Ok(())
     }
 
-    async fn trigger_parallelism_control(&self) -> MetaResult<()> {
+    async fn trigger_parallelism_control(&self) -> MetaResult<bool> {
+        tracing::info!("trigger parallelism control");
+
         let _reschedule_job_lock = self.reschedule_lock_write_guard().await;
 
-        match &self.metadata_manager {
+        let (schedulable_worker_ids, table_parallelisms) = match &self.metadata_manager {
             MetadataManager::V1(mgr) => {
-                let table_parallelisms = {
+                let table_parallelisms: HashMap<u32, TableParallelism> = {
                     let guard = mgr.fragment_manager.get_fragment_read_guard().await;
 
                     guard
@@ -2697,7 +2701,7 @@ impl GlobalStreamManager {
                     .list_active_streaming_compute_nodes()
                     .await;
 
-                let schedulable_worker_ids = workers
+                let schedulable_worker_ids: BTreeSet<_> = workers
                     .iter()
                     .filter(|worker| {
                         !worker
@@ -2709,26 +2713,7 @@ impl GlobalStreamManager {
                     .map(|worker| worker.id)
                     .collect();
 
-                let reschedules = self
-                    .scale_controller
-                    .generate_table_resize_plan(TableResizePolicy {
-                        worker_ids: schedulable_worker_ids,
-                        table_parallelisms,
-                    })
-                    .await?;
-
-                if reschedules.is_empty() {
-                    return Ok(());
-                }
-
-                self.reschedule_actors(
-                    reschedules,
-                    RescheduleOptions {
-                        resolve_no_shuffle_upstream: true,
-                    },
-                    None,
-                )
-                .await?;
+                (schedulable_worker_ids, table_parallelisms)
             }
             MetadataManager::V2(mgr) => {
                 let table_parallelisms: HashMap<_, _> = {
@@ -2768,33 +2753,90 @@ impl GlobalStreamManager {
                     .map(|worker| worker.id)
                     .collect();
 
-                let reschedules = self
-                    .scale_controller
-                    .generate_table_resize_plan(TableResizePolicy {
-                        worker_ids: schedulable_worker_ids,
-                        table_parallelisms: table_parallelisms.clone(),
-                    })
-                    .await?;
+                (schedulable_worker_ids, table_parallelisms)
+            }
+        };
 
-                if reschedules.is_empty() {
-                    return Ok(());
-                }
+        if table_parallelisms.is_empty() {
+            tracing::info!("no streaming jobs for scaling, maybe an empty cluster");
+            return Ok(false);
+        }
 
-                self.reschedule_actors(
-                    reschedules,
-                    RescheduleOptions {
-                        resolve_no_shuffle_upstream: true,
-                    },
-                    None,
-                )
+        let batch_size = match self.env.opts.parallelism_control_batch_size {
+            0 => table_parallelisms.len(),
+            n => n,
+        };
+
+        tracing::info!(
+            "total {} streaming jobs, batch size {}, schedulable worker ids: {:?}",
+            table_parallelisms.len(),
+            batch_size,
+            schedulable_worker_ids
+        );
+
+        let batches: Vec<_> = table_parallelisms
+            .into_iter()
+            .chunks(batch_size)
+            .into_iter()
+            .map(|chunk| chunk.collect_vec())
+            .collect();
+
+        let mut reschedules = None;
+
+        for batch in batches {
+            let parallelisms: HashMap<_, _> = batch.into_iter().collect();
+
+            let plan = self
+                .scale_controller
+                .generate_table_resize_plan(TableResizePolicy {
+                    worker_ids: schedulable_worker_ids.clone(),
+                    table_parallelisms: parallelisms.clone(),
+                })
                 .await?;
+
+            if !plan.is_empty() {
+                tracing::info!(
+                    "reschedule plan generated for streaming jobs {:?}",
+                    parallelisms
+                );
+                reschedules = Some(plan);
+                break;
             }
         }
 
-        Ok(())
+        let Some(reschedules) = reschedules else {
+            tracing::info!("no reschedule plan generated");
+            return Ok(false);
+        };
+
+        self.reschedule_actors(
+            reschedules,
+            RescheduleOptions {
+                resolve_no_shuffle_upstream: false,
+            },
+            None,
+        )
+        .await?;
+
+        Ok(true)
     }
 
     async fn run(&self, mut shutdown_rx: Receiver<()>) {
+        tracing::info!("starting automatic parallelism control monitor");
+
+        let check_period =
+            Duration::from_secs(self.env.opts.parallelism_control_trigger_period_sec);
+
+        let mut ticker = tokio::time::interval_at(
+            Instant::now()
+                + Duration::from_secs(self.env.opts.parallelism_control_trigger_first_delay_sec),
+            check_period,
+        );
+        ticker.set_missed_tick_behavior(MissedTickBehavior::Skip);
+
+        // waiting for first tick
+        ticker.tick().await;
+
         let (local_notification_tx, mut local_notification_rx) =
             tokio::sync::mpsc::unbounded_channel();
 
@@ -2803,11 +2845,6 @@ impl GlobalStreamManager {
             .insert_local_sender(local_notification_tx)
             .await;
 
-        let check_period = Duration::from_secs(10);
-        let mut ticker = tokio::time::interval(check_period);
-        ticker.set_missed_tick_behavior(MissedTickBehavior::Skip);
-        ticker.reset();
-
         let worker_nodes = self
             .metadata_manager
             .list_active_streaming_compute_nodes()
@@ -2819,7 +2856,7 @@ impl GlobalStreamManager {
             .map(|worker| (worker.id, worker))
             .collect();
 
-        let mut changed = true;
+        let mut should_trigger = false;
 
         loop {
             tokio::select! {
@@ -2830,18 +2867,18 @@ impl GlobalStreamManager {
                     break;
                 }
 
-                _ = ticker.tick(), if changed => {
+                _ = ticker.tick(), if should_trigger => {
                     let include_workers = worker_cache.keys().copied().collect_vec();
 
                     if include_workers.is_empty() {
                         tracing::debug!("no available worker nodes");
-                        changed = false;
+                        should_trigger = false;
                         continue;
                     }
 
                     match self.trigger_parallelism_control().await {
-                        Ok(_) => {
-                            changed = false;
+                        Ok(cont) => {
+                            should_trigger = cont;
                         }
                         Err(e) => {
                             tracing::warn!(error = %e.as_report(), "Failed to trigger scale out, waiting for next tick to retry after {}s", ticker.period().as_secs());
@@ -2855,13 +2892,26 @@ impl GlobalStreamManager {
 
                     match notification {
                         LocalNotification::WorkerNodeActivated(worker) => {
+                            match (worker.get_type(), worker.property.as_ref()) {
+                                (Ok(WorkerType::ComputeNode), Some(prop)) if prop.is_streaming => {
+                                    tracing::info!("worker {} activated notification received", worker.id);
+                                }
+                                _ => continue
+                            }
+
                             let prev_worker = worker_cache.insert(worker.id, worker.clone());
 
-                            if let Some(prev_worker) = prev_worker && prev_worker.parallel_units != worker.parallel_units {
-                                tracing::info!(worker = worker.id, "worker parallelism changed");
+                            match prev_worker {
+                                Some(prev_worker) if prev_worker.parallel_units != worker.parallel_units  => {
+                                    tracing::info!(worker = worker.id, "worker parallelism changed");
+                                    should_trigger = true;
+                                }
+                                None => {
+                                    tracing::info!(worker = worker.id, "new worker joined");
+                                    should_trigger = true;
+                                }
+                                _ => {}
                             }
-
-                            changed = true;
                         }
 
                         // Since our logic for handling passive scale-in is within the barrier manager,

src/tests/simulation/src/cluster.rs

@@ -189,6 +189,9 @@ impl Configuration {
                 r#"[meta]
 max_heartbeat_interval_secs = {max_heartbeat_interval_secs}
 disable_automatic_parallelism_control = {disable_automatic_parallelism_control}
+parallelism_control_trigger_first_delay_sec = 0
+parallelism_control_batch_size = 0
+parallelism_control_trigger_period_sec = 10
 
 [system]
 barrier_interval_ms = 250