feat: flownode to frontend load balance with guess

src/flow/src/batching_mode/frontend_client.rs

@@ -14,8 +14,9 @@
 
 //! Frontend client to run flow as batching task which is time-window-aware normal query triggered every tick set by user
 
-use std::sync::{Arc, Weak};
-use std::time::SystemTime;
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex, Weak};
+use std::time::{Duration, Instant, SystemTime};
 
 use api::v1::greptime_request::Request;
 use api::v1::CreateTableExpr;
@@ -26,20 +27,21 @@ use common_meta::cluster::{NodeInfo, NodeInfoKey, Role};
 use common_meta::peer::Peer;
 use common_meta::rpc::store::RangeRequest;
 use common_query::Output;
-use common_telemetry::warn;
+use common_telemetry::{debug, warn};
+use itertools::Itertools;
 use meta_client::client::MetaClient;
-use rand::rng;
-use rand::seq::SliceRandom;
 use servers::query_handler::grpc::GrpcQueryHandler;
 use session::context::{QueryContextBuilder, QueryContextRef};
 use snafu::{OptionExt, ResultExt};
 
+use crate::batching_mode::task::BatchingTask;
 use crate::batching_mode::{
     DEFAULT_BATCHING_ENGINE_QUERY_TIMEOUT, FRONTEND_ACTIVITY_TIMEOUT, GRPC_CONN_TIMEOUT,
     GRPC_MAX_RETRIES,
 };
 use crate::error::{ExternalSnafu, InvalidRequestSnafu, NoAvailableFrontendSnafu, UnexpectedSnafu};
-use crate::{Error, FlowAuthHeader};
+use crate::metrics::METRIC_FLOW_BATCHING_ENGINE_GUESS_FE_LOAD;
+use crate::{Error, FlowAuthHeader, FlowId};
 
 /// Just like [`GrpcQueryHandler`] but use BoxedError
 ///
@@ -74,6 +76,105 @@ impl<
 
 type HandlerMutable = Arc<std::sync::Mutex<Option<Weak<dyn GrpcQueryHandlerWithBoxedError>>>>;
 
+/// Statistics about running query on this frontend from flownode
+#[derive(Debug, Default, Clone)]
+struct FrontendStat {
+    /// The query for flow id has been running since this timestamp
+    since: HashMap<FlowId, Instant>,
+    /// The average query time for each flow id
+    /// This is used to calculate the average query time for each flow id
+    past_query_avg: HashMap<FlowId, (usize, Duration)>,
+}
+
+#[derive(Debug, Default, Clone)]
+pub struct FrontendStats {
+    /// The statistics for each flow id
+    stats: Arc<Mutex<HashMap<String, FrontendStat>>>,
+}
+
+impl FrontendStats {
+    pub fn observe(&self, frontend_addr: &str, flow_id: FlowId) -> FrontendStatsGuard {
+        let mut stats = self.stats.lock().expect("Failed to lock frontend stats");
+        let stat = stats.entry(frontend_addr.to_string()).or_default();
+        stat.since.insert(flow_id, Instant::now());
+
+        FrontendStatsGuard {
+            stats: self.stats.clone(),
+            frontend_addr: frontend_addr.to_string(),
+            cur: flow_id,
+        }
+    }
+
+    /// return frontend addrs sorted by load, from lightest to heaviest
+    /// The load is calculated as the total average query time for each flow id plus running query's total running time elapsed
+    pub fn sort_by_load(&self) -> Vec<String> {
+        let stats = self.stats.lock().expect("Failed to lock frontend stats");
+        let fe_load_factor = stats
+            .iter()
+            .map(|(node_addr, stat)| {
+                // total expected avg running time for all currently running queries
+                let total_expect_avg_run_time = stat
+                    .since
+                    .keys()
+                    .map(|f| {
+                        let (count, total_duration) =
+                            stat.past_query_avg.get(f).unwrap_or(&(0, Duration::ZERO));
+                        if *count == 0 {
+                            0.0
+                        } else {
+                            total_duration.as_secs_f64() / *count as f64
+                        }
+                    })
+                    .sum::<f64>();
+                let total_cur_running_time = stat
+                    .since
+                    .values()
+                    .map(|since| since.elapsed().as_secs_f64())
+                    .sum::<f64>();
+                (
+                    node_addr.to_string(),
+                    total_expect_avg_run_time + total_cur_running_time,
+                )
+            })
+            .sorted_by(|(_, load_a), (_, load_b)| {
+                load_a
+                    .partial_cmp(load_b)
+                    .unwrap_or(std::cmp::Ordering::Equal)
+            })
+            .collect::<Vec<_>>();
+        debug!("Frontend load factor: {:?}", fe_load_factor);
+        for (node_addr, load) in &fe_load_factor {
+            METRIC_FLOW_BATCHING_ENGINE_GUESS_FE_LOAD
+                .with_label_values(&[&node_addr.to_string()])
+                .observe(*load);
+        }
+        fe_load_factor
+            .into_iter()
+            .map(|(addr, _)| addr)
+            .collect::<Vec<_>>()
+    }
+}
+
+pub struct FrontendStatsGuard {
+    stats: Arc<Mutex<HashMap<String, FrontendStat>>>,
+    frontend_addr: String,
+    cur: FlowId,
+}
+
+impl Drop for FrontendStatsGuard {
+    fn drop(&mut self) {
+        let mut stats = self.stats.lock().expect("Failed to lock frontend stats");
+        if let Some(stat) = stats.get_mut(&self.frontend_addr) {
+            if let Some(since) = stat.since.remove(&self.cur) {
+                let elapsed = since.elapsed();
+                let (count, total_duration) = stat.past_query_avg.entry(self.cur).or_default();
+                *count += 1;
+                *total_duration += elapsed;
+            }
+        }
+    }
+}
+
 /// A simple frontend client able to execute sql using grpc protocol
 ///
 /// This is for computation-heavy query which need to offload computation to frontend, lifting the load from flownode
@@ -83,6 +184,7 @@ pub enum FrontendClient {
         meta_client: Arc<MetaClient>,
         chnl_mgr: ChannelManager,
         auth: Option<FlowAuthHeader>,
+        fe_stats: FrontendStats,
     },
     Standalone {
         /// for the sake of simplicity still use grpc even in standalone mode
@@ -114,6 +216,7 @@ impl FrontendClient {
                 ChannelManager::with_config(cfg)
             },
             auth,
+            fe_stats: Default::default(),
         }
     }
 
@@ -192,6 +295,7 @@ impl FrontendClient {
             meta_client: _,
             chnl_mgr,
             auth,
+            fe_stats,
         } = self
         else {
             return UnexpectedSnafu {
@@ -208,8 +312,21 @@ impl FrontendClient {
                 .duration_since(SystemTime::UNIX_EPOCH)
                 .unwrap()
                 .as_millis() as i64;
-            // shuffle the frontends to avoid always pick the same one
-            frontends.shuffle(&mut rng());
+            let node_addrs_by_load = fe_stats.sort_by_load();
+            // index+1 to load order asc, so that the lightest node has load 1 and non-existent node has load 0
+            let addr2load = node_addrs_by_load
+                .iter()
+                .enumerate()
+                .map(|(i, id)| (id.clone(), i + 1))
+                .collect::<HashMap<_, _>>();
+            // sort frontends by load, from lightest to heaviest
+            frontends.sort_by(|(_, a), (_, b)| {
+                // if not even in stats, treat as 0 load since never been queried
+                let load_a = addr2load.get(&a.peer.addr).unwrap_or(&0);
+                let load_b = addr2load.get(&b.peer.addr).unwrap_or(&0);
+                load_a.cmp(load_b)
+            });
+            debug!("Frontend nodes sorted by load: {:?}", frontends);
 
             // found node with maximum last_activity_ts
             for (_, node_info) in frontends
@@ -257,6 +374,7 @@ impl FrontendClient {
         create: CreateTableExpr,
         catalog: &str,
         schema: &str,
+        task: Option<&BatchingTask>,
     ) -> Result<u32, Error> {
         self.handle(
             Request::Ddl(api::v1::DdlRequest {
@@ -265,6 +383,7 @@ impl FrontendClient {
             catalog,
             schema,
             &mut None,
+            task,
         )
         .await
     }
@@ -276,15 +395,19 @@ impl FrontendClient {
         catalog: &str,
         schema: &str,
         peer_desc: &mut Option<PeerDesc>,
+        task: Option<&BatchingTask>,
     ) -> Result<u32, Error> {
         match self {
-            FrontendClient::Distributed { .. } => {
+            FrontendClient::Distributed { fe_stats, .. } => {
                 let db = self.get_random_active_frontend(catalog, schema).await?;
 
                 *peer_desc = Some(PeerDesc::Dist {
                     peer: db.peer.clone(),
                 });
 
+                let flow_id = task.map(|t| t.config.flow_id).unwrap_or_default();
+                let _guard = fe_stats.observe(&db.peer.addr, flow_id);
+
                 db.database
                     .handle_with_retry(req.clone(), GRPC_MAX_RETRIES)
                     .await

src/flow/src/batching_mode/task.rs

@@ -280,7 +280,7 @@ impl BatchingTask {
         let catalog = &self.config.sink_table_name[0];
         let schema = &self.config.sink_table_name[1];
         frontend_client
-            .create(expr.clone(), catalog, schema)
+            .create(expr.clone(), catalog, schema, Some(self))
             .await?;
         Ok(())
     }
@@ -361,7 +361,7 @@ impl BatchingTask {
             };
 
             frontend_client
-                .handle(req, catalog, schema, &mut peer_desc)
+                .handle(req, catalog, schema, &mut peer_desc, Some(self))
                 .await
         };
 

src/flow/src/metrics.rs

@@ -58,6 +58,14 @@ lazy_static! {
             vec![60., 4. * 60., 16. * 60., 64. * 60., 256. * 60.]
         )
         .unwrap();
+    pub static ref METRIC_FLOW_BATCHING_ENGINE_GUESS_FE_LOAD: HistogramVec =
+        register_histogram_vec!(
+            "greptime_flow_batching_engine_guess_fe_load",
+            "flow batching engine guessed frontend load",
+            &["fe_addr"],
+            vec![60., 4. * 60., 16. * 60., 64. * 60., 256. * 60.]
+        )
+        .unwrap();
     pub static ref METRIC_FLOW_RUN_INTERVAL_MS: IntGauge =
         register_int_gauge!("greptime_flow_run_interval_ms", "flow run interval in ms").unwrap();
     pub static ref METRIC_FLOW_ROWS: IntCounterVec = register_int_counter_vec!(