risingwave_stream/executor/
actor.rs

1// Copyright 2025 RisingWave Labs
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15use std::collections::{HashMap, HashSet};
16use std::sync::atomic::{AtomicUsize, Ordering};
17use std::sync::{Arc, LazyLock};
18
19use anyhow::anyhow;
20use await_tree::InstrumentAwait;
21use futures::FutureExt;
22use futures::future::join_all;
23use hytra::TrAdder;
24use risingwave_common::bitmap::Bitmap;
25use risingwave_common::config::StreamingConfig;
26use risingwave_common::hash::VirtualNode;
27use risingwave_common::log::LogSuppressor;
28use risingwave_common::metrics::{GLOBAL_ERROR_METRICS, IntGaugeExt};
29use risingwave_common::util::epoch::EpochPair;
30use risingwave_expr::ExprError;
31use risingwave_expr::expr_context::{FRAGMENT_ID, VNODE_COUNT, expr_context_scope};
32use risingwave_pb::id::SubscriberId;
33use risingwave_pb::plan_common::ExprContext;
34use risingwave_pb::stream_service::inject_barrier_request::BuildActorInfo;
35use risingwave_pb::stream_service::inject_barrier_request::build_actor_info::UpstreamActors;
36use risingwave_rpc_client::MetaClient;
37use thiserror_ext::AsReport;
38use tokio_stream::StreamExt;
39use tracing::Instrument;
40
41use super::StreamConsumer;
42use super::monitor::StreamingMetrics;
43use super::subtask::SubtaskHandle;
44use crate::CONFIG;
45use crate::error::StreamResult;
46use crate::task::{ActorId, FragmentId, LocalBarrierManager, StreamEnvironment};
47
48/// Shared by all operators of an actor.
49pub struct ActorContext {
50    pub id: ActorId,
51    pub fragment_id: FragmentId,
52    pub vnode_count: usize,
53    pub mview_definition: String,
54
55    // TODO(eric): these seem to be useless now?
56    last_mem_val: Arc<AtomicUsize>,
57    cur_mem_val: Arc<AtomicUsize>,
58    total_mem_val: Arc<TrAdder<i64>>,
59
60    pub streaming_metrics: Arc<StreamingMetrics>,
61
62    /// This is the number of dispatchers when the actor is created. It will not be updated during runtime when new downstreams are added.
63    pub initial_dispatch_num: usize,
64    // mv_table_id to subscription id
65    pub initial_subscriber_ids: HashSet<SubscriberId>,
66    pub initial_upstream_actors: HashMap<FragmentId, UpstreamActors>,
67
68    // Meta client. currently used for auto schema change. `None` for test only
69    pub meta_client: Option<MetaClient>,
70
71    /// The local streaming configuration for this specific actor.
72    ///
73    /// Compared to `stream_env.global_config`, this config can have some entries overridden by the user.
74    pub config: Arc<StreamingConfig>,
75
76    pub stream_env: StreamEnvironment,
77}
78
79pub type ActorContextRef = Arc<ActorContext>;
80
81impl ActorContext {
82    pub fn for_test(id: impl Into<ActorId>) -> ActorContextRef {
83        Arc::new(Self {
84            id: id.into(),
85            fragment_id: 0.into(),
86            vnode_count: VirtualNode::COUNT_FOR_TEST,
87            mview_definition: "".to_owned(),
88            cur_mem_val: Arc::new(0.into()),
89            last_mem_val: Arc::new(0.into()),
90            total_mem_val: Arc::new(TrAdder::new()),
91            streaming_metrics: Arc::new(StreamingMetrics::unused()),
92            // Set 1 for test to enable sanity check on table
93            initial_dispatch_num: 1,
94            initial_subscriber_ids: Default::default(),
95            initial_upstream_actors: Default::default(),
96            meta_client: None,
97            config: Arc::new(StreamingConfig::default()),
98            stream_env: StreamEnvironment::for_test(),
99        })
100    }
101
102    #[allow(clippy::too_many_arguments)]
103    pub fn create(
104        stream_actor: &BuildActorInfo,
105        fragment_id: FragmentId,
106        total_mem_val: Arc<TrAdder<i64>>,
107        streaming_metrics: Arc<StreamingMetrics>,
108        meta_client: Option<MetaClient>,
109        config: Arc<StreamingConfig>,
110        stream_env: StreamEnvironment,
111    ) -> ActorContextRef {
112        Arc::new(Self {
113            id: stream_actor.actor_id,
114            fragment_id,
115            mview_definition: stream_actor.mview_definition.clone(),
116            vnode_count: (stream_actor.vnode_bitmap.as_ref())
117                // An unset `vnode_bitmap` means the actor is a singleton,
118                // where only `SINGLETON_VNODE` is set.
119                .map_or(1, |b| Bitmap::from(b).len()),
120            cur_mem_val: Arc::new(0.into()),
121            last_mem_val: Arc::new(0.into()),
122            total_mem_val,
123            streaming_metrics,
124            initial_dispatch_num: stream_actor.dispatchers.len(),
125            initial_subscriber_ids: stream_actor
126                .initial_subscriber_ids
127                .iter()
128                .copied()
129                .collect(),
130            initial_upstream_actors: stream_actor.fragment_upstreams.clone(),
131            meta_client,
132            config,
133            stream_env,
134        })
135    }
136
137    pub fn on_compute_error(&self, err: ExprError, identity: &str) {
138        static LOG_SUPPRESSOR: LazyLock<LogSuppressor> = LazyLock::new(LogSuppressor::default);
139        if let Ok(suppressed_count) = LOG_SUPPRESSOR.check() {
140            tracing::error!(target: "stream_expr_error", identity, error = %err.as_report(), suppressed_count, "failed to evaluate expression");
141        }
142
143        let executor_name = identity.split(' ').next().unwrap_or("name_not_found");
144        GLOBAL_ERROR_METRICS.user_compute_error.report([
145            "ExprError".to_owned(),
146            executor_name.to_owned(),
147            self.fragment_id.to_string(),
148        ]);
149    }
150
151    pub fn store_mem_usage(&self, val: usize) {
152        // Record the last mem val.
153        // Calculate the difference between old val and new value, and apply the diff to total
154        // memory usage value.
155        let old_value = self.cur_mem_val.load(Ordering::Relaxed);
156        self.last_mem_val.store(old_value, Ordering::Relaxed);
157        let diff = val as i64 - old_value as i64;
158
159        self.total_mem_val.inc(diff);
160
161        self.cur_mem_val.store(val, Ordering::Relaxed);
162    }
163
164    pub fn mem_usage(&self) -> usize {
165        self.cur_mem_val.load(Ordering::Relaxed)
166    }
167}
168
169/// `Actor` is the basic execution unit in the streaming framework.
170pub struct Actor<C> {
171    /// The [`StreamConsumer`] of the actor.
172    consumer: C,
173    /// The subtasks to execute concurrently.
174    subtasks: Vec<SubtaskHandle>,
175
176    pub actor_context: ActorContextRef,
177    expr_context: ExprContext,
178    barrier_manager: LocalBarrierManager,
179}
180
181impl<C> Actor<C>
182where
183    C: StreamConsumer,
184{
185    pub fn new(
186        consumer: C,
187        subtasks: Vec<SubtaskHandle>,
188        _metrics: Arc<StreamingMetrics>,
189        actor_context: ActorContextRef,
190        expr_context: ExprContext,
191        barrier_manager: LocalBarrierManager,
192    ) -> Self {
193        Self {
194            consumer,
195            subtasks,
196            actor_context,
197            expr_context,
198            barrier_manager,
199        }
200    }
201
202    #[inline(always)]
203    pub async fn run(mut self) -> StreamResult<()> {
204        let expr_context = self.expr_context.clone();
205        let fragment_id = self.actor_context.fragment_id;
206        let vnode_count = self.actor_context.vnode_count;
207        let config = self.actor_context.config.clone();
208
209        let run = async move {
210            tokio::join!(
211                // Drive the subtasks concurrently.
212                join_all(std::mem::take(&mut self.subtasks)),
213                self.run_consumer(),
214            )
215            .1
216        }
217        .boxed();
218
219        // Attach contexts to the future.
220        let run = expr_context_scope(expr_context, run);
221        let run = FRAGMENT_ID::scope(fragment_id, run);
222        let run = VNODE_COUNT::scope(vnode_count, run);
223        let run = CONFIG.scope(config, run);
224
225        run.await
226    }
227
228    async fn run_consumer(self) -> StreamResult<()> {
229        fail::fail_point!("start_actors_err", |_| Err(anyhow::anyhow!(
230            "intentional start_actors_err"
231        )
232        .into()));
233
234        let id = self.actor_context.id;
235        let span_name = format!("Actor {id}");
236
237        let new_span = |epoch: Option<EpochPair>| {
238            tracing::info_span!(
239                parent: None,
240                "actor",
241                "otel.name" = span_name,
242                actor_id = %id,
243                prev_epoch = epoch.map(|e| e.prev),
244                curr_epoch = epoch.map(|e| e.curr),
245            )
246        };
247        let mut span = new_span(None);
248
249        let actor_count = self
250            .actor_context
251            .streaming_metrics
252            .actor_count
253            .with_guarded_label_values(&[&self.actor_context.fragment_id.to_string()]);
254        let _actor_count_guard = actor_count.inc_guard();
255
256        let current_epoch = self
257            .actor_context
258            .streaming_metrics
259            .actor_current_epoch
260            .with_guarded_label_values(&[
261                &self.actor_context.id.to_string(),
262                &self.actor_context.fragment_id.to_string(),
263            ]);
264
265        let mut last_epoch: Option<EpochPair> = None;
266        let mut stream = Box::pin(Box::new(self.consumer).execute());
267
268        // Drive the streaming task with an infinite loop
269        let result = loop {
270            let barrier = match stream
271                .try_next()
272                .instrument(span.clone())
273                .instrument_await(
274                    last_epoch.map_or(await_tree::span!("Epoch <initial>"), |e| {
275                        await_tree::span!("Epoch {}", e.curr)
276                    }),
277                )
278                .await
279            {
280                Ok(Some(barrier)) => barrier,
281                Ok(None) => break Err(anyhow!("actor exited unexpectedly").into()),
282                Err(err) => break Err(err),
283            };
284
285            fail::fail_point!("collect_actors_err", id == 10, |_| Err(anyhow::anyhow!(
286                "intentional collect_actors_err"
287            )
288            .into()));
289
290            // Then stop this actor if asked
291            if barrier.is_stop(id) {
292                debug!(actor_id = %id, epoch = ?barrier.epoch, "stop at barrier");
293                break Ok(barrier);
294            }
295
296            current_epoch.set(barrier.epoch.curr as i64);
297
298            // Collect barriers to local barrier manager
299            self.barrier_manager.collect(id, &barrier);
300
301            // Tracing related work
302            last_epoch = Some(barrier.epoch);
303            span = barrier.tracing_context().attach(new_span(last_epoch));
304        };
305
306        spawn_blocking_drop_stream(stream).await;
307
308        let result = result.map(|stop_barrier| {
309            // Collect the stop barrier after the stream has been dropped to ensure that all resources
310            self.barrier_manager.collect(id, &stop_barrier);
311        });
312
313        tracing::debug!(actor_id = %id, ok = result.is_ok(), "actor exit");
314        result
315    }
316}
317
318/// Drop the stream in a blocking task to avoid interfering with other actors.
319///
320/// Logically the actor is dropped after we send the barrier with `Drop` mutation to the
321/// downstream, thus making the `drop`'s progress asynchronous. However, there might be a
322/// considerable amount of data in the executors' in-memory cache, dropping these structures might
323/// be a CPU-intensive task. This may lead to the runtime being unable to schedule other actors if
324/// the `drop` is called on the current thread.
325pub async fn spawn_blocking_drop_stream<T: Send + 'static>(stream: T) {
326    let _ = tokio::task::spawn_blocking(move || drop(stream))
327        .instrument_await("drop_stream")
328        .await;
329}
risingwave_stream/executor/actor.rs

risingwave_stream/executor/
actor.rs
