flow/
adapter.rs

1// Copyright 2023 Greptime Team
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
15//! for getting data from source and sending results to sink
16//! and communicating with other parts of the database
17#![warn(unused_imports)]
18
19use std::collections::BTreeMap;
20use std::sync::Arc;
21use std::time::{Duration, Instant, SystemTime};
22
23use api::v1::{RowDeleteRequest, RowDeleteRequests, RowInsertRequest, RowInsertRequests};
24use common_base::memory_limit::MemoryLimit;
25use common_config::Configurable;
26use common_error::ext::BoxedError;
27use common_meta::key::TableMetadataManagerRef;
28use common_options::memory::MemoryOptions;
29use common_runtime::JoinHandle;
30use common_stat::get_total_cpu_cores;
31use common_telemetry::logging::{LoggingOptions, TracingOptions};
32use common_telemetry::{debug, info, trace};
33use datatypes::schema::ColumnSchema;
34use datatypes::value::Value;
35use greptime_proto::v1;
36use itertools::{EitherOrBoth, Itertools};
37use meta_client::MetaClientOptions;
38use query::QueryEngine;
39use query::options::QueryOptions;
40use serde::{Deserialize, Serialize};
41use servers::grpc::GrpcOptions;
42use servers::http::HttpOptions;
43use session::context::QueryContext;
44use snafu::{OptionExt, ResultExt, ensure};
45use store_api::storage::{ConcreteDataType, RegionId};
46use table::metadata::TableId;
47use tokio::sync::broadcast::error::TryRecvError;
48use tokio::sync::{Mutex, RwLock, broadcast, watch};
49
50pub(crate) use crate::adapter::node_context::FlownodeContext;
51use crate::adapter::refill::RefillTask;
52use crate::adapter::table_source::ManagedTableSource;
53use crate::adapter::util::relation_desc_to_column_schemas_with_fallback;
54pub(crate) use crate::adapter::worker::{Worker, WorkerHandle, create_worker};
55use crate::batching_mode::BatchingModeOptions;
56use crate::compute::ErrCollector;
57use crate::df_optimizer::sql_to_flow_plan;
58use crate::error::{EvalSnafu, ExternalSnafu, InternalSnafu, InvalidQuerySnafu, UnexpectedSnafu};
59use crate::expr::Batch;
60use crate::metrics::{METRIC_FLOW_INSERT_ELAPSED, METRIC_FLOW_ROWS, METRIC_FLOW_RUN_INTERVAL_MS};
61use crate::repr::{self, BATCH_SIZE, DiffRow, RelationDesc, Row};
62use crate::{CreateFlowArgs, FlowId, TableName};
63
64pub(crate) mod flownode_impl;
65mod parse_expr;
66pub(crate) mod refill;
67mod stat;
68#[cfg(test)]
69mod tests;
70pub(crate) mod util;
71mod worker;
72
73pub(crate) mod node_context;
74pub(crate) mod table_source;
75
76use crate::FrontendInvoker;
77use crate::error::Error;
78
79fn expire_after_secs_to_millis(expire_after_secs: i64) -> Result<repr::Duration, Error> {
80    ensure!(
81        expire_after_secs >= 0,
82        InvalidQuerySnafu {
83            reason: format!("EXPIRE AFTER must be non-negative, got {expire_after_secs} seconds"),
84        }
85    );
86
87    expire_after_secs
88        .checked_mul(1_000)
89        .with_context(|| InvalidQuerySnafu {
90            reason: format!(
91                "EXPIRE AFTER value {expire_after_secs} seconds cannot be represented in milliseconds"
92            ),
93        })
94}
95
96// `GREPTIME_TIMESTAMP` is not used to distinguish when table is created automatically by flow
97pub const AUTO_CREATED_PLACEHOLDER_TS_COL: &str = "__ts_placeholder";
98
99pub const AUTO_CREATED_UPDATE_AT_TS_COL: &str = "update_at";
100
101/// Flow config that exists both in standalone&distributed mode
102#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
103#[serde(default)]
104pub struct FlowConfig {
105    pub num_workers: usize,
106    pub batching_mode: BatchingModeOptions,
107}
108
109impl Default for FlowConfig {
110    fn default() -> Self {
111        Self {
112            num_workers: (get_total_cpu_cores() / 2).max(1),
113            batching_mode: BatchingModeOptions::default(),
114        }
115    }
116}
117
118/// Options for flow node
119#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
120#[serde(default)]
121pub struct FlownodeOptions {
122    pub node_id: Option<u64>,
123    pub flow: FlowConfig,
124    pub grpc: GrpcOptions,
125    pub http: HttpOptions,
126    pub meta_client: Option<MetaClientOptions>,
127    pub logging: LoggingOptions,
128    pub tracing: TracingOptions,
129    pub query: QueryOptions,
130    pub memory: MemoryOptions,
131}
132
133impl Default for FlownodeOptions {
134    fn default() -> Self {
135        Self {
136            node_id: None,
137            flow: FlowConfig::default(),
138            grpc: GrpcOptions::default().with_bind_addr("127.0.0.1:3004"),
139            http: HttpOptions::default(),
140            meta_client: None,
141            logging: LoggingOptions::default(),
142            tracing: TracingOptions::default(),
143            // flownode's query option is set to 1 to throttle flow's query so
144            // that it won't use too much cpu or memory
145            query: QueryOptions {
146                parallelism: 1,
147                allow_query_fallback: false,
148                memory_pool_size: MemoryLimit::default(),
149                enable_per_region_metrics: false,
150                ..Default::default()
151            },
152            memory: MemoryOptions::default(),
153        }
154    }
155}
156
157impl Configurable for FlownodeOptions {
158    fn validate_sanitize(&mut self) -> common_config::error::Result<()> {
159        if self.flow.num_workers == 0 {
160            self.flow.num_workers = (get_total_cpu_cores() / 2).max(1);
161        }
162        Ok(())
163    }
164}
165
166/// Arc-ed FlowNodeManager, cheaper to clone
167pub type FlowStreamingEngineRef = Arc<StreamingEngine>;
168
169/// FlowNodeManager manages the state of all tasks in the flow node, which should be run on the same thread
170///
171/// The choice of timestamp is just using current system timestamp for now
172///
173pub struct StreamingEngine {
174    /// The handler to the worker that will run the dataflow
175    /// which is `!Send` so a handle is used
176    pub worker_handles: Vec<WorkerHandle>,
177    /// The selector to select a worker to run the dataflow
178    worker_selector: Mutex<usize>,
179    /// The query engine that will be used to parse the query and convert it to a dataflow plan
180    pub query_engine: Arc<dyn QueryEngine>,
181    /// Getting table name and table schema from table info manager
182    table_info_source: ManagedTableSource,
183    frontend_invoker: RwLock<Option<FrontendInvoker>>,
184    /// contains mapping from table name to global id, and table schema
185    node_context: RwLock<FlownodeContext>,
186    /// Contains all refill tasks
187    refill_tasks: RwLock<BTreeMap<FlowId, RefillTask>>,
188    flow_err_collectors: RwLock<BTreeMap<FlowId, ErrCollector>>,
189    src_send_buf_lens: RwLock<BTreeMap<TableId, watch::Receiver<usize>>>,
190    tick_manager: FlowTickManager,
191    /// This node id is only available in distributed mode, on standalone mode this is guaranteed to be `None`
192    pub node_id: Option<u32>,
193    /// Lock for flushing, will be `read` by `handle_inserts` and `write` by `flush_flow`
194    ///
195    /// So that a series of event like `inserts -> flush` can be handled correctly
196    flush_lock: RwLock<()>,
197}
198
199/// Building FlownodeManager
200impl StreamingEngine {
201    /// set frontend invoker
202    pub async fn set_frontend_invoker(&self, frontend: FrontendInvoker) {
203        *self.frontend_invoker.write().await = Some(frontend);
204    }
205
206    /// Create **without** setting `frontend_invoker`
207    pub fn new(
208        node_id: Option<u32>,
209        query_engine: Arc<dyn QueryEngine>,
210        table_meta: TableMetadataManagerRef,
211    ) -> Self {
212        let srv_map = ManagedTableSource::new(
213            table_meta.table_info_manager().clone(),
214            table_meta.table_name_manager().clone(),
215        );
216        let node_context = FlownodeContext::new(Box::new(srv_map.clone()) as _);
217        let tick_manager = FlowTickManager::new();
218        let worker_handles = Vec::new();
219        StreamingEngine {
220            worker_handles,
221            worker_selector: Mutex::new(0),
222            query_engine,
223            table_info_source: srv_map,
224            frontend_invoker: RwLock::new(None),
225            node_context: RwLock::new(node_context),
226            refill_tasks: Default::default(),
227            flow_err_collectors: Default::default(),
228            src_send_buf_lens: Default::default(),
229            tick_manager,
230            node_id,
231            flush_lock: RwLock::new(()),
232        }
233    }
234
235    /// Create a flownode manager with one worker
236    pub fn new_with_workers<'s>(
237        node_id: Option<u32>,
238        query_engine: Arc<dyn QueryEngine>,
239        table_meta: TableMetadataManagerRef,
240        num_workers: usize,
241    ) -> (Self, Vec<Worker<'s>>) {
242        let mut zelf = Self::new(node_id, query_engine, table_meta);
243
244        let workers: Vec<_> = (0..num_workers)
245            .map(|_| {
246                let (handle, worker) = create_worker();
247                zelf.add_worker_handle(handle);
248                worker
249            })
250            .collect();
251        (zelf, workers)
252    }
253
254    /// add a worker handler to manager, meaning this corresponding worker is under it's manage
255    pub fn add_worker_handle(&mut self, handle: WorkerHandle) {
256        self.worker_handles.push(handle);
257    }
258}
259
260#[derive(Debug)]
261pub enum DiffRequest {
262    Insert(Vec<(Row, repr::Timestamp)>),
263    Delete(Vec<(Row, repr::Timestamp)>),
264}
265
266impl DiffRequest {
267    pub fn len(&self) -> usize {
268        match self {
269            Self::Insert(v) => v.len(),
270            Self::Delete(v) => v.len(),
271        }
272    }
273
274    pub fn is_empty(&self) -> bool {
275        self.len() == 0
276    }
277}
278
279pub fn batches_to_rows_req(batches: Vec<Batch>) -> Result<Vec<DiffRequest>, Error> {
280    let mut reqs = Vec::new();
281    for batch in batches {
282        let mut rows = Vec::with_capacity(batch.row_count());
283        for i in 0..batch.row_count() {
284            let row = batch.get_row(i).context(EvalSnafu)?;
285            rows.push((Row::new(row), 0));
286        }
287        reqs.push(DiffRequest::Insert(rows));
288    }
289    Ok(reqs)
290}
291
292/// This impl block contains methods to send writeback requests to frontend
293impl StreamingEngine {
294    /// Return the number of requests it made
295    pub async fn send_writeback_requests(&self) -> Result<usize, Error> {
296        let all_reqs = self.generate_writeback_request().await?;
297        if all_reqs.is_empty() || all_reqs.iter().all(|v| v.1.is_empty()) {
298            return Ok(0);
299        }
300        let mut req_cnt = 0;
301        for (table_name, reqs) in all_reqs {
302            if reqs.is_empty() {
303                continue;
304            }
305            let (catalog, schema) = (table_name[0].clone(), table_name[1].clone());
306            let ctx = Arc::new(QueryContext::with(&catalog, &schema));
307
308            let (is_ts_placeholder, proto_schema) = match self
309                .try_fetch_existing_table(&table_name)
310                .await?
311                .context(UnexpectedSnafu {
312                    reason: format!("Table not found: {}", table_name.join(".")),
313                }) {
314                Ok(r) => r,
315                Err(e) => {
316                    if self
317                        .table_info_source
318                        .get_opt_table_id_from_name(&table_name)
319                        .await?
320                        .is_none()
321                    {
322                        // deal with both flow&sink table no longer exists
323                        // but some output is still in output buf
324                        common_telemetry::warn!(e; "Table `{}` no longer exists, skip writeback", table_name.join("."));
325                        continue;
326                    } else {
327                        return Err(e);
328                    }
329                }
330            };
331            let schema_len = proto_schema.len();
332
333            let total_rows = reqs.iter().map(|r| r.len()).sum::<usize>();
334            trace!(
335                "Sending {} writeback requests to table {}, reqs total rows={}",
336                reqs.len(),
337                table_name.join("."),
338                reqs.iter().map(|r| r.len()).sum::<usize>()
339            );
340
341            METRIC_FLOW_ROWS
342                .with_label_values(&["out-streaming"])
343                .inc_by(total_rows as u64);
344
345            let now = self.tick_manager.tick();
346            for req in reqs {
347                match req {
348                    DiffRequest::Insert(insert) => {
349                        let rows_proto: Vec<v1::Row> = insert
350                            .into_iter()
351                            .map(|(mut row, _ts)| {
352                                // extend `update_at` col if needed
353                                // if schema include a millisecond timestamp here, and result row doesn't have it, add it
354                                if row.len() < proto_schema.len()
355                                    && proto_schema[row.len()].datatype
356                                        == greptime_proto::v1::ColumnDataType::TimestampMillisecond
357                                            as i32
358                                {
359                                    row.extend([Value::from(
360                                        common_time::Timestamp::new_millisecond(now),
361                                    )]);
362                                }
363                                // ts col, if auto create
364                                if is_ts_placeholder {
365                                    ensure!(
366                                        row.len() == schema_len - 1,
367                                        InternalSnafu {
368                                            reason: format!(
369                                                "Row len mismatch, expect {} got {}",
370                                                schema_len - 1,
371                                                row.len()
372                                            )
373                                        }
374                                    );
375                                    row.extend([Value::from(
376                                        common_time::Timestamp::new_millisecond(0),
377                                    )]);
378                                }
379                                if row.len() != proto_schema.len() {
380                                    UnexpectedSnafu {
381                                        reason: format!(
382                                            "Flow output row length mismatch, expect {} got {}, the columns in schema are: {:?}",
383                                            proto_schema.len(),
384                                            row.len(),
385                                            proto_schema.iter().map(|c|&c.column_name).collect_vec()
386                                        ),
387                                    }
388                                    .fail()?;
389                                }
390                                Ok(row.into())
391                            })
392                            .collect::<Result<Vec<_>, Error>>()?;
393                        let table_name = table_name.last().unwrap().clone();
394                        let req = RowInsertRequest {
395                            table_name,
396                            rows: Some(v1::Rows {
397                                schema: proto_schema.clone(),
398                                rows: rows_proto,
399                            }),
400                        };
401                        req_cnt += 1;
402                        self.frontend_invoker
403                            .read()
404                            .await
405                            .as_ref()
406                            .with_context(|| UnexpectedSnafu {
407                                reason: "Expect a frontend invoker for flownode to write back",
408                            })?
409                            .row_inserts(RowInsertRequests { inserts: vec![req] }, ctx.clone())
410                            .await
411                            .map_err(BoxedError::new)
412                            .with_context(|_| ExternalSnafu {})?;
413                    }
414                    DiffRequest::Delete(remove) => {
415                        info!("original remove rows={:?}", remove);
416                        let rows_proto: Vec<v1::Row> = remove
417                            .into_iter()
418                            .map(|(mut row, _ts)| {
419                                row.extend(Some(Value::from(
420                                    common_time::Timestamp::new_millisecond(0),
421                                )));
422                                row.into()
423                            })
424                            .collect::<Vec<_>>();
425                        let table_name = table_name.last().unwrap().clone();
426                        let req = RowDeleteRequest {
427                            table_name,
428                            rows: Some(v1::Rows {
429                                schema: proto_schema.clone(),
430                                rows: rows_proto,
431                            }),
432                        };
433
434                        req_cnt += 1;
435                        self.frontend_invoker
436                            .read()
437                            .await
438                            .as_ref()
439                            .with_context(|| UnexpectedSnafu {
440                                reason: "Expect a frontend invoker for flownode to write back",
441                            })?
442                            .row_deletes(RowDeleteRequests { deletes: vec![req] }, ctx.clone())
443                            .await
444                            .map_err(BoxedError::new)
445                            .with_context(|_| ExternalSnafu {})?;
446                    }
447                }
448            }
449        }
450        Ok(req_cnt)
451    }
452
453    /// Generate writeback request for all sink table
454    pub async fn generate_writeback_request(
455        &self,
456    ) -> Result<BTreeMap<TableName, Vec<DiffRequest>>, Error> {
457        trace!("Start to generate writeback request");
458        let mut output = BTreeMap::new();
459        let mut total_row_count = 0;
460        for (name, sink_recv) in self
461            .node_context
462            .write()
463            .await
464            .sink_receiver
465            .iter_mut()
466            .map(|(n, (_s, r))| (n, r))
467        {
468            let mut batches = Vec::new();
469            while let Ok(batch) = sink_recv.try_recv() {
470                total_row_count += batch.row_count();
471                batches.push(batch);
472            }
473            let reqs = batches_to_rows_req(batches)?;
474            output.insert(name.clone(), reqs);
475        }
476        trace!("Prepare writeback req: total row count={}", total_row_count);
477        Ok(output)
478    }
479
480    /// Fetch table schema and primary key from table info source, if table not exist return None
481    async fn fetch_table_pk_schema(
482        &self,
483        table_name: &TableName,
484    ) -> Result<Option<(Vec<String>, Option<usize>, Vec<ColumnSchema>)>, Error> {
485        if let Some(table_id) = self
486            .table_info_source
487            .get_opt_table_id_from_name(table_name)
488            .await?
489        {
490            let table_info = self
491                .table_info_source
492                .get_table_info_value(&table_id)
493                .await?
494                .unwrap();
495            let meta = table_info.table_info.meta;
496            let schema = meta.schema.column_schemas().to_vec();
497            let primary_keys = meta
498                .primary_key_indices
499                .into_iter()
500                .map(|i| schema[i].name.clone())
501                .collect_vec();
502            let time_index = meta.schema.timestamp_index();
503            Ok(Some((primary_keys, time_index, schema)))
504        } else {
505            Ok(None)
506        }
507    }
508
509    /// return (primary keys, schema and if the table have a placeholder timestamp column)
510    /// schema of the table comes from flow's output plan
511    ///
512    /// adjust to add `update_at` column and ts placeholder if needed
513    async fn adjust_auto_created_table_schema(
514        &self,
515        schema: &RelationDesc,
516    ) -> Result<(Vec<String>, Vec<ColumnSchema>, bool), Error> {
517        // TODO(discord9): consider remove buggy auto create by schema
518
519        // TODO(discord9): use default key from schema
520        let primary_keys = schema
521            .typ()
522            .keys
523            .first()
524            .map(|v| {
525                v.column_indices
526                    .iter()
527                    .map(|i| {
528                        schema
529                            .get_name(*i)
530                            .clone()
531                            .unwrap_or_else(|| format!("col_{i}"))
532                    })
533                    .collect_vec()
534            })
535            .unwrap_or_default();
536        let update_at = ColumnSchema::new(
537            AUTO_CREATED_UPDATE_AT_TS_COL,
538            ConcreteDataType::timestamp_millisecond_datatype(),
539            true,
540        );
541
542        let original_schema = relation_desc_to_column_schemas_with_fallback(schema);
543
544        let mut with_auto_added_col = original_schema.clone();
545        with_auto_added_col.push(update_at);
546
547        // if no time index, add one as placeholder
548        let no_time_index = schema.typ().time_index.is_none();
549        if no_time_index {
550            let ts_col = ColumnSchema::new(
551                AUTO_CREATED_PLACEHOLDER_TS_COL,
552                ConcreteDataType::timestamp_millisecond_datatype(),
553                true,
554            )
555            .with_time_index(true);
556            with_auto_added_col.push(ts_col);
557        }
558
559        Ok((primary_keys, with_auto_added_col, no_time_index))
560    }
561}
562
563/// Flow Runtime related methods
564impl StreamingEngine {
565    /// run in common_runtime background runtime
566    pub fn run_background(
567        self: Arc<Self>,
568        shutdown: Option<broadcast::Receiver<()>>,
569    ) -> JoinHandle<()> {
570        info!("Starting flownode manager's background task");
571        common_runtime::spawn_global(async move { self.run(shutdown).await })
572    }
573
574    /// log all flow errors
575    pub async fn log_all_errors(&self) {
576        for (f_id, f_err) in self.flow_err_collectors.read().await.iter() {
577            let all_errors = f_err.get_all().await;
578            if !all_errors.is_empty() {
579                let all_errors = all_errors
580                    .into_iter()
581                    .map(|i| format!("{:?}", i))
582                    .join("\n");
583                common_telemetry::error!("Flow {} has following errors: {}", f_id, all_errors);
584            }
585        }
586    }
587
588    /// Trigger dataflow running, and then send writeback request to the source sender
589    ///
590    /// note that this method didn't handle input mirror request, as this should be handled by grpc server
591    pub async fn run(&self, mut shutdown: Option<broadcast::Receiver<()>>) {
592        debug!("Starting to run");
593        let default_interval = Duration::from_secs(1);
594        let mut tick_interval = tokio::time::interval(default_interval);
595        // burst mode, so that if we miss a tick, we will run immediately to fully utilize the cpu
596        tick_interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Burst);
597        let mut avg_spd = 0; // rows/sec
598        let mut since_last_run = tokio::time::Instant::now();
599        let run_per_trace = 10;
600        let mut run_cnt = 0;
601        loop {
602            // TODO(discord9): only run when new inputs arrive or scheduled to
603            let row_cnt = self.run_available(false).await.unwrap_or_else(|err| {
604                common_telemetry::error!(err;"Run available errors");
605                0
606            });
607
608            if let Err(err) = self.send_writeback_requests().await {
609                common_telemetry::error!(err;"Send writeback request errors");
610            };
611            self.log_all_errors().await;
612
613            // determine if need to shutdown
614            match &shutdown.as_mut().map(|s| s.try_recv()) {
615                Some(Ok(())) => {
616                    info!("Shutdown flow's main loop");
617                    break;
618                }
619                Some(Err(TryRecvError::Empty)) => (),
620                Some(Err(TryRecvError::Closed)) => {
621                    common_telemetry::error!("Shutdown channel is closed");
622                    break;
623                }
624                Some(Err(TryRecvError::Lagged(num))) => {
625                    common_telemetry::error!(
626                        "Shutdown channel is lagged by {}, meaning multiple shutdown cmd have been issued",
627                        num
628                    );
629                    break;
630                }
631                None => (),
632            }
633
634            // for now we want to batch rows until there is around `BATCH_SIZE` rows in send buf
635            // before trigger a run of flow's worker
636            let wait_for = since_last_run.elapsed();
637
638            // last runs insert speed
639            let cur_spd = row_cnt * 1000 / wait_for.as_millis().max(1) as usize;
640            // rapid increase, slow decay
641            avg_spd = if cur_spd > avg_spd {
642                cur_spd
643            } else {
644                (9 * avg_spd + cur_spd) / 10
645            };
646            let new_wait = BATCH_SIZE * 1000 / avg_spd.max(1); //in ms
647            let new_wait = Duration::from_millis(new_wait as u64).min(default_interval);
648
649            // print trace every `run_per_trace` times so that we can see if there is something wrong
650            // but also not get flooded with trace
651            if run_cnt >= run_per_trace {
652                trace!("avg_spd={} r/s, cur_spd={} r/s", avg_spd, cur_spd);
653                trace!("Wait for {} ms, row_cnt={}", new_wait.as_millis(), row_cnt);
654                run_cnt = 0;
655            } else {
656                run_cnt += 1;
657            }
658
659            METRIC_FLOW_RUN_INTERVAL_MS.set(new_wait.as_millis() as i64);
660            since_last_run = tokio::time::Instant::now();
661            tokio::select! {
662                _ = tick_interval.tick() => (),
663                _ = tokio::time::sleep(new_wait) => ()
664            }
665        }
666        // flow is now shutdown, drop frontend_invoker early so a ref cycle(in standalone mode) can be prevent:
667        // FlowWorkerManager.frontend_invoker -> FrontendInvoker.inserter
668        // -> Inserter.node_manager -> NodeManager.flownode -> Flownode.flow_streaming_engine.frontend_invoker
669        self.frontend_invoker.write().await.take();
670    }
671
672    /// Run all available subgraph in the flow node
673    /// This will try to run all dataflow in this node
674    ///
675    /// set `blocking` to true to wait until worker finish running
676    /// false to just trigger run and return immediately
677    /// return numbers of rows send to worker(Inaccuary)
678    /// TODO(discord9): add flag for subgraph that have input since last run
679    pub async fn run_available(&self, blocking: bool) -> Result<usize, Error> {
680        let mut row_cnt = 0;
681
682        let now = self.tick_manager.tick();
683        for worker in self.worker_handles.iter() {
684            // TODO(discord9): consider how to handle error in individual worker
685            worker.run_available(now, blocking).await?;
686        }
687        // check row send and rows remain in send buf
688        let flush_res = if blocking {
689            let ctx = self.node_context.read().await;
690            ctx.flush_all_sender().await
691        } else {
692            match self.node_context.try_read() {
693                Ok(ctx) => ctx.flush_all_sender().await,
694                Err(_) => return Ok(row_cnt),
695            }
696        };
697        match flush_res {
698            Ok(r) => {
699                common_telemetry::trace!("Total flushed {} rows", r);
700                row_cnt += r;
701            }
702            Err(err) => {
703                common_telemetry::error!("Flush send buf errors: {:?}", err);
704            }
705        };
706
707        Ok(row_cnt)
708    }
709
710    /// send write request to related source sender
711    pub async fn handle_write_request(
712        &self,
713        region_id: RegionId,
714        rows: Vec<DiffRow>,
715        batch_datatypes: &[ConcreteDataType],
716    ) -> Result<(), Error> {
717        let rows_len = rows.len();
718        let table_id = region_id.table_id();
719        let _timer = METRIC_FLOW_INSERT_ELAPSED
720            .with_label_values(&[table_id.to_string().as_str()])
721            .start_timer();
722        self.node_context
723            .read()
724            .await
725            .send(table_id, rows, batch_datatypes)
726            .await?;
727        trace!(
728            "Handling write request for table_id={} with {} rows",
729            table_id, rows_len
730        );
731        Ok(())
732    }
733}
734
735/// Create&Remove flow
736impl StreamingEngine {
737    /// remove a flow by it's id
738    pub async fn remove_flow_inner(&self, flow_id: FlowId) -> Result<(), Error> {
739        for handle in self.worker_handles.iter() {
740            if handle.contains_flow(flow_id).await? {
741                handle.remove_flow(flow_id).await?;
742                break;
743            }
744        }
745        self.node_context.write().await.remove_flow(flow_id);
746        Ok(())
747    }
748
749    /// Return task id if a new task is created, otherwise return None
750    ///
751    /// steps to create task:
752    /// 1. parse query into typed plan(and optional parse expire_after expr)
753    /// 2. render source/sink with output table id and used input table id
754    pub async fn create_flow_inner(&self, args: CreateFlowArgs) -> Result<Option<FlowId>, Error> {
755        let CreateFlowArgs {
756            flow_id,
757            sink_table_name,
758            source_table_ids,
759            create_if_not_exists,
760            or_replace,
761            expire_after: expire_after_secs,
762            eval_interval: _,
763            comment,
764            sql,
765            flow_options,
766            query_ctx,
767            ..
768        } = args;
769        let expire_after = expire_after_secs
770            .map(expire_after_secs_to_millis)
771            .transpose()?;
772
773        let mut node_ctx = self.node_context.write().await;
774        // assign global id to source and sink table
775        for source in &source_table_ids {
776            node_ctx
777                .assign_global_id_to_table(&self.table_info_source, None, Some(*source))
778                .await?;
779        }
780        node_ctx
781            .assign_global_id_to_table(&self.table_info_source, Some(sink_table_name.clone()), None)
782            .await?;
783
784        node_ctx.register_task_src_sink(flow_id, &source_table_ids, sink_table_name.clone());
785
786        node_ctx.query_context = query_ctx.map(Arc::new);
787        // construct a active dataflow state with it
788        let flow_plan = sql_to_flow_plan(&mut node_ctx, &self.query_engine, &sql).await?;
789
790        debug!("Flow {:?}'s Plan is {:?}", flow_id, flow_plan);
791
792        // check schema against actual table schema if exists
793        // if not exist create sink table immediately
794        if let Some((_, _, real_schema)) = self.fetch_table_pk_schema(&sink_table_name).await? {
795            let auto_schema = relation_desc_to_column_schemas_with_fallback(&flow_plan.schema);
796
797            // for column schema, only `data_type` need to be check for equality
798            // since one can omit flow's column name when write flow query
799            // print a user friendly error message about mismatch and how to correct them
800            for (idx, zipped) in auto_schema
801                .iter()
802                .zip_longest(real_schema.iter())
803                .enumerate()
804            {
805                match zipped {
806                    EitherOrBoth::Both(auto, real) => {
807                        if auto.data_type != real.data_type {
808                            InvalidQuerySnafu {
809                                    reason: format!(
810                                        "Column {}(name is '{}', flow inferred name is '{}')'s data type mismatch, expect {:?} got {:?}",
811                                        idx,
812                                        real.name,
813                                        auto.name,
814                                        real.data_type,
815                                        auto.data_type
816                                    ),
817                                }
818                                .fail()?;
819                        }
820                    }
821                    EitherOrBoth::Right(real) if real.data_type.is_timestamp() => {
822                        // if table is auto created, the last one or two column should be timestamp(update at and ts placeholder)
823                        continue;
824                    }
825                    _ => InvalidQuerySnafu {
826                        reason: format!(
827                            "schema length mismatched, expected {} found {}",
828                            real_schema.len(),
829                            auto_schema.len()
830                        ),
831                    }
832                    .fail()?,
833                }
834            }
835        } else {
836            // assign inferred schema to sink table
837            // create sink table
838            let did_create = self
839                .create_table_from_relation(
840                    &format!("flow-id={flow_id}"),
841                    &sink_table_name,
842                    &flow_plan.schema,
843                )
844                .await?;
845            if !did_create {
846                UnexpectedSnafu {
847                    reason: format!("Failed to create table {:?}", sink_table_name),
848                }
849                .fail()?;
850            }
851        }
852
853        node_ctx.add_flow_plan(flow_id, flow_plan.clone());
854
855        let _ = comment;
856        let _ = flow_options;
857
858        // TODO(discord9): add more than one handles
859        let sink_id = node_ctx.table_repr.get_by_name(&sink_table_name).unwrap().1;
860        let sink_sender = node_ctx.get_sink_by_global_id(&sink_id)?;
861
862        let source_ids = source_table_ids
863            .iter()
864            .map(|id| node_ctx.table_repr.get_by_table_id(id).unwrap().1)
865            .collect_vec();
866        let source_receivers = source_ids
867            .iter()
868            .map(|id| {
869                node_ctx
870                    .get_source_by_global_id(id)
871                    .map(|s| s.get_receiver())
872            })
873            .collect::<Result<Vec<_>, _>>()?;
874        let err_collector = ErrCollector::default();
875        self.flow_err_collectors
876            .write()
877            .await
878            .insert(flow_id, err_collector.clone());
879        // TODO(discord9): load balance?
880        let handle = self.get_worker_handle_for_create_flow().await;
881        let create_request = worker::Request::Create {
882            flow_id,
883            plan: flow_plan,
884            sink_id,
885            sink_sender,
886            source_ids,
887            src_recvs: source_receivers,
888            expire_after,
889            or_replace,
890            create_if_not_exists,
891            err_collector,
892        };
893
894        handle.create_flow(create_request).await?;
895        info!("Successfully create flow with id={}", flow_id);
896        Ok(Some(flow_id))
897    }
898
899    pub async fn flush_flow_inner(&self, flow_id: FlowId) -> Result<usize, Error> {
900        debug!("Starting to flush flow_id={:?}", flow_id);
901        // lock to make sure writes before flush are written to flow
902        // and immediately drop to prevent following writes to be blocked
903        drop(self.flush_lock.write().await);
904        let flushed_input_rows = self.node_context.read().await.flush_all_sender().await?;
905        let rows_send = self.run_available(true).await?;
906        let row = self.send_writeback_requests().await?;
907        debug!(
908            "Done to flush flow_id={:?} with {} input rows flushed, {} rows sent and {} output rows flushed",
909            flow_id, flushed_input_rows, rows_send, row
910        );
911        Ok(row)
912    }
913
914    pub async fn flow_exist_inner(&self, flow_id: FlowId) -> Result<bool, Error> {
915        let mut exist = false;
916        for handle in self.worker_handles.iter() {
917            if handle.contains_flow(flow_id).await? {
918                exist = true;
919                break;
920            }
921        }
922        Ok(exist)
923    }
924}
925
926/// FlowTickManager is a manager for flow tick, which trakc flow execution progress
927///
928/// TODO(discord9): better way to do it, and not expose flow tick even to other flow to avoid
929/// TSO coord mess
930#[derive(Clone, Debug)]
931pub struct FlowTickManager {
932    /// The starting instant of the flow, used with `start_timestamp` to calculate the current timestamp
933    start: Instant,
934    /// The timestamp when the flow started
935    start_timestamp: repr::Timestamp,
936}
937
938impl Default for FlowTickManager {
939    fn default() -> Self {
940        Self::new()
941    }
942}
943
944impl FlowTickManager {
945    pub fn new() -> Self {
946        FlowTickManager {
947            start: Instant::now(),
948            start_timestamp: SystemTime::now()
949                .duration_since(SystemTime::UNIX_EPOCH)
950                .unwrap()
951                .as_millis() as repr::Timestamp,
952        }
953    }
954
955    /// Return the current timestamp in milliseconds
956    ///
957    /// TODO(discord9): reconsider since `tick()` require a monotonic clock and also need to survive recover later
958    pub fn tick(&self) -> repr::Timestamp {
959        let current = Instant::now();
960        let since_the_epoch = current - self.start;
961        since_the_epoch.as_millis() as repr::Timestamp + self.start_timestamp
962    }
963}
flow/adapter.rs

flow/
adapter.rs
