[D] Advice on best way to handle streaming record batches in memory aggregation [datafusion]

["erchirag (via GitHub)" <gi...@apache.org>]

GitHub user erchirag created a discussion: Advice on best way to handle streaming record batches in memory aggregation

Hi I have scenario where in one node gets bunch of record batches from worker nodes and then it needs to aggregate them. I have a working code but I want to understand if this is the best way to do it or there are other abstractions available in the datafusion library which can reduce the boiler plate code here.

My approach is as follows:

It DataFusion's [StreamingTable](https://docs.rs/datafusion/latest/datafusion/catalog/streaming/struct.StreamingTable.html) to process data arriving from a Tokio MPSC channel.

A custom struct StreamPartition implements the PartitionStream trait. Its execute() method is key:
It takes a ReceiverStream<DFResult<RecordBatch>> (from the MPSC channel).
It uses RecordBatchStreamAdapter to wrap this stream, making it a SendableRecordBatchStream that DataFusion's execution engine can consume.
An instance of StreamingTable is created using:
The SchemaRef of the incoming data.
A Vec containing Arcs of our StreamPartition (one for each partition, though here it's effectively a single stream source).
This Arc<StreamingTable> (which is a TableProvider) is then registered with the SessionContext using ctx.register_table("table_name", table_provider).
This setup allows DataFusion to query the MPSC channel as if it were a SQL table, with StreamingTable and our StreamPartition implementation handling the mechanics of pulling data in RecordBatches when the query executes.


Here is the working code sample:

```rust
use std::fs::File;
use std::sync::{Arc, Mutex};
use std::time::SystemTime;
use std::fmt::Debug;
use std::env; // Added for command-line arguments
use std::path::Path; // Added for path checking

// Explicit imports from arrow sub-crates
use arrow_array::builder::{Float32Builder, Float64Builder, Int32Builder, Int64Builder, StringBuilder, TimestampNanosecondBuilder};
use arrow_array::RecordBatch;
use arrow_ipc::writer::{StreamWriter, IpcWriteOptions};
use arrow_ipc::reader::StreamReader;
use arrow_schema::{DataType, Field, Schema, SchemaRef, TimeUnit, ArrowError};
use datafusion::arrow::compute::concat_batches;
use datafusion::error::{DataFusionError, Result as DFResult};
use datafusion::prelude::*;
use datafusion::physical_plan::stream::{EmptyRecordBatchStream, RecordBatchStreamAdapter};
use datafusion::physical_plan::streaming::PartitionStream;
use datafusion::physical_plan::SendableRecordBatchStream;
use datafusion::execution::context::TaskContext;
use datafusion::catalog::streaming::StreamingTable;

use futures::{StreamExt, TryStreamExt};
use tokio::sync::mpsc;
use tokio_stream::wrappers::ReceiverStream;
use async_trait::async_trait;

const DEFAULT_ARROW_STREAM_FILE_PATH: &str = "input.arrow_stream"; // Renamed from ARROW_STREAM_FILE_PATH
const NUM_STREAM_REPETITIONS: usize = 10;
const CHANNEL_BUFFER_SIZE: usize = 100;

/// Reads the schema from the Arrow IPC stream file.
fn get_schema_from_file(file_path: &str) -> DFResult<SchemaRef> {
    let file = File::open(file_path)?;
    let reader = StreamReader::try_new(file, None)?;
    Ok(reader.schema())
}

/// Producer task: Reads RecordBatches and sends them over MPSC channel.
async fn producer_task(tx: mpsc::Sender<DFResult<RecordBatch>>, file_path: String) -> DFResult<()> { // Added file_path argument
    println!("Producer: Starting to read {} and send batches.", file_path); // Use argument
    for i in 0..NUM_STREAM_REPETITIONS {
        let file = File::open(&file_path)?; // Use argument
        let mut reader = StreamReader::try_new(file, None)?;
        println!("Producer: Sending repetition #{}", i + 1);
        while let Some(batch_result) = reader.next() {
            match batch_result {
                Ok(batch) => {
                    if tx.send(Ok(batch)).await.is_err() {
                        eprintln!("Producer: Receiver closed, stopping.");
                        return Ok(());
                    }
                }
                Err(e) => {
                    eprintln!("Producer: Error reading batch: {:?}", e);
                    let df_error = DataFusionError::ArrowError(e, None);
                    if tx.send(Err(df_error)).await.is_err() {
                        eprintln!("Producer: Receiver closed while sending error, stopping.");
                        return Ok(());
                    }
                }
            }
        }
    }
    println!("Producer: Finished sending all batches.");
    Ok(())
}

/// Builds the SQL query string for aggregation based on the schema and rules.
fn build_sql_query(schema: &Schema, table_name: &str) -> Result<String, String> {
    let mut select_clauses = Vec::new();
    let mut groupby_clauses = Vec::new();
    for field in schema.fields() {
        let name = field.name();
        let quoted_name = format!("\"{}\"", name.replace("\"", "\"\""));
        match field.data_type() {
            DataType::Utf8 | DataType::LargeUtf8 | DataType::Timestamp(_, _) => {
                select_clauses.push(quoted_name.clone());
                groupby_clauses.push(quoted_name);
            }
            DataType::Int8 | DataType::Int16 | DataType::Int32 | DataType::Int64 |
            DataType::UInt8 | DataType::UInt16 | DataType::UInt32 | DataType::UInt64 |
            DataType::Float16 | DataType::Float32 | DataType::Float64 => {
                if name.ends_with("_sum") || name.ends_with("_count") {
                    select_clauses.push(format!("SUM({}) AS {}", quoted_name, quoted_name));
                } else if name.ends_with("_min") {
                    select_clauses.push(format!("MIN({}) AS {}", quoted_name, quoted_name));
                } else if name.ends_with("_max") {
                    select_clauses.push(format!("MAX({}) AS {}", quoted_name, quoted_name));
                } else {
                    select_clauses.push(format!("SUM({}) AS {}", quoted_name, quoted_name));
                }
            }
            _ => { println!("Skipping column '{}' with unhandled type: {:?}", name, field.data_type()); }
        }
    }
    if select_clauses.is_empty() { return Err("No columns suitable for selecting or grouping found.".to_string()); }
    let select_statement = select_clauses.join(", ");
    if groupby_clauses.is_empty() {
        Ok(format!("SELECT {} FROM \"{}\"", select_statement, table_name))
    } else {
        Ok(format!(
            "SELECT {} FROM \"{}\" GROUP BY {}",
            select_statement,
            table_name,
            groupby_clauses.join(", ")
        ))
    }
}

/// PartitionStream wrapper that holds the stream within a Mutex
/// to ensure it's Sync and consumed only once.
/// It takes the stream yielding DataFusionError and uses RecordBatchStreamAdapter.
struct StreamPartition {
    schema: SchemaRef,
    // Store the original stream yielding DataFusionError
    stream_mutex: Arc<Mutex<Option<ReceiverStream<DFResult<RecordBatch>>>>>,
}

impl Debug for StreamPartition {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("StreamPartition")
         .field("schema", &self.schema)
         .field("stream_mutex", &"<Mutex containing Option<ReceiverStream<DFResult<RecordBatch>>>>")
         .finish()
    }
}

#[async_trait]
impl PartitionStream for StreamPartition {
    fn schema(&self) -> &SchemaRef {
        &self.schema
    }

    fn execute(&self, _ctx: Arc<TaskContext>) -> SendableRecordBatchStream {
        let mut stream_guard = self.stream_mutex.lock().unwrap();
        match stream_guard.take() {
            Some(stream) => {
                let record_batch_stream_adapter = RecordBatchStreamAdapter::new(
                    self.schema.clone(),
                    stream
                );

                Box::pin(record_batch_stream_adapter) as SendableRecordBatchStream
            },
            None => {
                println!("Warning: PartitionStream executed more than once. Returning empty stream.");
                Box::pin(EmptyRecordBatchStream::new(self.schema.clone()))
            }
        }
    }
}

/// Aggregator task: Receives batches, sets up DataFusion streaming table,
/// executes aggregation query, and prints results.
async fn aggregator_task(
    rx: mpsc::Receiver<DFResult<RecordBatch>>,
    input_schema: SchemaRef,
) -> DFResult<()> {
    println!("Aggregator: Task started. Waiting for batches...");

    let config = SessionConfig::new().with_batch_size(8192);
    let ctx = SessionContext::new_with_config(config);

    let stream_from_channel = ReceiverStream::new(rx);

    let partition_stream: Arc<dyn PartitionStream> = Arc::new(StreamPartition {
        schema: input_schema.clone(),
        stream_mutex: Arc::new(Mutex::new(Some(stream_from_channel))),
    });

    let table_name = "streaming_sensor_data";

    let table_provider = Arc::new(StreamingTable::try_new(
        input_schema.clone(),
        vec![partition_stream]
    )?);

    ctx.register_table(table_name, table_provider)?;
    println!("Aggregator: Streaming table '{}' registered.", table_name);

    let sql_query = match build_sql_query(&input_schema, table_name) {
        Ok(q) => q,
        Err(e) => {
            eprintln!("Aggregator: Failed to build SQL query: {}", e);
            return Err(DataFusionError::Execution(e));
        }
    };
    println!("Aggregator: Executing SQL query:\n{}", sql_query);

    let df = ctx.sql(&sql_query).await?;

    let mut result_batches = df.collect().await?;
    println!("\n--- Aggregated Results ---");
    let mut total_result_rows = 0;

    // merge the results into a single batch
    let merged_batch = concat_batches(&input_schema, &result_batches)?;
    println!("\n--- Merged Aggregated Results ---");
    if merged_batch.num_rows() > 0 {
        println!("Merged batch schema: {:?}", merged_batch.schema());
        arrow::util::pretty::print_batches(&[merged_batch])?;
    } else {
        println!("No rows in the merged aggregated result.");
    }

    if total_result_rows == 0 {
        println!("Aggregator: No rows in the aggregated result.");
    } else {
        println!("Aggregator: Finished processing. Total aggregated rows: {}", total_result_rows);
    }
    Ok(())
}

#[tokio::main]
async fn main() -> DFResult<()> {
    let file_path_to_use: String;

    if let Some(arg_path) = env::args().nth(1) {
        file_path_to_use = arg_path;
    } else {
        file_path_to_use = DEFAULT_ARROW_STREAM_FILE_PATH.to_string();
        println!("No file path argument provided. Using default path: {}", file_path_to_use);
        return Err();
    }

    let input_schema = match get_schema_from_file(&file_path_to_use) {
        Ok(s) => s,
        Err(e) => {
            eprintln!("Critical: Failed to read schema from {}: {:?}", file_path_to_use, e);
            return Err(e.into());
        }
    };
    println!("Read input schema: {:?} from {}", input_schema.fields(), file_path_to_use);

    let (tx, rx) = mpsc::channel::<DFResult<RecordBatch>>(CHANNEL_BUFFER_SIZE);

    let producer_file_path_clone = file_path_to_use.clone();
    let producer_handle = tokio::spawn(async move {
        if let Err(e) = producer_task(tx, producer_file_path_clone).await {
            eprintln!("Producer task error: {:?}", e);
        }
    });

    let aggregator_handle = tokio::spawn(async move {
        if let Err(e) = aggregator_task(rx, input_schema).await {
            eprintln!("Aggregator task error: {:?}", e);
        }
    });

    let (producer_res, aggregator_res) = tokio::join!(producer_handle, aggregator_handle);

    if producer_res.is_err() { eprintln!("Producer task panicked or encountered an unhandled error."); }
    if aggregator_res.is_err() { eprintln!("Aggregator task panicked or encountered an unhandled error."); }

    println!("Processing complete.");
    Ok(())
}
```


GitHub link: https://github.com/apache/datafusion/discussions/16035

----
This is an automatically sent email for github@datafusion.apache.org.
To unsubscribe, please send an email to: github-unsubscribe@datafusion.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: github-unsubscribe@datafusion.apache.org
For additional commands, e-mail: github-help@datafusion.apache.org