bug: drains output channel when batching results (#763)

Drains the output channel. Previously was causing queries to block on
send.

Puts the tokio runtime into an `Arc` to share it -- the `handle()` was
not working like I wanted since it was being dropped (a blocking method)
from an asynchronous context. Unsure why, but swapping to use `Arc`, as
is also recommended, works fine.

https://docs.rs/tokio/latest/tokio/runtime/struct.Runtime.html#sharing

Closes #775

---------

Co-authored-by: Ben Chambers <35960+bjchambers@users.noreply.github.com>

crates/sparrow-runtime/src/execute/output.rs

@@ -109,7 +109,6 @@ pub(super) fn write(
                 batch
             };
 
-
             if batch.num_rows() > max_batch_size {
                 for start in (0..batch.num_rows()).step_by(max_batch_size) {
                     let end = (start + max_batch_size).min(batch.num_rows());
@@ -121,7 +120,6 @@ pub(super) fn write(
                 yield post_process_batch(&sink_schema, batch, &key_hash_inverse).await;
             }
 
-
             if limit_rows && remaining == 0 {
                 break;
             }

crates/sparrow-session/src/execution.rs

@@ -2,17 +2,17 @@ use arrow_array::RecordBatch;
 use arrow_schema::SchemaRef;
 use futures::future::BoxFuture;
 use futures::stream::BoxStream;
-use futures::{StreamExt, TryStreamExt};
+use futures::StreamExt;
 use sparrow_api::kaskada::v1alpha::ExecuteResponse;
 
 use crate::Error;
 
 pub struct Execution {
-    /// Tokio runtme managing this execution.
-    rt: tokio::runtime::Runtime,
+    /// Tokio handle managing this execution.
+    handle: tokio::runtime::Handle,
     /// Channel to receive output on.
     output: tokio_stream::wrappers::ReceiverStream<RecordBatch>,
-    /// Future which resolves to the first error or  None.
+    // Future that resolves to the first error, if one occurred.
     status: Status,
     /// Stop signal. Send `true` to stop execution.
     stop_signal_rx: tokio::sync::watch::Sender<bool>,
@@ -27,21 +27,31 @@ enum Status {
 
 impl Execution {
     pub(super) fn new(
-        rt: tokio::runtime::Runtime,
+        handle: tokio::runtime::Handle,
         output_rx: tokio::sync::mpsc::Receiver<RecordBatch>,
         progress: BoxStream<'static, error_stack::Result<ExecuteResponse, Error>>,
         stop_signal_rx: tokio::sync::watch::Sender<bool>,
         schema: SchemaRef,
     ) -> Self {
         let output = tokio_stream::wrappers::ReceiverStream::new(output_rx);
+
+        // Constructs a futures that resolves to the first error, if one occurred.
         let status = Status::Running(Box::pin(async move {
-            let mut progress = progress;
-            while (progress.try_next().await?).is_some() {}
-            Ok(())
+            let mut errors = progress
+                .filter_map(|result| {
+                    futures::future::ready(if let Err(e) = result { Some(e) } else { None })
+                })
+                .boxed();
+            let first_error = errors.next().await;
+            if let Some(first_error) = first_error {
+                Err(first_error)
+            } else {
+                Ok(())
+            }
         }));
 
         Self {
-            rt,
+            handle,
             output,
             status,
             stop_signal_rx,
@@ -56,12 +66,12 @@ impl Execution {
     /// status (and return) accordingly.
     fn is_done(&mut self) -> error_stack::Result<(), Error> {
         let result = match &mut self.status {
-            Status::Running(future) => {
+            Status::Running(progress) => {
                 // Based on the implementation of `FutureExt::now_or_never`:
                 let noop_waker = futures::task::noop_waker();
                 let mut cx = std::task::Context::from_waker(&noop_waker);
 
-                match future.as_mut().poll(&mut cx) {
+                match progress.as_mut().poll(&mut cx) {
                     std::task::Poll::Ready(x) => x,
                     _ => return Ok(()),
                 }
@@ -99,29 +109,36 @@ impl Execution {
 
     pub fn next_blocking(&mut self) -> error_stack::Result<Option<RecordBatch>, Error> {
         self.is_done()?;
-        Ok(self.rt.block_on(self.output.next()))
+        Ok(self.handle.block_on(self.output.next()))
     }
 
     pub async fn collect_all(self) -> error_stack::Result<Vec<RecordBatch>, Error> {
-        // TODO: For large outputs, we likely need to drain the output while waiting for the future.
-        match self.status {
-            Status::Running(future) => future.await?,
+        let progress = match self.status {
+            Status::Running(progress) => progress,
             Status::Failed => error_stack::bail!(Error::ExecutionFailed),
-            _ => {}
+            Status::Completed => {
+                // If the progress channel has completed without error, we know that the output channel
+                // hasn't filled up, so we can go ahead and collect the output
+                return Ok(self.output.collect().await);
+            }
         };
 
-        Ok(self.output.collect().await)
+        let output = self.output.collect::<Vec<_>>();
+
+        let (first_error, output) = futures::join!(progress, output);
+        if let Err(e) = first_error {
+            Err(e)
+        } else {
+            Ok(output)
+        }
     }
 
     pub fn collect_all_blocking(self) -> error_stack::Result<Vec<RecordBatch>, Error> {
-        // TODO: For large outputs, we likely need to drain the output while waiting for the future.
-        match self.status {
-            Status::Running(future) => self.rt.block_on(future)?,
-            Status::Failed => error_stack::bail!(Error::ExecutionFailed),
-            _ => {}
-        };
-
-        Ok(self.rt.block_on(self.output.collect()))
+        // In order to check the running status, we have to enter the runtime regardless,
+        // so there's no reason to check the status prior to entering the runtime
+        // here.
+        let handle = self.handle.clone();
+        handle.block_on(self.collect_all())
     }
 }
 

crates/sparrow-session/src/session.rs

@@ -21,7 +21,6 @@ use uuid::Uuid;
 use crate::execution::Execution;
 use crate::{Error, Expr, Literal, Table};
 
-#[derive(Default)]
 pub struct Session {
     data_context: DataContext,
     dfg: Dfg,
@@ -34,6 +33,20 @@ pub struct Session {
     /// udf as well.
     udfs: HashMap<Uuid, Arc<dyn Udf>>,
     object_store_registry: Arc<ObjectStoreRegistry>,
+    rt: tokio::runtime::Runtime,
+}
+
+impl Default for Session {
+    fn default() -> Self {
+        Self {
+            data_context: Default::default(),
+            dfg: Default::default(),
+            key_hash_inverse: Default::default(),
+            udfs: Default::default(),
+            object_store_registry: Default::default(),
+            rt: tokio::runtime::Runtime::new().expect("tokio runtime"),
+        }
+    }
 }
 
 #[derive(Default)]
@@ -454,13 +467,6 @@ impl Session {
         .into_report()
         .change_context(Error::Compile)?;
 
-        // Switch to the Tokio async pool. This seems gross.
-        // Create the runtime.
-        //
-        // TODO: Figure out how to asynchronously pass results back to Python?
-        let rt = tokio::runtime::Runtime::new()
-            .into_report()
-            .change_context(Error::Execute)?;
         let (output_tx, output_rx) = tokio::sync::mpsc::channel(10);
         let destination = Destination::Channel(output_tx);
 
@@ -481,7 +487,8 @@ impl Session {
             });
 
         // Hacky. Use the existing execution logic. This weird things with downloading checkpoints, etc.
-        let progress = rt
+        let progress = self
+            .rt
             .block_on(sparrow_runtime::execute::execute_new(
                 plan,
                 destination,
@@ -495,8 +502,9 @@ impl Session {
             .map_err(|e| e.change_context(Error::Execute))
             .boxed();
 
+        let handle = self.rt.handle().clone();
         Ok(Execution::new(
-            rt,
+            handle,
             output_rx,
             progress,
             stop_signal_tx,

python/pysrc/kaskada/_timestream.py

@@ -565,7 +565,9 @@ def select(self, *args: str) -> Timestream:
         """
         return Timestream._call("select_fields", self, *args)
 
-    def substring(self, start: Optional[int] = None, end: Optional[int] = None) -> Timestream:
+    def substring(
+        self, start: Optional[int] = None, end: Optional[int] = None
+    ) -> Timestream:
         """Return a Timestream with a substring between the start and end indices.
 
         Args:

python/pysrc/kaskada/sources/__init__.py

@@ -3,4 +3,12 @@
 from .source import Source
 
 
-__all__ = ["Source", "CsvString", "Pandas", "JsonlFile", "JsonlString", "PyDict", "Parquet"]
+__all__ = [
+    "Source",
+    "CsvString",
+    "Pandas",
+    "JsonlFile",
+    "JsonlString",
+    "PyDict",
+    "Parquet",
+]

python/pysrc/kaskada/sources/arrow.py

@@ -280,7 +280,9 @@ async def create(
         if schema is None:
             if csv_string is None:
                 raise ValueError("Must provide schema or csv_string")
-            schema = pa.csv.read_csv(csv_string, parse_options=CsvString._parse_options).schema
+            schema = pa.csv.read_csv(
+                csv_string, parse_options=CsvString._parse_options
+            ).schema
             csv_string.seek(0)
 
         source = CsvString(
@@ -303,7 +305,7 @@ async def add_string(self, csv_string: str | BytesIO) -> None:
         content = pa.csv.read_csv(
             csv_string,
             convert_options=self._convert_options,
-            parse_options=CsvString._parse_options
+            parse_options=CsvString._parse_options,
         )
         for batch in content.to_batches():
             await self._ffi_table.add_pyarrow(batch)
@@ -396,8 +398,7 @@ async def create(
     async def add_file(self, path: str) -> None:
         """Add data to the source."""
         batches = pa.json.read_json(
-            Source._get_absolute_path(path),
-            parse_options=self._parse_options
+            Source._get_absolute_path(path), parse_options=self._parse_options
         )
         for batch in batches.to_batches():
             await self._ffi_table.add_pyarrow(batch)

python/pysrc/kaskada/sources/source.py

@@ -1,6 +1,6 @@
 """Provide the base-class for Kaskada sources."""
-from typing import Literal, Optional
 import os
+from typing import Literal, Optional
 
 import kaskada._ffi as _ffi
 import pyarrow as pa

python/pytests/filter_test.py

@@ -57,10 +57,6 @@ async def test_filter_to_merge_preserves_interpolation(source, golden) -> None:
     predicate = n < 9
     golden.jsonl(
         kd.record(
-            {
-                "n": n,
-                "predicate": predicate,
-                "filter_sum": n.sum().filter(predicate)
-            }
+            {"n": n, "predicate": predicate, "filter_sum": n.sum().filter(predicate)}
         )
     )

python/pytests/golden/parquet_source_test/test_large_parquet_file.jsonl

@@ -0,0 +1,18 @@
+{"_time":"2022-01-04T14:38:31.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-04T14:38:31.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":499.48,"item":"0da9b3fd-2c92-4b87-92b0-5137eaf6ff75"}
+{"_time":"2022-01-05T20:40:03.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-05T20:40:03.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":498.16,"item":"f9cdde05-40f9-48fd-812e-1c3936589184"}
+{"_time":"2022-01-06T04:54:59.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-06T04:54:59.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":491.38,"item":"64cd0de2-02b8-4420-8b8a-57ad4d0b9aa2"}
+{"_time":"2022-01-06T07:14:07.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-06T07:14:07.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":498.32,"item":"3940b205-50a6-4141-ab68-aa0464ae0f3d"}
+{"_time":"2022-01-07T02:32:48.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-07T02:32:48.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":498.4,"item":"3c1f2b17-8bb0-43b9-8a52-e3d8d81fe311"}
+{"_time":"2022-01-08T08:17:28.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-08T08:17:28.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":492.91,"item":"b5a8d6e1-9070-410d-bf44-72754b485faa"}
+{"_time":"2022-01-09T15:12:23.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-09T15:12:23.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":493.92,"item":"9976f04f-3faf-46bd-80f6-1dc102632ec6"}
+{"_time":"2022-01-10T02:11:28.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-10T02:11:28.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":495.04,"item":"ca02d3d3-a309-4b7b-ac12-29fa4a1a8704"}
+{"_time":"2022-01-14T15:06:56.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-14T15:06:56.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":492.81,"item":"3940b205-50a6-4141-ab68-aa0464ae0f3d"}
+{"_time":"2022-01-16T05:08:53.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-16T05:08:53.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":492.92,"item":"5a86942a-5bcc-41f7-9286-937b248caccc"}
+{"_time":"2022-01-20T03:28:47.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-20T03:28:47.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":495.65,"item":"bcfd7a57-f36e-4b37-9b2d-795401f36459"}
+{"_time":"2022-01-21T13:25:25.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-21T13:25:25.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":499.51,"item":"d6789f76-7ac6-415b-a2fa-8b56f80eef74"}
+{"_time":"2022-01-23T06:10:21.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-23T06:10:21.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":491.33,"item":"d988eedb-2f3c-4ad5-82ab-7b1c25754ea0"}
+{"_time":"2022-01-24T16:50:58.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-24T16:50:58.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":494.25,"item":"69718e27-44e6-4cb1-86ff-fc5b5d4c50a1"}
+{"_time":"2022-01-26T20:56:58.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-26T20:56:58.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":492.23,"item":"87c91aeb-dba3-431e-bbda-f65f9164c64d"}
+{"_time":"2022-01-26T22:57:18.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-26T22:57:18.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":491.84,"item":"804488a1-9724-465d-a596-1b6930510640"}
+{"_time":"2022-01-29T08:46:35.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-29T08:46:35.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":490.18,"item":"bc323957-93e4-4aa8-8fc1-c73411e9ca0b"}
+{"_time":"2022-01-29T17:21:29.000000000","_key":"5fec83d4-f5c6-4943-ab05-2b6760330daf","time":"2022-01-29T17:21:29.000000000","user":"5fec83d4-f5c6-4943-ab05-2b6760330daf","amount":491.56,"item":"ad811380-ac9c-4f6a-9015-ba2441abbff0"}

python/pytests/len_test.py

@@ -22,10 +22,4 @@ async def source() -> kd.sources.CsvString:
 
 async def test_len(source, golden) -> None:
     s = source.col("s")
-    golden.jsonl(
-        kd.record(
-            {
-                "len": s.len()
-            }
-        )
-    )
+    golden.jsonl(kd.record({"len": s.len()}))

python/pytests/parquet_source_test.py

@@ -24,3 +24,25 @@ async def test_read_parquet_with_subsort(golden) -> None:
 
     await source.add_file("../testdata/purchases/purchases_part2.parquet")
     golden.jsonl(source)
+
+
+# Verifies that we drain the output and progress channels correctly.
+#
+# When the parquet file contains more rows than
+# (CHANNEL_SIZE / MAX_BATCH_SIZE), the channels previously filled
+# up, causing the sender to block. This test verifies that the
+# channels correctly drain, allowing the sender to continue.
+# See https://github.com/kaskada-ai/kaskada/issues/775
+async def test_large_parquet_file(golden) -> None:
+    source = await kd.sources.Parquet.create(
+        "../testdata/parquet/purchases_100k.parquet",
+        time_column="time",
+        key_column="user",
+    )
+    user = source.col("user")
+    amount = source.col("amount")
+
+    # Add a filter to reduce the output file size while ensuring the entire
+    # file is still processed
+    predicate = user.eq("5fec83d4-f5c6-4943-ab05-2b6760330daf").and_(amount.gt(490))
+    golden.jsonl(source.filter(predicate))