feat: allowing injecting custom join_set tracer to avoid dependency on tracing crate

Author: geoffreyclaude

Diff for: Cargo.lock

@@ -129,7 +129,6 @@ Optional features:
 - `backtrace`: include backtrace information in error messages
 - `pyarrow`: conversions between PyArrow and DataFusion types
 - `serde`: enable arrow-schema's `serde` feature
-- `tracing`: propagates the current span across thread boundaries
 
 [apache avro]: https://avro.apache.org/
 [apache parquet]: https://parquet.apache.org/

Diff for: README.md

@@ -61,7 +61,7 @@ async-trait = { workspace = true }
 bytes = { workspace = true }
 dashmap = { workspace = true }
 # note only use main datafusion crate for examples
-datafusion = { workspace = true, default-features = true, features = ["avro", "tracing"] }
+datafusion = { workspace = true, default-features = true, features = ["avro"] }
 datafusion-proto = { workspace = true }
 env_logger = { workspace = true }
 futures = { workspace = true }

Diff for: datafusion-examples/Cargo.toml

@@ -15,113 +15,83 @@
 // specific language governing permissions and limitations
 // under the License.
 
-//! This example demonstrates the trace feature in DataFusion’s runtime.
-//! When the `tracing` feature is enabled, spawned tasks in DataFusion (such as those
-//! created during repartitioning or when reading Parquet files) are instrumented
-//! with the current tracing span, allowing to propagate any existing tracing context.
-//!
-//! In this example we create a session configured to use multiple partitions,
-//! register a Parquet table (based on the `alltypes_tiny_pages_plain.parquet` file),
-//! and run a query that should trigger parallel execution on multiple threads.
-//! We wrap the entire query execution within a custom span and log messages.
-//! By inspecting the tracing output, we should see that the tasks spawned
-//! internally inherit the span context.
+//! This example demonstrates the tracing injection feature for the DataFusion runtime.
+//! Tasks spawned on new threads behave differently depending on whether a tracer is
+//! injected. The log output clearly distinguishes the two cases.
 
-use arrow::util::pretty::pretty_format_batches;
-use datafusion::arrow::record_batch::RecordBatch;
+use std::any::Any;
 use datafusion::datasource::file_format::parquet::ParquetFormat;
 use datafusion::datasource::listing::ListingOptions;
 use datafusion::error::Result;
 use datafusion::prelude::*;
 use datafusion::test_util::parquet_test_data;
 use std::sync::Arc;
-use tracing::{info, instrument, Level};
+use futures::future::BoxFuture;
+use futures::FutureExt;
+use tracing::{info, instrument, Instrument, Level, Span};
+use datafusion::common::runtime::set_join_set_tracer;
 
 #[tokio::main]
 async fn main() -> Result<()> {
-    // Initialize a tracing subscriber that prints to stdout.
+    // Initialize tracing subscriber with thread info.
     tracing_subscriber::fmt()
         .with_thread_ids(true)
         .with_thread_names(true)
         .with_max_level(Level::DEBUG)
         .init();
 
-    log::info!("Starting example, this log is not captured by tracing");
+    // Run query WITHOUT tracer injection.
+    info!("***** RUNNING WITHOUT INJECTED TRACER *****");
+    run_instrumented_query().await?;
+    info!("***** WITHOUT tracer: `tokio-runtime-worker` tasks did NOT inherit the `run_instrumented_query` span *****");
 
-    // execute the query within a tracing span
-    let result = run_instrumented_query().await;
+    // Inject custom tracer so tasks run in the current span.
+    info!("Injecting custom tracer...");
+    set_join_set_tracer(instrument_future, instrument_block)
+        .expect("Failed to set tracer");
 
-    info!(
-        "Finished example. Check the logs above for tracing span details showing \
-that tasks were spawned within the 'run_instrumented_query' span on different threads."
-    );
+    // Run query WITH tracer injection.
+    info!("***** RUNNING WITH INJECTED TRACER *****");
+    run_instrumented_query().await?;
+    info!("***** WITH tracer: `tokio-runtime-worker` tasks DID inherit the `run_instrumented_query` span *****");
 
-    result
+    Ok(())
+}
+
+/// Instruments a boxed future to run in the current span.
+fn instrument_future(
+    fut: BoxFuture<'static, Box<dyn Any + Send>>
+) -> BoxFuture<'static, Box<dyn Any + Send>> {
+    fut.in_current_span().boxed()
+}
+
+/// Instruments a boxed blocking closure to execute within the current span.
+fn instrument_block(
+    f: Box<dyn FnOnce() -> Box<dyn Any + Send> + Send>
+) -> Box<dyn FnOnce() -> Box<dyn Any + Send> + Send> {
+    let span = Span::current();
+    Box::new(move || span.in_scope(|| f()))
 }
 
 #[instrument(level = "info")]
 async fn run_instrumented_query() -> Result<()> {
-    info!("Starting query execution within the custom tracing span");
+    info!("Starting query execution");
 
-    // The default session will set the number of partitions to `std::thread::available_parallelism()`.
     let ctx = SessionContext::new();
-
-    // Get the path to the test parquet data.
     let test_data = parquet_test_data();
-    // Build listing options that pick up only the "alltypes_tiny_pages_plain.parquet" file.
     let file_format = ParquetFormat::default().with_enable_pruning(true);
     let listing_options = ListingOptions::new(Arc::new(file_format))
         .with_file_extension("alltypes_tiny_pages_plain.parquet");
 
-    info!("Registering Parquet table 'alltypes' from {test_data} in {listing_options:?}");
-
-    // Register a listing table using an absolute URL.
     let table_path = format!("file://{test_data}/");
-    ctx.register_listing_table(
-        "alltypes",
-        &table_path,
-        listing_options.clone(),
-        None,
-        None,
-    )
-    .await
-    .expect("register_listing_table failed");
-
-    info!("Registered Parquet table 'alltypes' from {table_path}");
-
-    // Run a query that will trigger parallel execution on multiple threads.
-    let sql = "SELECT COUNT(*), bool_col, date_string_col, string_col
-                    FROM (
-                      SELECT bool_col, date_string_col, string_col FROM alltypes
-                      UNION ALL
-                      SELECT bool_col, date_string_col, string_col FROM alltypes
-                    ) AS t
-                    GROUP BY bool_col, date_string_col, string_col
-                    ORDER BY 1,2,3,4 DESC
-                    LIMIT 5;";
-    info!(%sql, "Executing SQL query");
-    let df = ctx.sql(sql).await?;
-
-    let results: Vec<RecordBatch> = df.collect().await?;
-    info!("Query execution complete");
-
-    // Print out the results and tracing output.
-    datafusion::common::assert_batches_eq!(
-        [
-            "+----------+----------+-----------------+------------+",
-            "| count(*) | bool_col | date_string_col | string_col |",
-            "+----------+----------+-----------------+------------+",
-            "| 2        | false    | 01/01/09        | 9          |",
-            "| 2        | false    | 01/01/09        | 7          |",
-            "| 2        | false    | 01/01/09        | 5          |",
-            "| 2        | false    | 01/01/09        | 3          |",
-            "| 2        | false    | 01/01/09        | 1          |",
-            "+----------+----------+-----------------+------------+",
-        ],
-        &results
-    );
-
-    info!("Query results:\n{}", pretty_format_batches(&results)?);
-
+    info!("Registering table 'alltypes' from {}", table_path);
+    ctx.register_listing_table("alltypes", &table_path, listing_options, None, None)
+        .await
+        .expect("Failed to register table");
+
+    let sql = "SELECT COUNT(*), string_col FROM alltypes GROUP BY string_col";
+    info!(sql, "Executing SQL query");
+    let result = ctx.sql(sql).await?.collect().await?;
+    info!("Query complete: {} batches returned", result.len());
     Ok(())
-}
+}

Diff for: datafusion-examples/examples/tracing.rs

@@ -34,17 +34,13 @@ all-features = true
 [lints]
 workspace = true
 
-[features]
-tracing = ["dep:tracing", "dep:tracing-futures"]
-
 [lib]
 name = "datafusion_common_runtime"
 
 [dependencies]
+futures = { workspace = true }
 log = { workspace = true }
 tokio = { workspace = true }
-tracing = { version = "0.1", optional = true }
-tracing-futures = { version = "0.2", optional = true }
 
 [dev-dependencies]
 tokio = { version = "1.36", features = ["rt", "rt-multi-thread", "time"] }

Diff for: datafusion/common-runtime/Cargo.toml

@@ -53,15 +53,15 @@ impl<R: 'static> SpawnedTask<R> {
     }
 
     /// Joins the task, returning the result of join (`Result<R, JoinError>`).
-    pub async fn join(mut self) -> Result<R, JoinError> {
+    pub async fn join(mut self) -> Result<R, JoinError> where R: Send {
         self.inner
             .join_next()
             .await
             .expect("`SpawnedTask` instance always contains exactly 1 task")
     }
 
     /// Joins the task and unwinds the panic if it happens.
-    pub async fn join_unwind(self) -> Result<R, JoinError> {
+    pub async fn join_unwind(self) -> Result<R, JoinError> where R: Send {
         self.join().await.map_err(|e| {
             // `JoinError` can be caused either by panic or cancellation. We have to handle panics:
             if e.is_panic() {