Route DataFusion scan tasks by file affinity

Author: alexanderbianchi

quickwit/Cargo.lock

@@ -39,7 +39,7 @@ datafusion-substrait = "53"
 datafusion-datasource = "53"
 datafusion-physical-plan = "53"
 datafusion-datasource-parquet = "53"
-datafusion-distributed = "1.0"
+datafusion-distributed = { git = "https://github.com/datafusion-contrib/datafusion-distributed.git", rev = "d7de9c3fb99fd7fa99da925b460dc2d4529c04a2" }
 object_store = "0.13"
 
 [dev-dependencies]

quickwit/quickwit-datafusion/Cargo.toml

@@ -15,13 +15,14 @@ async-trait = { workspace = true }
 futures = { workspace = true }
 prost = { workspace = true }            # substrait Plan::decode (runtime only)
 serde_json = { workspace = true }
+siphasher = { workspace = true }
 tracing = { workspace = true }
 url = "2"
 
 arrow = { workspace = true }
 datafusion = "53"
 datafusion-substrait = "53"
-datafusion-distributed = "1.0"
+datafusion-distributed = { git = "https://github.com/datafusion-contrib/datafusion-distributed.git", rev = "d7de9c3fb99fd7fa99da925b460dc2d4529c04a2" }
 object_store = "0.13"
 
 # gRPC surface is opt-in. The extension traits + session builder do not need

quickwit/quickwit-df-core/Cargo.toml

@@ -45,7 +45,7 @@ use datafusion::execution::object_store::ObjectStoreRegistry;
 use datafusion::execution::runtime_env::{RuntimeEnv, RuntimeEnvBuilder};
 use datafusion::prelude::{SessionConfig, SessionContext};
 use datafusion_distributed::{
-    DistributedExt, DistributedPhysicalOptimizerRule, TaskEstimator, WorkerResolver,
+    DistributedExt, SessionStateBuilderExt, TaskEstimator, TaskRoutingContext, WorkerResolver,
 };
 
 use crate::data_source::{
@@ -346,7 +346,7 @@ impl DataFusionSessionBuilder {
             builder = builder
                 .with_distributed_worker_resolver(ArcWorkerResolver(Arc::clone(resolver)))
                 .with_distributed_task_estimator(ArcTaskEstimator(Arc::clone(&self.task_estimator)))
-                .with_physical_optimizer_rule(Arc::new(DistributedPhysicalOptimizerRule));
+                .with_distributed_planner();
         }
 
         Ok(SessionContext::new_with_state(builder.build()))
@@ -386,6 +386,10 @@ impl TaskEstimator for ArcTaskEstimator {
     ) -> Option<Arc<dyn datafusion::physical_plan::ExecutionPlan>> {
         self.0.scale_up_leaf_node(plan, task_count, cfg)
     }
+
+    fn route_tasks(&self, routing_ctx: &TaskRoutingContext<'_>) -> DFResult<Option<Vec<url::Url>>> {
+        self.0.route_tasks(routing_ctx)
+    }
 }
 
 #[cfg(test)]

quickwit/quickwit-df-core/src/session.rs

@@ -19,12 +19,19 @@
 //! count equals the file-group count; for tantivy sources it equals the total
 //! segment count across splits.
 
+use std::hash::{Hash, Hasher};
 use std::sync::Arc;
 
+use datafusion::common::Result as DFResult;
 use datafusion::config::ConfigOptions;
+use datafusion::datasource::physical_plan::FileScanConfig;
 use datafusion::datasource::source::DataSourceExec;
 use datafusion::physical_plan::ExecutionPlan;
-use datafusion_distributed::{PartitionIsolatorExec, TaskEstimation, TaskEstimator};
+use datafusion_distributed::{
+    PartitionIsolatorExec, TaskEstimation, TaskEstimator, TaskRoutingContext,
+};
+use siphasher::sip::SipHasher;
+use url::Url;
 
 /// Estimates distributed task count from the partition count of a
 /// `DataSourceExec` leaf. Returns `None` for any other plan node, letting the
@@ -61,4 +68,210 @@ impl TaskEstimator for DataSourceExecPartitionEstimator {
             task_count,
         )))
     }
+
+    fn route_tasks(&self, routing_ctx: &TaskRoutingContext<'_>) -> DFResult<Option<Vec<Url>>> {
+        if routing_ctx.available_urls.is_empty() {
+            return Ok(None);
+        }
+        let Some(file_scan) = single_file_scan_config(routing_ctx.plan) else {
+            return Ok(None);
+        };
+        let task_affinities = file_scan_task_affinities(file_scan, routing_ctx.task_count);
+        if task_affinities.len() != routing_ctx.task_count {
+            return Ok(None);
+        }
+        Ok(Some(route_by_rendezvous(
+            &task_affinities,
+            routing_ctx.available_urls,
+        )))
+    }
+}
+
+#[derive(Debug, Clone, Eq, PartialEq)]
+struct TaskAffinity {
+    task_index: usize,
+    affinity_key: String,
+}
+
+fn single_file_scan_config(plan: &Arc<dyn ExecutionPlan>) -> Option<&FileScanConfig> {
+    let mut file_scans = Vec::new();
+    collect_file_scan_configs(plan, &mut file_scans);
+    match file_scans.as_slice() {
+        [file_scan] => Some(*file_scan),
+        _ => None,
+    }
+}
+
+fn collect_file_scan_configs<'a>(
+    plan: &'a Arc<dyn ExecutionPlan>,
+    file_scans: &mut Vec<&'a FileScanConfig>,
+) {
+    if let Some(data_source_exec) = plan.as_any().downcast_ref::<DataSourceExec>()
+        && let Some(file_scan) = data_source_exec
+            .data_source()
+            .as_any()
+            .downcast_ref::<FileScanConfig>()
+    {
+        file_scans.push(file_scan);
+    }
+
+    for child in plan.children() {
+        collect_file_scan_configs(child, file_scans);
+    }
+}
+
+fn file_scan_task_affinities(file_scan: &FileScanConfig, task_count: usize) -> Vec<TaskAffinity> {
+    partition_groups(file_scan.file_groups.len(), task_count)
+        .into_iter()
+        .enumerate()
+        .map(|(task_index, partition_group)| {
+            let mut file_keys = Vec::new();
+            for partition_idx in partition_group {
+                if let Some(file_group) = file_scan.file_groups.get(partition_idx) {
+                    for file in file_group.files() {
+                        file_keys.push(format!(
+                            "{}{}",
+                            file_scan.object_store_url.as_str(),
+                            file.object_meta.location.as_ref()
+                        ));
+                    }
+                }
+            }
+            file_keys.sort_unstable();
+            let affinity_key = if file_keys.is_empty() {
+                format!("empty-task-{task_index}")
+            } else {
+                file_keys.join("\n")
+            };
+            TaskAffinity {
+                task_index,
+                affinity_key,
+            }
+        })
+        .collect()
+}
+
+fn partition_groups(input_partitions: usize, task_count: usize) -> Vec<Vec<usize>> {
+    if task_count == 0 {
+        return Vec::new();
+    }
+    let q = input_partitions / task_count;
+    let r = input_partitions % task_count;
+    let mut off = 0;
+    (0..task_count)
+        .map(|i| q + usize::from(i < r))
+        .map(|n| {
+            let result = (off..(off + n)).collect();
+            off += n;
+            result
+        })
+        .collect()
+}
+
+fn route_by_rendezvous(task_affinities: &[TaskAffinity], available_urls: &[Url]) -> Vec<Url> {
+    let mut routes: Vec<Option<Url>> = vec![None; task_affinities.len()];
+    let mut task_order: Vec<usize> = (0..task_affinities.len()).collect();
+    task_order.sort_unstable_by(|&left, &right| {
+        task_affinities[left]
+            .affinity_key
+            .cmp(&task_affinities[right].affinity_key)
+            .then_with(|| {
+                task_affinities[left]
+                    .task_index
+                    .cmp(&task_affinities[right].task_index)
+            })
+    });
+
+    for task_idx in task_order {
+        let task = &task_affinities[task_idx];
+        let mut candidates: Vec<usize> = (0..available_urls.len()).collect();
+        candidates.sort_unstable_by(|&left, &right| {
+            let left_score = rendezvous_affinity(&available_urls[left], &task.affinity_key);
+            let right_score = rendezvous_affinity(&available_urls[right], &task.affinity_key);
+            right_score.cmp(&left_score).then_with(|| {
+                available_urls[left]
+                    .as_str()
+                    .cmp(available_urls[right].as_str())
+            })
+        });
+        let chosen_idx = candidates[0];
+        routes[task.task_index] = Some(available_urls[chosen_idx].clone());
+    }
+
+    routes
+        .into_iter()
+        .map(|route| route.expect("every task should be routed"))
+        .collect()
+}
+
+fn rendezvous_affinity(url: &Url, key: &str) -> u64 {
+    let mut state = SipHasher::new();
+    key.hash(&mut state);
+    url.as_str().hash(&mut state);
+    state.finish()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn url(value: &str) -> Url {
+        Url::parse(value).unwrap()
+    }
+
+    fn task(task_index: usize, affinity_key: &str) -> TaskAffinity {
+        TaskAffinity {
+            task_index,
+            affinity_key: affinity_key.to_string(),
+        }
+    }
+
+    #[test]
+    fn partition_groups_match_partition_isolator_layout() {
+        assert_eq!(partition_groups(2, 1), vec![vec![0, 1]]);
+        assert_eq!(partition_groups(6, 2), vec![vec![0, 1, 2], vec![3, 4, 5]]);
+        assert_eq!(
+            partition_groups(10, 3),
+            vec![vec![0, 1, 2, 3], vec![4, 5, 6], vec![7, 8, 9]]
+        );
+    }
+
+    #[test]
+    fn rendezvous_routing_is_stable_across_url_order() {
+        let tasks = vec![
+            task(0, "s3://bucket/metrics-0.parquet"),
+            task(1, "s3://bucket/metrics-1.parquet"),
+            task(2, "s3://bucket/metrics-2.parquet"),
+            task(3, "s3://bucket/metrics-3.parquet"),
+            task(4, "s3://bucket/metrics-4.parquet"),
+            task(5, "s3://bucket/metrics-5.parquet"),
+        ];
+        let urls = vec![
+            url("http://node-0:7281"),
+            url("http://node-1:7281"),
+            url("http://node-2:7281"),
+        ];
+        let mut reversed_urls = urls.clone();
+        reversed_urls.reverse();
+
+        let routes = route_by_rendezvous(&tasks, &urls);
+        let reversed_routes = route_by_rendezvous(&tasks, &reversed_urls);
+
+        assert_eq!(routes, reversed_routes);
+    }
+
+    #[test]
+    fn rendezvous_routing_uses_best_affinity_node_per_task() {
+        let tasks = vec![task(0, "s3://bucket/metrics.parquet")];
+        let urls = vec![url("http://node-0:7281"), url("http://node-1:7281")];
+
+        let routes = route_by_rendezvous(&tasks, &urls);
+        let expected = urls
+            .iter()
+            .max_by_key(|url| rendezvous_affinity(url, "s3://bucket/metrics.parquet"))
+            .unwrap()
+            .clone();
+
+        assert_eq!(routes, vec![expected]);
+    }
 }