keep knowledge of ongoing merges across merge pipelines

quickwit/Cargo.toml

@@ -90,6 +90,7 @@ binggan = { version = "0.14" }
 bytes = { version = "1", features = ["serde"] }
 bytesize = { version = "1.3.0", features = ["serde"] }
 bytestring = "1.3.0"
+census = "0.4.2"
 chitchat = { git = "https://github.com/quickwit-oss/chitchat.git", rev = "54cbc70" }
 chrono = { version = "0.4", default-features = false, features = [
   "clock",

quickwit/quickwit-common/Cargo.toml

@@ -15,6 +15,7 @@ anyhow = { workspace = true }
 async-speed-limit = { workspace = true }
 async-trait = { workspace = true }
 bytesize = { workspace = true }
+census = { workspace = true }
 coarsetime = { workspace = true }
 dyn-clone = { workspace = true }
 env_logger = { workspace = true }

quickwit/quickwit-common/src/lib.rs

@@ -45,6 +45,7 @@ pub mod temp_dir;
 pub mod test_utils;
 pub mod thread_pool;
 pub mod tower;
+pub mod tracker;
 pub mod type_map;
 pub mod uri;
 

quickwit/quickwit-common/src/tracker.rs

@@ -0,0 +1,108 @@
+use std::mem::MaybeUninit;
+use std::ops::Deref;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::mpsc::{channel, Receiver, Sender};
+use std::sync::{Arc, Mutex};
+
+use census::{Inventory, TrackedObject as InventoredObject};
+
+pub type TrackedObject<T> = InventoredObject<RecordUnacknoledgedDrop<T>>;
+
+#[derive(Clone)]
+pub struct Tracker<T> {
+    inner_inventory: Inventory<RecordUnacknoledgedDrop<T>>,
+    unacknoledged_drop_receiver: Arc<Mutex<Receiver<T>>>,
+    return_channel: Sender<T>,
+}
+
+#[derive(Debug)]
+pub struct RecordUnacknoledgedDrop<T> {
+    // safety: this is always kept initialized except after Self::drop, where we move that
+    // that value away to either send it through the return channel, or drop it manually
+    inner: MaybeUninit<T>,
+    acknoledged: AtomicBool,
+    return_channel: Sender<T>,
+}
+
+impl<T> RecordUnacknoledgedDrop<T> {
+    pub fn acknoledge(&self) {
+        self.acknoledged.store(true, Ordering::Relaxed);
+    }
+
+    pub fn untracked(value: T) -> Self {
+        let (sender, _receiver) = channel();
+        RecordUnacknoledgedDrop {
+            inner: MaybeUninit::new(value),
+            acknoledged: true.into(),
+            return_channel: sender,
+        }
+    }
+}
+
+impl<T> Deref for RecordUnacknoledgedDrop<T> {
+    type Target = T;
+    fn deref(&self) -> &T {
+        unsafe {
+            // safety: see struct definition, this operation is valid except after drop.
+            self.inner.assume_init_ref()
+        }
+    }
+}
+
+impl<T> Drop for RecordUnacknoledgedDrop<T> {
+    fn drop(&mut self) {
+        let item = unsafe {
+            // safety: see struct definition. Additionally, we don't touch to self.inner
+            // after this point so there is no risk of making a 2nd copy and cause a
+            // double-free
+            self.inner.assume_init_read()
+        };
+        if !*self.acknoledged.get_mut() {
+            // if send fails, no one cared about getting that notification, it's fine to
+            // drop item
+            let _ = self.return_channel.send(item);
+        }
+    }
+}
+
+impl<T> Default for Tracker<T> {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl<T> Tracker<T> {
+    pub fn new() -> Self {
+        let (sender, receiver) = channel();
+        Tracker {
+            inner_inventory: Inventory::new(),
+            unacknoledged_drop_receiver: Arc::new(Mutex::new(receiver)),
+            return_channel: sender,
+        }
+    }
+
+    pub fn rebuildable_from_the_void(&self) -> bool {
+        Arc::strong_count(&self.unacknoledged_drop_receiver) == 1 && self.inner_inventory.len() == 0
+    }
+
+    pub fn list_ongoing(&self) -> Vec<TrackedObject<T>> {
+        self.inner_inventory.list()
+    }
+
+    pub fn take_dead(&self) -> Vec<T> {
+        let mut res = Vec::new();
+        let receiver = self.unacknoledged_drop_receiver.lock().unwrap();
+        while let Ok(dead_entry) = receiver.try_recv() {
+            res.push(dead_entry);
+        }
+        res
+    }
+
+    pub fn track(&self, value: T) -> TrackedObject<T> {
+        self.inner_inventory.track(RecordUnacknoledgedDrop {
+            inner: MaybeUninit::new(value),
+            acknoledged: false.into(),
+            return_channel: self.return_channel.clone(),
+        })
+    }
+}

quickwit/quickwit-indexing/failpoints/mod.rs

@@ -36,6 +36,7 @@
 //! Below we test panics at different steps in the indexing pipeline.
 
 use std::path::Path;
+use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Barrier, Mutex};
 use std::time::Duration;
 
@@ -47,15 +48,17 @@ use quickwit_common::split_file;
 use quickwit_common::temp_dir::TempDirectory;
 use quickwit_indexing::actors::MergeExecutor;
 use quickwit_indexing::merge_policy::{MergeOperation, MergeTask};
-use quickwit_indexing::models::MergeScratch;
+use quickwit_indexing::models::{
+    DetachIndexingPipeline, DetachMergePipeline, MergeScratch, SpawnPipeline,
+};
 use quickwit_indexing::{get_tantivy_directory_from_split_bundle, TestSandbox};
 use quickwit_metastore::{
     ListSplitsQuery, ListSplitsRequestExt, MetastoreServiceStreamSplitsExt, SplitMetadata,
     SplitState,
 };
 use quickwit_proto::indexing::MergePipelineId;
 use quickwit_proto::metastore::{ListSplitsRequest, MetastoreService};
-use quickwit_proto::types::{IndexUid, NodeId};
+use quickwit_proto::types::{IndexUid, NodeId, PipelineUid};
 use serde_json::Value as JsonValue;
 use tantivy::Directory;
 
@@ -351,3 +354,122 @@ async fn test_merge_executor_controlled_directory_kill_switch() -> anyhow::Resul
 
     Ok(())
 }
+
+#[tokio::test]
+async fn test_no_duplicate_merge_on_pipeline_restart() -> anyhow::Result<()> {
+    quickwit_common::setup_logging_for_tests();
+    let doc_mapper_yaml = r#"
+        field_mappings:
+          - name: body
+            type: text
+          - name: ts
+            type: datetime
+            fast: true
+        timestamp_field: ts
+        "#;
+    let indexing_setting_yaml = r#"
+        split_num_docs_target: 2500
+        merge_policy:
+          type: "limit_merge"
+          max_merge_ops: 1
+          merge_factor: 4
+          max_merge_factor: 4
+          max_finalize_merge_operations: 1
+    "#;
+    let search_fields = ["body"];
+    let index_id = "test-index-merge-duplication";
+    let mut test_index_builder = TestSandbox::create(
+        index_id,
+        doc_mapper_yaml,
+        indexing_setting_yaml,
+        &search_fields,
+    )
+    .await?;
+
+    //  0: start
+    //  1: 1st merge reached the failpoint
+    // 11: 1st merge failed
+    // 12: 2nd merge reached the failpoint
+    // 22: 2nd merge failed (we don't care about this state)
+    let state = Arc::new(AtomicU32::new(0));
+    let state_clone = state.clone();
+
+    fail::cfg_callback("before-merge-split", move || {
+        use std::sync::atomic::Ordering;
+        state_clone.fetch_add(1, Ordering::Relaxed);
+        std::thread::sleep(std::time::Duration::from_millis(300));
+        state_clone.fetch_add(10, Ordering::Relaxed);
+        panic!("kill merge pipeline");
+    })
+    .unwrap();
+
+    let batch: Vec<JsonValue> =
+        std::iter::repeat_with(|| serde_json::json!({"body ": TEST_TEXT, "ts": 1631072713 }))
+            .take(500)
+            .collect();
+    // this sometime fails because the ingest api isn't aware of the index yet?!
+    tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+    for _ in 0..4 {
+        test_index_builder
+            .add_documents_through_api(batch.clone())
+            .await?;
+    }
+
+    let (indexing_pipeline, merge_pipeline) = test_index_builder
+        .take_indexing_and_merge_pipeline()
+        .await?;
+
+    // stop the pipeline
+    indexing_pipeline.kill().await;
+    merge_pipeline
+        .mailbox()
+        .ask(quickwit_indexing::FinishPendingMergesAndShutdownPipeline)
+        .await?;
+
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    let pipeline_id = test_index_builder
+        .indexing_service()
+        .ask_for_res(SpawnPipeline {
+            index_id: index_id.to_string(),
+            source_config: quickwit_config::SourceConfig::ingest_api_default(),
+            pipeline_uid: PipelineUid::for_test(1u128),
+        })
+        .await?;
+
+    tokio::time::sleep(std::time::Duration::from_millis(200)).await;
+    // we shouldn't have had a 2nd split run yet (the 1st one hasn't panicked just yet)
+    assert_eq!(state.load(Ordering::Relaxed), 1);
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    assert_eq!(state.load(Ordering::Relaxed), 11);
+
+    let merge_pipeline_id = pipeline_id.merge_pipeline_id();
+    let indexing_pipeline = test_index_builder
+        .indexing_service()
+        .ask_for_res(DetachIndexingPipeline { pipeline_id })
+        .await?;
+    let merge_pipeline = test_index_builder
+        .indexing_service()
+        .ask_for_res(DetachMergePipeline {
+            pipeline_id: merge_pipeline_id,
+        })
+        .await?;
+
+    indexing_pipeline.kill().await;
+    merge_pipeline
+        .mailbox()
+        .ask(quickwit_indexing::FinishPendingMergesAndShutdownPipeline)
+        .await?;
+
+    // stoping the merge pipeline makes it recheck for possible dead merge
+    // (alternatively, it does that sooner when rebuilding the known split list)
+    tokio::time::sleep(std::time::Duration::from_millis(200)).await;
+    // timing-wise, we can't have reached 22, but it would be logically correct to get that state
+    assert_eq!(state.load(Ordering::Relaxed), 12);
+
+    let universe = test_index_builder.universe();
+    universe.kill();
+    fail::cfg("before-merge-split", "off").unwrap();
+    universe.quit().await;
+
+    Ok(())
+}

quickwit/quickwit-indexing/src/actors/merge_pipeline.rs

@@ -355,7 +355,8 @@ impl MergePipeline {
             self.params.merge_policy.clone(),
             merge_split_downloader_mailbox,
             self.params.merge_scheduler_service.clone(),
-        );
+        )
+        .await?;
         let (_, merge_planner_handle) = ctx
             .spawn_actor()
             .set_kill_switch(self.kill_switch.clone())