diff --git a/quickwit/quickwit-indexing/src/actors/merge_planner.rs b/quickwit/quickwit-indexing/src/actors/merge_planner.rs
index 6375186d19d..c00fb1a71a4 100644
--- a/quickwit/quickwit-indexing/src/actors/merge_planner.rs
+++ b/quickwit/quickwit-indexing/src/actors/merge_planner.rs
@@ -18,7 +18,7 @@
 // along with this program. If not, see <http://www.gnu.org/licenses/>.
 
 use std::cmp::Reverse;
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::sync::Arc;
 use std::time::Instant;
 
@@ -31,7 +31,7 @@ use quickwit_proto::indexing::IndexingPipelineId;
 use serde::Serialize;
 use tantivy::Inventory;
 use time::OffsetDateTime;
-use tracing::info;
+use tracing::{info, warn};
 
 use crate::actors::MergeSplitDownloader;
 use crate::merge_policy::MergeOperation;
@@ -42,15 +42,37 @@ use crate::MergePolicy;
 /// The merge planner decides when to start a merge task.
 pub struct MergePlanner {
     pipeline_id: IndexingPipelineId,
+
     /// A young split is a split that has not reached maturity
     /// yet and can be candidate to merge operations.
     partitioned_young_splits: HashMap<u64, Vec<SplitMetadata>>,
+    num_young_splits: usize,
+
+    /// This set contains all of the split ids that we "acknowledged".
+    /// The point of this set is to rapidly dismiss redundant `NewSplit` message.
+    ///
+    /// Complex scenarii that can result in the reemission of
+    /// such messages are described in #3627.
+    ///
+    /// At any given point in time, the set must contains at least:
+    /// - young splits (non-mature)
+    /// - splits that are currently in merge.
+    ///
+    /// It also contain other splits, that have gone through a successful
+    /// merge and have been deleted for instance.
+    ///
+    /// We incrementally build this set, by adding new splits to it.
+    /// When it becomes too large, we entirely rebuild it.
+    known_split_ids: HashSet<String>,
+
     merge_policy: Arc<dyn MergePolicy>,
     merge_split_downloader_mailbox: Mailbox<MergeSplitDownloader>,
+
     /// Inventory of ongoing merge operations. If everything goes well,
     /// a merge operation is dropped after the publish of the merged split.
     /// Used for observability.
     ongoing_merge_operations_inventory: Inventory<MergeOperation>,
+
     /// We use the actor start_time as a way to identify incarnations.
     /// Incarnations are useful to avoid a nasty bug:
     /// Since we recycle mailbox between MergePlanner instantiation,
@@ -75,6 +97,10 @@ impl Actor for MergePlanner {
             .iter()
             .map(|tracked_operation| tracked_operation.as_ref().clone())
             .collect_vec();
+        #[cfg(test)]
+        {
+            self.check_invariants();
+        }
         MergePlannerState {
             ongoing_merge_operations,
         }
@@ -126,6 +152,10 @@ impl Handler<PlanMerge> for MergePlanner {
             // (See comment on `Self::incarnation_start_at`.)
             self.send_merge_ops(ctx).await?;
         }
+        #[cfg(test)]
+        {
+            self.check_invariants();
+        }
         Ok(())
     }
 }
@@ -141,6 +171,10 @@ impl Handler<NewSplits> for MergePlanner {
     ) -> Result<(), ActorExitStatus> {
         self.record_splits(new_splits.new_splits);
         self.send_merge_ops(ctx).await?;
+        #[cfg(test)]
+        {
+            self.check_invariants();
+        }
         Ok(())
     }
 }
@@ -173,6 +207,8 @@ impl MergePlanner {
             .collect();
         let mut merge_planner = MergePlanner {
             pipeline_id,
+            num_young_splits: 0,
+            known_split_ids: Default::default(),
             partitioned_young_splits: Default::default(),
             merge_policy,
             merge_split_downloader_mailbox,
@@ -183,23 +219,110 @@ impl MergePlanner {
         merge_planner
     }
 
+    fn rebuild_known_split_ids(&self) -> HashSet<String> {
+        let mut known_split_ids: HashSet<String> =
+            HashSet::with_capacity(self.num_known_splits_rebuild_threshold());
+        // Add splits that in `partitioned_young_splits`.
+        for young_split_partition in self.partitioned_young_splits.values() {
+            for split in young_split_partition {
+                known_split_ids.insert(split.split_id().to_string());
+            }
+        }
+        let ongoing_merge_operations = self.ongoing_merge_operations_inventory.list();
+        // Add splits that are known as in merge.
+        for merge_op in ongoing_merge_operations {
+            for split in &merge_op.splits {
+                known_split_ids.insert(split.split_id().to_string());
+            }
+        }
+        if known_split_ids.len() * 2 >= self.known_split_ids.len() {
+            warn!(
+                known_split_ids_len_after = known_split_ids.len(),
+                known_split_ids_len_before = self.known_split_ids.len(),
+                "Rebuilding the known split ids set ended up not halving its size. Please report. \
+                 This is likely a bug, please report."
+            );
+        }
+        known_split_ids
+    }
+
+    /// Updates `known_split_ids` and return true if the split was not
+    /// previously known and should be recorded.
+    fn acknownledge_split(&mut self, split_id: &str) -> bool {
+        if self.known_split_ids.contains(split_id) {
+            return false;
+        }
+        // We only rebuild `known_split_ids` when it seems too large.
+        if self.known_split_ids.len() >= self.num_known_splits_rebuild_threshold() {
+            self.known_split_ids = self.rebuild_known_split_ids();
+        }
+        self.known_split_ids.insert(split_id.to_string());
+        true
+    }
+
+    #[cfg(test)]
+    fn check_invariants(&self) {
+        assert_eq!(
+            self.partitioned_young_splits
+                .values()
+                .map(Vec::len)
+                .sum::<usize>(),
+            self.num_young_splits,
+        );
+        let known_split_ids = self.rebuild_known_split_ids();
+        for split_id in known_split_ids {
+            self.known_split_ids.contains(&split_id);
+        }
+        let merge_operation = self.ongoing_merge_operations_inventory.list();
+        let mut young_splits = HashSet::new();
+        for (&partition_id, young_splits_in_partition) in &self.partitioned_young_splits {
+            for split_metadata in young_splits_in_partition {
+                assert_eq!(split_metadata.partition_id, partition_id);
+                young_splits.insert(split_metadata.split_id());
+            }
+        }
+        for merge_op in merge_operation {
+            assert!(!self.known_split_ids.contains(&merge_op.merge_split_id));
+            for split_in_merge in merge_op.splits_as_slice() {
+                assert!(self.known_split_ids.contains(split_in_merge.split_id()));
+            }
+        }
+        assert!(self.known_split_ids.len() <= self.num_known_splits_rebuild_threshold() + 1);
+    }
+
+    /// Whenever the number of known splits exceeds this threshold, we rebuild the `known_split_ids`
+    /// set.
+    ///
+    /// We have this function to return a number that is higher than 2 times the len of
+    /// `known_split_ids` after a rebuild to get amortization.
+    fn num_known_splits_rebuild_threshold(&self) -> usize {
+        // The idea behind this formula is that we expect the max legitimate of splits after a
+        // rebuild to be  `num_young_splits` + `num_splits_merge`.
+        // The capacity of `partioned_young_splits` is a good upper bound for the number of
+        // partition.
+        //
+        // We can expect a maximum of 100 ongoing splits in merge per partition. (We oversize this
+        // because it actually depends on the merge factor.
+        1 + self.num_young_splits + (1 + self.partitioned_young_splits.capacity()) * 20
+    }
+
     fn record_split(&mut self, new_split: SplitMetadata) {
         if new_split.is_mature(OffsetDateTime::now_utc()) {
             return;
         }
-        let splits_for_partition: &mut Vec<SplitMetadata> = self
-            .partitioned_young_splits
-            .entry(new_split.partition_id)
-            .or_default();
         // Due to the recycling of the mailbox of the merge planner, it is possible for
         // a split already in store to be received.
-        if splits_for_partition
-            .iter()
-            .any(|split| split.split_id() == new_split.split_id())
-        {
+        //
+        // See `known_split_ids`.
+        if !self.acknownledge_split(new_split.split_id()) {
             return;
         }
+        let splits_for_partition: &mut Vec<SplitMetadata> = self
+            .partitioned_young_splits
+            .entry(new_split.partition_id)
+            .or_default();
         splits_for_partition.push(new_split);
+        self.num_young_splits += 1;
     }
 
     // Records a list of splits.
@@ -214,7 +337,6 @@ impl MergePlanner {
             self.record_split(new_split);
         }
     }
-
     async fn compute_merge_ops(
         &mut self,
         ctx: &ActorContext<Self>,
@@ -229,6 +351,12 @@ impl MergePlanner {
         }
         self.partitioned_young_splits
             .retain(|_, splits| !splits.is_empty());
+        // We recompute the number of young splits.
+        self.num_young_splits = self
+            .partitioned_young_splits
+            .values()
+            .map(|splits| splits.len())
+            .sum();
         Ok(merge_operations)
     }
 
@@ -677,7 +805,24 @@ mod tests {
             .recv_typed_message::<TrackedObject<MergeOperation>>()
             .await;
         assert!(merge_op.is_some());
+
+        // We make sure that the known splits filtering set filters out splits are currently in
+        // merge.
+        merge_planner_mailbox
+            .ask(NewSplits {
+                new_splits: pre_existing_splits,
+            })
+            .await?;
+
+        let _ = merge_planner_handle.process_pending_and_observe().await;
+
+        let merge_ops =
+            merge_split_downloader_inbox.drain_for_test_typed::<TrackedObject<MergeOperation>>();
+
+        assert!(merge_ops.is_empty());
+
         merge_planner_mailbox.send_message(Command::Quit).await?;
+
         let (exit_status, _last_state) = merge_planner_handle.join().await;
         assert!(matches!(exit_status, ActorExitStatus::Quit));
         let merge_ops =
@@ -812,7 +957,7 @@ mod tests {
 
         // The low capacity of the queue of the merge planner prevents us from sending a Command in
         // the low priority queue. It would take the single slot and prevent the message
-        // sent in the initiliaze method.
+        // sent in the initialize method.
 
         // Instead, we wait for the first merge ops.
         let merge_ops = merge_split_downloader_inbox
@@ -832,4 +977,72 @@ mod tests {
         universe.assert_quit().await;
         Ok(())
     }
+
+    #[tokio::test]
+    async fn test_merge_planner_known_splits_set_size_stays_bounded() -> anyhow::Result<()> {
+        let universe = Universe::with_accelerated_time();
+        let (merge_split_downloader_mailbox, merge_split_downloader_inbox) = universe
+            .spawn_ctx()
+            .create_mailbox("MergeSplitDownloader", QueueCapacity::Unbounded);
+        let index_uid = IndexUid::new("test-index");
+        let pipeline_id = IndexingPipelineId {
+            index_uid: index_uid.clone(),
+            source_id: "test-source".to_string(),
+            node_id: "test-node".to_string(),
+            pipeline_ord: 0,
+        };
+        let merge_policy_config = ConstWriteAmplificationMergePolicyConfig {
+            merge_factor: 2,
+            max_merge_factor: 2,
+            max_merge_ops: 3,
+            ..Default::default()
+        };
+        let indexing_settings = IndexingSettings {
+            merge_policy: MergePolicyConfig::ConstWriteAmplification(merge_policy_config),
+            ..Default::default()
+        };
+        let merge_policy: Arc<dyn MergePolicy> = merge_policy_from_settings(&indexing_settings);
+        let merge_planner = MergePlanner::new(
+            pipeline_id,
+            Vec::new(),
+            merge_policy,
+            merge_split_downloader_mailbox,
+        );
+        let universe = Universe::with_accelerated_time();
+
+        // We spawn our merge planner with this recycled mailbox.
+        let (merge_planner_mailbox, merge_planner_handle) =
+            universe.spawn_builder().spawn(merge_planner);
+
+        for j in 0..100 {
+            for i in 0..10 {
+                merge_planner_mailbox
+                    .ask(NewSplits {
+                        new_splits: vec![split_metadata_for_test(
+                            &index_uid,
+                            &format!("split_{}", j * 10 + i),
+                            0,
+                            1_000_000,
+                            1,
+                        )],
+                    })
+                    .await
+                    .unwrap();
+            }
+            // We drain the merge_ops to make sure merge ops are dropped (as if merges where
+            // successful) and that we are properly testing that the known_splits_set is
+            // bounded.
+            let merge_ops = merge_split_downloader_inbox
+                .drain_for_test_typed::<TrackedObject<MergeOperation>>();
+            assert_eq!(merge_ops.len(), 5);
+        }
+
+        // At this point, our merge has been initialized.
+        merge_planner_mailbox.send_message(Command::Quit).await?;
+        let (exit_status, _last_state) = merge_planner_handle.join().await;
+        assert!(matches!(exit_status, ActorExitStatus::Quit));
+
+        universe.assert_quit().await;
+        Ok(())
+    }
 }