diff --git a/crates/dbsp/src/circuit/dbsp_handle.rs b/crates/dbsp/src/circuit/dbsp_handle.rs
index 7fed588398a..aa762e386cd 100644
--- a/crates/dbsp/src/circuit/dbsp_handle.rs
+++ b/crates/dbsp/src/circuit/dbsp_handle.rs
@@ -3587,10 +3587,19 @@ pub(crate) mod tests {
     fn test_is_compaction_complete() {
         use crate::circuit::GlobalNodeId;
         use crate::circuit::metadata::{MetaItem, SPINE_BATCHES_COUNT};
+        use crate::trace::spine_async::MIN_LEVEL0_MERGE_BATCHES;
         use crate::utils::Tup2;
         use std::time::Duration;
 
-        const BATCHES: usize = 30;
+        // Stay below the number of batches that makes a spine merge on its own,
+        // so that every batch is still waiting when compaction is requested.
+        // Feeding more than that races the background merger, which is free to
+        // reduce a spine to a single batch and leave compaction nothing to do.
+        const BATCHES: i32 = MIN_LEVEL0_MERGE_BATCHES as i32 - 1;
+
+        // Small batches keep every batch in level 0, where
+        // `MIN_LEVEL0_MERGE_BATCHES` applies.  500 records also spread over
+        // both workers, so both spines hold `BATCHES` batches.
         const RECORDS_PER_BATCH: i32 = 500;
 
         let (mut dbsp, input_handle) =
@@ -3602,7 +3611,7 @@ pub(crate) mod tests {
             .unwrap();
 
         // Feed enough batches to give each spine more than one batch to merge.
-        for batch in 0..BATCHES as i32 {
+        for batch in 0..BATCHES {
             let mut tuples: Vec<_> = (0..RECORDS_PER_BATCH)
                 .map(|r| Tup2(batch * RECORDS_PER_BATCH + r, Tup2(r, 1)))
                 .collect();
diff --git a/crates/dbsp/src/trace/spine_async.rs b/crates/dbsp/src/trace/spine_async.rs
index 79b204d7f1c..a9b879c3b50 100644
--- a/crates/dbsp/src/trace/spine_async.rs
+++ b/crates/dbsp/src/trace/spine_async.rs
@@ -112,6 +112,13 @@ static LEVEL_NAMES: [&str; MAX_LEVELS] = [
 /// Configurable via `dev_tweaks.max_level0_batch_size_records`.
 pub(crate) const MAX_LEVEL0_BATCH_SIZE_RECORDS: u16 = 14_999;
 
+/// Number of loose level-0 batches at which the spine starts a background
+/// merge, i.e., `MERGE_COUNTS[0].start()`.
+///
+/// A spine holding fewer level-0 batches than this merges nothing until
+/// something else asks it to, such as a compaction request or memory pressure.
+pub(crate) const MIN_LEVEL0_MERGE_BATCHES: usize = 8;
+
 fn scope_tokio_merger_locals<F>(
     worker_index: usize,
     buffer_cache: Arc<BufferCache>,
@@ -264,7 +271,7 @@ where
         /// The minimum number of batches to merge is key to performance.  The
         /// maximum number seems much less important.
         const MERGE_COUNTS: [RangeInclusive<usize>; MAX_LEVELS] = [
-            8..=128,
+            MIN_LEVEL0_MERGE_BATCHES..=128,
             8..=64,
             3..=64,
             3..=64,