diff --git a/tsdb/head.go b/tsdb/head.go
index 6fe42c8cf2..e6d75e109f 100644
--- a/tsdb/head.go
+++ b/tsdb/head.go
@@ -2156,15 +2156,24 @@ func (h *Head) deleteSeriesByID(refs []chunks.HeadSeriesRef) {
 	)
 
 	for _, ref := range refs {
+		// Delete the reference from the series map.
+		// Copying getByID here to avoid locking and unlocking twice.
 		refShard := int(ref) & (h.series.size - 1)
 		h.series.locks[refShard].Lock()
-
-		// Copying getByID here to avoid locking and unlocking twice.
 		series := h.series.series[refShard][ref]
 		if series == nil {
 			h.series.locks[refShard].Unlock()
 			continue
 		}
+		delete(h.series.series[refShard], series.ref)
+		h.series.locks[refShard].Unlock()
+
+		// Delete the reference from the hash.
+		hash := series.lset.Hash()
+		hashShard := int(hash) & (h.series.size - 1)
+		h.series.locks[hashShard].Lock()
+		h.series.hashes[hashShard].del(hash, series.ref)
+		h.series.locks[hashShard].Unlock()
 
 		if value.IsStaleNaN(series.lastValue) ||
 			(series.lastHistogramValue != nil && value.IsStaleNaN(series.lastHistogramValue.Sum)) ||
@@ -2172,9 +2181,6 @@ func (h *Head) deleteSeriesByID(refs []chunks.HeadSeriesRef) {
 			staleSeriesDeleted++
 		}
 
-		hash := series.lset.Hash()
-		hashShard := int(hash) & (h.series.size - 1)
-
 		chunksRemoved += len(series.mmappedChunks)
 		if series.headChunks != nil {
 			chunksRemoved += series.headChunks.len()
@@ -2182,10 +2188,6 @@ func (h *Head) deleteSeriesByID(refs []chunks.HeadSeriesRef) {
 
 		deleted[storage.SeriesRef(series.ref)] = struct{}{}
 		series.lset.Range(func(l labels.Label) { affected[l] = struct{}{} })
-		h.series.hashes[hashShard].del(hash, series.ref)
-		delete(h.series.series[refShard], series.ref)
-
-		h.series.locks[refShard].Unlock()
 	}
 
 	h.metrics.seriesRemoved.Add(float64(len(deleted)))
diff --git a/tsdb/head_test.go b/tsdb/head_test.go
index 7b8ae0ecbd..6dbe3599e4 100644
--- a/tsdb/head_test.go
+++ b/tsdb/head_test.go
@@ -7246,3 +7246,77 @@ func TestHistogramStalenessConversionMetrics(t *testing.T) {
 		})
 	}
 }
+
+// TestWALReplayRaceWithStaleSeriesCompaction verifies that deleteSeriesByID correctly locks the
+// hash shard (not only the ref shard) when deleting from the hashes map.
+// The race only occurs when Prometheus restarts after having done a stale series compaction because
+// deleteSeriesByID is not used otherwise.
+func TestWALReplayRaceWithStaleSeriesCompaction(t *testing.T) {
+	opts := newTestHeadDefaultOptions(1000, false)
+	// A small stripe size ensures many series share hash shards, increasing
+	// the likelihood that deleteSeriesByID and getOrCreateWithOptionalID
+	// contend on the same shard during WAL replay.
+	opts.StripeSize = 32
+	head, _ := newTestHeadWithOptions(t, compression.None, opts)
+	require.NoError(t, head.Init(0))
+
+	appendSample := func(lbls labels.Labels, ts int64, val float64) {
+		app := head.Appender(context.Background())
+		_, err := app.Append(0, lbls, ts, val)
+		require.NoError(t, err)
+		require.NoError(t, app.Commit())
+	}
+
+	// Step 1: Create a batch of series and make them stale.
+	const numStaleSeries = 500
+	staleLbls := make([]labels.Labels, numStaleSeries)
+	for i := range numStaleSeries {
+		staleLbls[i] = labels.FromStrings("__name__", "stale_metric", "i", strconv.Itoa(i))
+		appendSample(staleLbls[i], 100, float64(i))
+	}
+	for _, lbl := range staleLbls {
+		appendSample(lbl, 200, math.Float64frombits(value.StaleNaN))
+	}
+	require.Equal(t, uint64(numStaleSeries), head.NumStaleSeries())
+
+	// Step 2: Truncate stale series. This removes them from the Head and
+	// writes tombstone records (with Mint=MinInt64, Maxt=MaxInt64) to the WAL.
+	staleRefs := make([]storage.SeriesRef, 0, numStaleSeries)
+	for i := range numStaleSeries {
+		ms := head.series.getByHash(staleLbls[i].Hash(), staleLbls[i])
+		require.NotNil(t, ms)
+		staleRefs = append(staleRefs, storage.SeriesRef(ms.ref))
+	}
+	require.NoError(t, head.truncateStaleSeries(staleRefs, 300))
+	require.Equal(t, uint64(0), head.NumStaleSeries())
+	require.Equal(t, uint64(0), head.NumSeries())
+
+	// Step 3: Add new series AFTER the truncation. In the WAL, these series
+	// records appear after the tombstone records. During replay, the main
+	// goroutine will create these series (via getOrCreateWithOptionalID, which
+	// accesses hashes[hashShard] under locks[hashShard]) concurrently with
+	// the walSubsetProcessor goroutines deleting the stale series (via
+	// deleteSeriesByID, which must also lock the correct hashShard).
+	const numNewSeries = 500
+	for i := range numNewSeries {
+		lbl := labels.FromStrings("__name__", "new_metric", "i", strconv.Itoa(i))
+		appendSample(lbl, 300, float64(i))
+	}
+	require.Equal(t, uint64(numNewSeries), head.NumSeries())
+
+	// Step 4: Close and re-open the Head to trigger WAL replay.
+	// With the buggy locking, the race detector should catch the data race
+	// between the main goroutine (creating series) and worker goroutines
+	// (deleting stale series) during replay.
+	require.NoError(t, head.Close())
+
+	wal, err := wlog.NewSize(nil, nil, filepath.Join(head.opts.ChunkDirRoot, "wal"), 32768, compression.None)
+	require.NoError(t, err)
+	head, err = NewHead(nil, nil, wal, nil, head.opts, nil)
+	require.NoError(t, err)
+	require.NoError(t, head.Init(0)) // Should not cause a race here.
+
+	require.Equal(t, uint64(0), head.NumStaleSeries())
+	require.Equal(t, uint64(numNewSeries), head.NumSeries())
+	require.NoError(t, head.Close())
+}