Merge branch 'main' into nv/exec-stats

pkg/querier/bucket_cache.go

@@ -26,8 +26,9 @@ const bucketCacheCellSteps = 60
 type BucketCache interface {
 	// cached portion + list of gaps to fetch
 	GetMissRanges(ctx context.Context, orgID valuer.UUID, q qbtypes.Query, step qbtypes.Step) (cached *qbtypes.Result, missing []*qbtypes.TimeRange)
-	// store fresh buckets for future hits
-	Put(ctx context.Context, orgID valuer.UUID, q qbtypes.Query, step qbtypes.Step, fresh *qbtypes.Result)
+	// store fresh buckets for future hits; only the ranges in the map were
+	// queried for fresh, the rest of it came from the cache
+	Put(ctx context.Context, orgID valuer.UUID, q qbtypes.Query, step qbtypes.Step, fresh *qbtypes.Result, execStatsForQueriedTimeRanges map[qbtypes.TimeRange]qbtypes.ExecStats)
 }
 
 // bucketCache implements the BucketCache interface.
@@ -116,7 +117,7 @@ func (bc *bucketCache) GetMissRanges(
 	}
 
 	// Merge buckets into a single result
-	mergedResult := bc.mergeBuckets(ctx, relevantBuckets, cachedWarnings)
+	mergedResult := bc.mergeBuckets(ctx, relevantBuckets, cachedWarnings, startMs, endMs)
 
 	// Filter the merged result to only include values within the requested time range
 	mergedResult = bc.filterResultToTimeRange(mergedResult, q, startMs, endMs, stepMs)
@@ -125,7 +126,7 @@ func (bc *bucketCache) GetMissRanges(
 }
 
 // Put stores fresh query results in the cache.
-func (bc *bucketCache) Put(ctx context.Context, orgID valuer.UUID, q qbtypes.Query, step qbtypes.Step, fresh *qbtypes.Result) {
+func (bc *bucketCache) Put(ctx context.Context, orgID valuer.UUID, q qbtypes.Query, step qbtypes.Step, fresh *qbtypes.Result, execStatsForQueriedTimeRanges map[qbtypes.TimeRange]qbtypes.ExecStats) {
 	// Get query window
 	startMs, endMs := q.Window()
 
@@ -188,7 +189,7 @@ func (bc *bucketCache) Put(ctx context.Context, orgID valuer.UUID, q qbtypes.Que
 
 	// Convert trimmed result to buckets, one per grid cell it spans
 	cellWidthMs := bucketCacheCellSteps * stepMs
-	freshBuckets := bc.resultToBuckets(ctx, trimmedResult, cachableStartMs, cachableEndMs, cellWidthMs)
+	freshBuckets := bc.resultToBuckets(ctx, trimmedResult, cachableStartMs, cachableEndMs, cellWidthMs, execStatsForQueriedTimeRanges)
 
 	cellKeys := make([]string, len(freshBuckets))
 	cellEntries := make([]cachetypes.Cacheable, len(freshBuckets))
@@ -210,7 +211,7 @@ func (bc *bucketCache) Put(ctx context.Context, orgID valuer.UUID, q qbtypes.Que
 			cellEntry = cellEntries[i].(*qbtypes.CachedData)
 		}
 
-		cellBuckets, changed := bc.mergeFreshBucketIntoCell(ctx, cellEntry.Buckets, freshBucket)
+		cellBuckets, changed := bc.mergeFreshBucketIntoCell(ctx, cellEntry.Buckets, freshBucket, execStatsForQueriedTimeRanges)
 		if !changed {
 			continue
 		}
@@ -463,7 +464,7 @@ func (bc *bucketCache) filterRelevantBuckets(buckets []*qbtypes.CachedBucket, st
 }
 
 // mergeBuckets combines multiple cached buckets into a single result.
-func (bc *bucketCache) mergeBuckets(ctx context.Context, buckets []*qbtypes.CachedBucket, warnings []string) *qbtypes.Result {
+func (bc *bucketCache) mergeBuckets(ctx context.Context, buckets []*qbtypes.CachedBucket, warnings []string, startMs, endMs uint64) *qbtypes.Result {
 	if len(buckets) == 0 {
 		return &qbtypes.Result{}
 	}
@@ -471,12 +472,13 @@ func (bc *bucketCache) mergeBuckets(ctx context.Context, buckets []*qbtypes.Cach
 	// All buckets should have the same type
 	resultType := buckets[0].Type
 
-	// Aggregate stats
+	// A bucket can reach past the window; only the part inside it is served
 	var totalStats qbtypes.ExecStats
 	for _, bucket := range buckets {
-		totalStats.RowsScanned += bucket.Stats.RowsScanned
-		totalStats.BytesScanned += bucket.Stats.BytesScanned
-		totalStats.DurationMS += bucket.Stats.DurationMS
+		servedStats := bucket.ProrateStatsToWindow(startMs, endMs)
+		totalStats.RowsScanned += servedStats.RowsScanned
+		totalStats.BytesScanned += servedStats.BytesScanned
+		totalStats.DurationMS += servedStats.DurationMS
 	}
 
 	// Merge values based on type
@@ -674,7 +676,7 @@ func (bc *bucketCache) isEmptyResult(result *qbtypes.Result) (isEmpty bool, isFi
 }
 
 // resultToBuckets converts a query result into time-based buckets.
-func (bc *bucketCache) resultToBuckets(ctx context.Context, result *qbtypes.Result, startMs, endMs, cellWidthMs uint64) []*qbtypes.CachedBucket {
+func (bc *bucketCache) resultToBuckets(ctx context.Context, result *qbtypes.Result, startMs, endMs, cellWidthMs uint64, execStatsForQueriedTimeRanges map[qbtypes.TimeRange]qbtypes.ExecStats) []*qbtypes.CachedBucket {
 	// Check if result is empty
 	isEmpty, isFiltered := bc.isEmptyResult(result)
 
@@ -686,7 +688,7 @@ func (bc *bucketCache) resultToBuckets(ctx context.Context, result *qbtypes.Resu
 	}
 
 	// Empty filtered results still become buckets, so the range is not re-queried
-	perCellBuckets, err := result.SplitIntoCacheCellBuckets(startMs, endMs, cellWidthMs)
+	perCellBuckets, err := result.SplitIntoCacheCellBuckets(startMs, endMs, cellWidthMs, execStatsForQueriedTimeRanges)
 	if err != nil {
 		bc.logger.ErrorContext(ctx, "failed to convert result to buckets", errors.Attr(err))
 		return nil
@@ -698,7 +700,7 @@ func (bc *bucketCache) resultToBuckets(ctx context.Context, result *qbtypes.Resu
 // and reports whether the cell changed: a bucket already covering the fresh
 // one leaves it as is, buckets inside the fresh one are replaced, buckets
 // overlapping or edge to edge with it are unioned, and the rest are kept apart.
-func (bc *bucketCache) mergeFreshBucketIntoCell(ctx context.Context, cellBuckets []*qbtypes.CachedBucket, freshBucket *qbtypes.CachedBucket) ([]*qbtypes.CachedBucket, bool) {
+func (bc *bucketCache) mergeFreshBucketIntoCell(ctx context.Context, cellBuckets []*qbtypes.CachedBucket, freshBucket *qbtypes.CachedBucket, execStatsForQueriedTimeRanges map[qbtypes.TimeRange]qbtypes.ExecStats) ([]*qbtypes.CachedBucket, bool) {
 	keptBuckets := make([]*qbtypes.CachedBucket, 0, len(cellBuckets)+1)
 	touchingBuckets := []*qbtypes.CachedBucket{}
 	for _, cellBucket := range cellBuckets {
@@ -711,7 +713,7 @@ func (bc *bucketCache) mergeFreshBucketIntoCell(ctx context.Context, cellBuckets
 			keptBuckets = append(keptBuckets, cellBucket)
 		}
 	}
-	keptBuckets = append(keptBuckets, bc.unionBuckets(ctx, freshBucket, touchingBuckets))
+	keptBuckets = append(keptBuckets, bc.unionBuckets(ctx, freshBucket, touchingBuckets, execStatsForQueriedTimeRanges))
 
 	slices.SortFunc(keptBuckets, func(a, b *qbtypes.CachedBucket) int {
 		return cmp.Compare(a.StartMs, b.StartMs)
@@ -722,7 +724,16 @@ func (bc *bucketCache) mergeFreshBucketIntoCell(ctx context.Context, cellBuckets
 // unionBuckets widens the fresh bucket by the buckets it touches. Buckets
 // inside it are replaced outright, so only a genuine partial overlap decodes
 // anything.
-func (bc *bucketCache) unionBuckets(ctx context.Context, freshBucket *qbtypes.CachedBucket, touchingBuckets []*qbtypes.CachedBucket) *qbtypes.CachedBucket {
+func (bc *bucketCache) unionBuckets(ctx context.Context, freshBucket *qbtypes.CachedBucket, touchingBuckets []*qbtypes.CachedBucket, execStatsForQueriedTimeRanges map[qbtypes.TimeRange]qbtypes.ExecStats) *qbtypes.CachedBucket {
+	// Data queried again takes the new query's cost, which the fresh bucket
+	// carries; a touching bucket keeps its cost only for what was not queried again
+	for _, touchingBucket := range touchingBuckets {
+		keptStats := touchingBucket.ProrateStatsToUnqueried(execStatsForQueriedTimeRanges)
+		freshBucket.Stats.RowsScanned += keptStats.RowsScanned
+		freshBucket.Stats.BytesScanned += keptStats.BytesScanned
+		freshBucket.Stats.DurationMS += keptStats.DurationMS
+	}
+
 	bucketsToMerge := []*qbtypes.CachedBucket{freshBucket}
 	for _, touchingBucket := range touchingBuckets {
 		if !freshBucket.Contains(touchingBucket) {
@@ -743,14 +754,11 @@ func (bc *bucketCache) unionBuckets(ctx context.Context, freshBucket *qbtypes.Ca
 		EndMs:   freshBucket.EndMs,
 		Type:    freshBucket.Type,
 		Value:   value,
-		Stats:   qbtypes.ExecStats{StepIntervals: freshBucket.Stats.StepIntervals},
+		Stats:   freshBucket.Stats,
 	}
 	for _, mergedBucket := range bucketsToMerge {
 		unionBucket.StartMs = min(unionBucket.StartMs, mergedBucket.StartMs)
 		unionBucket.EndMs = max(unionBucket.EndMs, mergedBucket.EndMs)
-		unionBucket.Stats.RowsScanned += mergedBucket.Stats.RowsScanned
-		unionBucket.Stats.BytesScanned += mergedBucket.Stats.BytesScanned
-		unionBucket.Stats.DurationMS += mergedBucket.Stats.DurationMS
 	}
 	return unionBucket
 }

pkg/querier/bucket_cache_bench_test.go

@@ -30,7 +30,7 @@ func BenchmarkBucketCache_GetMissRanges(b *testing.B) {
 			endMs:       uint64((i + 1) * 10000),
 		}
 		result := createBenchmarkResult(query.startMs, query.endMs, 1000)
-		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: query.startMs, To: query.endMs}: result.Stats})
 	}
 
 	// Create test queries with varying cache hit patterns
@@ -121,7 +121,7 @@ func BenchmarkBucketCache_Put(b *testing.B) {
 
 			for i := 0; i < b.N; i++ {
 				for j := 0; j < tc.numQueries; j++ {
-					bc.Put(ctx, orgID, queries[j], qbtypes.Step{Duration: 1000 * time.Millisecond}, results[j])
+					bc.Put(ctx, orgID, queries[j], qbtypes.Step{Duration: 1000 * time.Millisecond}, results[j], map[qbtypes.TimeRange]qbtypes.ExecStats{{From: queries[j].startMs, To: queries[j].endMs}: results[j].Stats})
 				}
 			}
 		})
@@ -259,7 +259,7 @@ func BenchmarkBucketCache_ConcurrentOperations(b *testing.B) {
 			endMs:       uint64((i + 1) * 10000),
 		}
 		result := createBenchmarkResult(query.startMs, query.endMs, 1000)
-		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: query.startMs, To: query.endMs}: result.Stats})
 	}
 
 	b.ResetTimer()
@@ -284,7 +284,7 @@ func BenchmarkBucketCache_ConcurrentOperations(b *testing.B) {
 					endMs:       uint64((i + 1) * 10000),
 				}
 				result := createBenchmarkResult(query.startMs, query.endMs, 1000)
-				bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+				bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: query.startMs, To: query.endMs}: result.Stats})
 			case 2: // Partial read
 				query := &mockQuery{
 					fingerprint: fmt.Sprintf("concurrent-query-%d", i%100),

pkg/querier/bucket_cache_step_test.go

@@ -59,7 +59,7 @@ func TestBucketCacheStepAlignment(t *testing.T) {
 	}
 
 	// Put result in cache
-	bc.Put(ctx, orgID, query, step, result)
+	bc.Put(ctx, orgID, query, step, result, nil)
 
 	// Get cached data
 	cached, missing := bc.GetMissRanges(ctx, orgID, query, step)
@@ -109,7 +109,7 @@ func TestBucketCacheNoStepInterval(t *testing.T) {
 	}
 
 	// Should cache the entire range when step is 0
-	bc.Put(ctx, orgID, query, step, result)
+	bc.Put(ctx, orgID, query, step, result, nil)
 
 	cached, missing := bc.GetMissRanges(ctx, orgID, query, step)
 	assert.NotNil(t, cached)

pkg/querier/bucket_cache_test.go

@@ -156,7 +156,7 @@ func TestBucketCache_Put_And_Get(t *testing.T) {
 	}
 
 	// Store in cache
-	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: query.startMs, To: query.endMs}: result.Stats})
 
 	// Wait a bit for cache to be written
 	time.Sleep(10 * time.Millisecond)
@@ -189,7 +189,7 @@ func TestBucketCache_PartialHit(t *testing.T) {
 		Type:  qbtypes.RequestTypeTimeSeries,
 		Value: createTestTimeSeries("A", 1000, 3000, 1000),
 	}
-	bc.Put(context.Background(), valuer.UUID{}, query1, qbtypes.Step{Duration: 1000 * time.Millisecond}, result1)
+	bc.Put(context.Background(), valuer.UUID{}, query1, qbtypes.Step{Duration: 1000 * time.Millisecond}, result1, nil)
 
 	// Wait for cache write
 	time.Sleep(10 * time.Millisecond)
@@ -225,7 +225,7 @@ func TestBucketCache_MultipleBuckets(t *testing.T) {
 	bc.Put(context.Background(), valuer.UUID{}, query1, qbtypes.Step{Duration: 100 * time.Millisecond}, &qbtypes.Result{
 		Type:  qbtypes.RequestTypeTimeSeries,
 		Value: createTestTimeSeries("A", 1000, 2000, 100),
-	})
+	}, nil)
 
 	query2 := &mockQuery{
 		fingerprint: "test-query",
@@ -235,7 +235,7 @@ func TestBucketCache_MultipleBuckets(t *testing.T) {
 	bc.Put(context.Background(), valuer.UUID{}, query2, qbtypes.Step{Duration: 100 * time.Millisecond}, &qbtypes.Result{
 		Type:  qbtypes.RequestTypeTimeSeries,
 		Value: createTestTimeSeries("A", 3000, 4000, 100),
-	})
+	}, nil)
 
 	// Wait for cache writes
 	time.Sleep(10 * time.Millisecond)
@@ -280,7 +280,7 @@ func TestBucketCache_FluxInterval(t *testing.T) {
 	}
 
 	// This should not be cached due to flux interval
-	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 
 	// Wait a bit
 	time.Sleep(10 * time.Millisecond)
@@ -358,7 +358,7 @@ func TestBucketCache_MergeTimeSeriesResults(t *testing.T) {
 				{Series: []*qbtypes.TimeSeries{series1}},
 			},
 		},
-	})
+	}, nil)
 
 	// Cache second part
 	query2 := &mockQuery{
@@ -374,7 +374,7 @@ func TestBucketCache_MergeTimeSeriesResults(t *testing.T) {
 				{Series: []*qbtypes.TimeSeries{series2}},
 			},
 		},
-	})
+	}, nil)
 
 	// Wait for cache writes
 	time.Sleep(10 * time.Millisecond)
@@ -441,7 +441,7 @@ func TestBucketCache_RawData(t *testing.T) {
 		Value: rawData,
 	}
 
-	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 	time.Sleep(10 * time.Millisecond)
 
 	cached, missing := bc.GetMissRanges(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond})
@@ -481,7 +481,7 @@ func TestBucketCache_ScalarData(t *testing.T) {
 		Value: scalarData,
 	}
 
-	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 	time.Sleep(10 * time.Millisecond)
 
 	cached, missing := bc.GetMissRanges(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond})
@@ -509,7 +509,7 @@ func TestBucketCache_EmptyFingerprint(t *testing.T) {
 		Value: createTestTimeSeries("A", 1000, 5000, 1000),
 	}
 
-	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 	time.Sleep(10 * time.Millisecond)
 
 	// Should still be able to retrieve
@@ -564,7 +564,7 @@ func TestBucketCache_ConcurrentAccess(t *testing.T) {
 				Type:  qbtypes.RequestTypeTimeSeries,
 				Value: createTestTimeSeries(fmt.Sprintf("Q%d", id), query.startMs, query.endMs, 100),
 			}
-			bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 100 * time.Microsecond}, result)
+			bc.Put(context.Background(), valuer.UUID{}, query, qbtypes.Step{Duration: 100 * time.Microsecond}, result, nil)
 			done <- true
 		}(i)
 	}
@@ -624,7 +624,7 @@ func TestBucketCache_GetMissRanges_FluxInterval(t *testing.T) {
 		},
 	}
 
-	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, cachedResult)
+	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, cachedResult, nil)
 
 	// Get miss ranges
 	cached, missing := bc.GetMissRanges(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond})
@@ -686,7 +686,7 @@ func TestBucketCache_Put_FluxIntervalTrimming(t *testing.T) {
 	}
 
 	// Put the result
-	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: query.startMs, To: query.endMs}: result.Stats})
 
 	// Retrieve cached data
 	cached, missing := bc.GetMissRanges(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond})
@@ -756,7 +756,7 @@ func TestBucketCache_Put_EntireRangeInFluxInterval(t *testing.T) {
 	}
 
 	// Put the result - should not cache anything
-	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 
 	// Try to get cached data - should have no cached data
 	cached, missing := bc.GetMissRanges(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond})
@@ -862,7 +862,7 @@ func TestBucketCache_EmptyDataHandling(t *testing.T) {
 			}
 
 			// Put the result
-			bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, tt.result)
+			bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, tt.result, nil)
 
 			// Wait a bit for cache to be written
 			time.Sleep(10 * time.Millisecond)
@@ -927,7 +927,7 @@ func TestBucketCache_PartialValues(t *testing.T) {
 	}
 
 	// Put the result
-	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: query.startMs, To: query.endMs}: result.Stats})
 
 	// Wait for cache to be written
 	time.Sleep(10 * time.Millisecond)
@@ -997,7 +997,7 @@ func TestBucketCache_AllPartialValues(t *testing.T) {
 	}
 
 	// Put the result
-	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 
 	// Wait for cache to be written
 	time.Sleep(10 * time.Millisecond)
@@ -1058,7 +1058,7 @@ func TestBucketCache_FilteredCachedResults(t *testing.T) {
 	}
 
 	// Cache the wide range
-	bc.Put(ctx, orgID, query1, qbtypes.Step{Duration: 1000 * time.Millisecond}, result1)
+	bc.Put(ctx, orgID, query1, qbtypes.Step{Duration: 1000 * time.Millisecond}, result1, nil)
 	time.Sleep(10 * time.Millisecond)
 
 	// Now query for a smaller range (2000-3500ms)
@@ -1125,7 +1125,7 @@ func TestBucketCache_PromQLKeepsTheValueAtTheWindowEnd(t *testing.T) {
 				}},
 			}},
 		},
-	})
+	}, nil)
 	time.Sleep(10 * time.Millisecond)
 
 	cached, missing := bc.GetMissRanges(ctx, orgID, query, step)
@@ -1291,7 +1291,7 @@ func TestBucketCache_PartialValueDetection(t *testing.T) {
 		}
 
 		// Put the result
-		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 		time.Sleep(10 * time.Millisecond)
 
 		// Get cached data
@@ -1345,7 +1345,7 @@ func TestBucketCache_PartialValueDetection(t *testing.T) {
 		}
 
 		// Put the result
-		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 		time.Sleep(10 * time.Millisecond)
 
 		// Get cached data
@@ -1397,7 +1397,7 @@ func TestBucketCache_PartialValueDetection(t *testing.T) {
 		}
 
 		// Put the result
-		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+		bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 		time.Sleep(10 * time.Millisecond)
 
 		// Get cached data
@@ -1454,7 +1454,7 @@ func TestBucketCache_NoCache(t *testing.T) {
 	}
 
 	// Put the result in cache
-	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result)
+	bc.Put(ctx, orgID, query, qbtypes.Step{Duration: 1000 * time.Millisecond}, result, nil)
 	time.Sleep(10 * time.Millisecond)
 
 	// Verify data is cached
@@ -1475,11 +1475,12 @@ func TestBucketCachePutSlicesAtCellBoundaries(t *testing.T) {
 
 	// a 1s step gives 60s cells, so 30s to 150s spans three of them
 	query := &mockQuery{fingerprint: "sliced-query", startMs: 30000, endMs: 150000}
-	bc.Put(ctx, orgID, query, step, &qbtypes.Result{
+	result := &qbtypes.Result{
 		Type:  qbtypes.RequestTypeTimeSeries,
 		Value: createTestTimeSeries("A", 30000, 150000, 1000),
 		Stats: qbtypes.ExecStats{RowsScanned: 120, BytesScanned: 1200, DurationMS: 12},
-	})
+	}
+	bc.Put(ctx, orgID, query, step, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: 30000, To: 150000}: result.Stats})
 
 	type bucketShape struct {
 		cell, startMs, endMs, rowsScanned uint64
@@ -1539,7 +1540,7 @@ func TestBucketCachePutAbsorbsWritesWithinACell(t *testing.T) {
 			name:                  "OverlappingWrites_BecomeOne",
 			firstWindow:           qbtypes.TimeRange{From: 1000, To: 4000},
 			secondWindow:          qbtypes.TimeRange{From: 3000, To: 6000},
-			expectedBuckets:       []bucketShape{{1000, 6000, 6}},
+			expectedBuckets:       []bucketShape{{1000, 6000, 5}},
 			expectedMissingRanges: []*qbtypes.TimeRange{},
 			expectedValueCount:    5,
 		},
@@ -1570,12 +1571,13 @@ func TestBucketCachePutAbsorbsWritesWithinACell(t *testing.T) {
 
 			for _, window := range []qbtypes.TimeRange{testCase.firstWindow, testCase.secondWindow} {
 				query := &mockQuery{fingerprint: "absorb-query", startMs: window.From, endMs: window.To}
-				bc.Put(ctx, orgID, query, step, &qbtypes.Result{
+				result := &qbtypes.Result{
 					Type:  qbtypes.RequestTypeTimeSeries,
 					Value: createTestTimeSeries("A", window.From, window.To, 1000),
-					// one row per second, so a bucket's stats read as the seconds it was scanned for
+					// one row per second, so a bucket's stats read as the seconds of data it holds
 					Stats: qbtypes.ExecStats{RowsScanned: (window.To - window.From) / 1000},
-				})
+				}
+				bc.Put(ctx, orgID, query, step, result, map[qbtypes.TimeRange]qbtypes.ExecStats{window: result.Stats})
 			}
 
 			query := &mockQuery{
@@ -1609,7 +1611,7 @@ func TestBucketCachePutLeavesAnUnchangedCellAlone(t *testing.T) {
 		Type:     qbtypes.RequestTypeTimeSeries,
 		Value:    createTestTimeSeries("A", 1000, 10000, 1000),
 		Warnings: []string{"from the wide write"},
-	})
+	}, nil)
 
 	// a write already covered by the cell's bucket must not rewrite the entry,
 	// which shows as its warnings never arriving
@@ -1618,7 +1620,7 @@ func TestBucketCachePutLeavesAnUnchangedCellAlone(t *testing.T) {
 		Type:     qbtypes.RequestTypeTimeSeries,
 		Value:    createTestTimeSeries("A", 3000, 5000, 1000),
 		Warnings: []string{"from the inside write"},
-	})
+	}, nil)
 
 	var cellEntry qbtypes.CachedData
 	require.NoError(t, bc.cache.Get(ctx, orgID, bc.generateCacheKey(wide, 0), &cellEntry))

pkg/querier/querier.go

@@ -823,7 +823,7 @@ func (q *querier) executeWithCache(ctx context.Context, orgID valuer.UUID, query
 				return nil, err
 			}
 			// Store in cache for future use
-			q.bucketCache.Put(ctx, orgID, query, step, result)
+			q.bucketCache.Put(ctx, orgID, query, step, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: startMs, To: endMs}: result.Stats})
 			return result, nil
 		}
 	}
@@ -879,16 +879,19 @@ func (q *querier) executeWithCache(ctx context.Context, orgID valuer.UUID, query
 			if err != nil {
 				return nil, err
 			}
-			q.bucketCache.Put(ctx, orgID, query, step, result)
+			startMs, endMs := query.Window()
+			q.bucketCache.Put(ctx, orgID, query, step, result, map[qbtypes.TimeRange]qbtypes.ExecStats{{From: startMs, To: endMs}: result.Stats})
 			return result, nil
 		}
 	}
 
 	// Calculate total stats and filter out nil results
 	validResults := make([]*qbtypes.Result, 0, len(freshResults))
-	for _, result := range freshResults {
+	execStatsForQueriedTimeRanges := make(map[qbtypes.TimeRange]qbtypes.ExecStats, len(freshResults))
+	for i, result := range freshResults {
 		if result != nil {
 			validResults = append(validResults, result)
+			execStatsForQueriedTimeRanges[*missingRanges[i]] = result.Stats
 			totalStats.RowsScanned += result.Stats.RowsScanned
 			totalStats.BytesScanned += result.Stats.BytesScanned
 			totalStats.DurationMS += result.Stats.DurationMS
@@ -903,7 +906,7 @@ func (q *querier) executeWithCache(ctx context.Context, orgID valuer.UUID, query
 	mergedResult.Stats.DurationMS += totalStats.DurationMS
 
 	// Store merged result in cache
-	q.bucketCache.Put(ctx, orgID, query, step, mergedResult)
+	q.bucketCache.Put(ctx, orgID, query, step, mergedResult, execStatsForQueriedTimeRanges)
 
 	return mergedResult, nil
 }

pkg/types/querybuildertypes/querybuildertypesv5/cached.go

@@ -43,6 +43,23 @@ func (c *CachedBucket) Touches(other *CachedBucket) bool {
 	return c.StartMs <= other.EndMs && other.StartMs <= c.EndMs
 }
 
+// ProrateStatsToWindow returns the share of the stats for the part of the
+// bucket inside [startMs, endMs).
+func (c *CachedBucket) ProrateStatsToWindow(startMs, endMs uint64) ExecStats {
+	return c.Stats.prorate(overlapMs(c.StartMs, c.EndMs, startMs, endMs), c.EndMs-c.StartMs)
+}
+
+// ProrateStatsToUnqueried returns the share of the stats for the part of the
+// bucket no queried range covers.
+func (c *CachedBucket) ProrateStatsToUnqueried(execStatsForQueriedTimeRanges map[TimeRange]ExecStats) ExecStats {
+	widthMs := c.EndMs - c.StartMs
+	queriedMs := uint64(0)
+	for queried := range execStatsForQueriedTimeRanges {
+		queriedMs += overlapMs(c.StartMs, c.EndMs, queried.From, queried.To)
+	}
+	return c.Stats.prorate(widthMs-min(queriedMs, widthMs), widthMs)
+}
+
 // CachedData represents the full cached data for a query.
 type CachedData struct {
 	Buckets  []*CachedBucket `json:"buckets"`
@@ -88,7 +105,7 @@ func (c *CachedData) Cost() int64 {
 	return size
 }
 
-func (r *Result) SplitIntoCacheCellBuckets(startMs, endMs, cellWidthMs uint64) ([]*CachedBucket, error) {
+func (r *Result) SplitIntoCacheCellBuckets(startMs, endMs, cellWidthMs uint64, execStatsForQueriedTimeRanges map[TimeRange]ExecStats) ([]*CachedBucket, error) {
 	tsData, ok := r.Value.(*TimeSeriesData)
 	if !ok || tsData == nil {
 		return nil, errors.NewInternalf(errors.CodeInternal, "result of type %s cannot be bucketed", r.Type.StringValue())
@@ -111,18 +128,39 @@ func (r *Result) SplitIntoCacheCellBuckets(startMs, endMs, cellWidthMs uint64) (
 		if err != nil {
 			return nil, err
 		}
+		// A slice costs what was queried inside it; the rest of the result came
+		// out of the cache and is costed there already
+		sliceStats := ExecStats{StepIntervals: maps.Clone(r.Stats.StepIntervals)}
+		for queried, stats := range execStatsForQueriedTimeRanges {
+			queriedMs := overlapMs(from, to, queried.From, queried.To)
+			if queriedMs == 0 {
+				continue
+			}
+			share := stats.prorate(queriedMs, queried.To-queried.From)
+			sliceStats.RowsScanned += share.RowsScanned
+			sliceStats.BytesScanned += share.BytesScanned
+			sliceStats.DurationMS += share.DurationMS
+		}
 		cellBuckets = append(cellBuckets, &CachedBucket{
 			StartMs: from,
 			EndMs:   to,
 			Type:    r.Type,
 			Value:   value,
-			Stats:   r.Stats.prorate(to-from, endMs-startMs),
+			Stats:   sliceStats,
 		})
 		from = to
 	}
 	return cellBuckets, nil
 }
 
+// overlapMs is the length of [aFromMs, aToMs) ∩ [bFromMs, bToMs).
+func overlapMs(aFromMs, aToMs, bFromMs, bToMs uint64) uint64 {
+	if fromMs, toMs := max(aFromMs, bFromMs), min(aToMs, bToMs); fromMs < toMs {
+		return toMs - fromMs
+	}
+	return 0
+}
+
 // prorate returns the share of the stats that partMs out of wholeMs of the
 // scanned window accounts for.
 func (e ExecStats) prorate(partMs, wholeMs uint64) ExecStats {

tests/integration/tests/queriermetrics/17_cache.py

@@ -442,3 +442,55 @@ def test_builder_refreshing_a_sliding_time_range(
     assert uncached.status_code == HTTPStatus.OK, uncached.text
 
     assert_results_equal(from_cache.json(), uncached.json(), "A", "the twentieth refresh")
+
+
+def test_builder_refreshing_does_not_inflate_rows_scanned(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_metrics: Callable[[list[Metrics]], None],
+) -> None:
+    metric_name = f"cache_rows_scanned_{uuid4().hex[:8]}"
+
+    # 90 minutes back so even the last refresh closes clear of the flux
+    # interval, which holds recent data out of the cache
+    start_time = datetime.fromtimestamp(int((datetime.now(tz=UTC) - timedelta(minutes=90)).timestamp()) // 60 * 60, tz=UTC)
+    start_time_ms = int(start_time.timestamp() * 1000)
+
+    query = [build_builder_query("A", metric_name, "max", "max")]
+
+    insert_metrics(
+        [
+            Metrics(
+                metric_name=metric_name,
+                labels={"service": "api"},
+                timestamp=start_time + timedelta(minutes=minute),
+                value=minute,
+                type_="Gauge",
+                is_monotonic=False,
+            )
+            for minute in range(70)
+        ]
+    )
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+
+    # a dashboard left open on a thirty minute range, re-running a minute later
+    # each time. rowsScanned is a cost indicator, so a refresh served mostly out
+    # of the cache cannot claim more than every read made so far: the first
+    # full read and then one new minute per refresh, measured with the cache off
+    uncached_rows_scanned_so_far = 0
+    for refresh in range(10):
+        refresh_start_ms = start_time_ms + refresh * MINUTE_MS
+        refresh_end_ms = refresh_start_ms + 30 * MINUTE_MS
+
+        uncached_start_ms = refresh_start_ms if refresh == 0 else refresh_end_ms - MINUTE_MS
+        uncached_response = make_query_request(signoz, token, uncached_start_ms, refresh_end_ms, query, no_cache=True)
+        assert uncached_response.status_code == HTTPStatus.OK, f"refresh {refresh}: uncached read failed: {uncached_response.text}"
+        uncached_rows_scanned_so_far += uncached_response.json()["data"]["meta"]["rowsScanned"]
+        assert uncached_rows_scanned_so_far > 0, f"refresh {refresh}: uncached read reported no rows scanned"
+
+        cached_response = make_query_request(signoz, token, refresh_start_ms, refresh_end_ms, query, no_cache=False)
+        assert cached_response.status_code == HTTPStatus.OK, f"refresh {refresh}: cached read failed: {cached_response.text}"
+        cached_rows_scanned = cached_response.json()["data"]["meta"]["rowsScanned"]
+        assert cached_rows_scanned <= uncached_rows_scanned_so_far, f"refresh {refresh}: cached read reported {cached_rows_scanned} rows scanned, more than the {uncached_rows_scanned_so_far} rows all uncached reads scanned so far"