fix: sum without over an unwrap range function keeps a series per remaining label set, like Loki

CHANGELOG.md

@@ -7,6 +7,41 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Fixed
+
+- **`sum without (...)` over a range function with `unwrap`, and over `rate` at
+  an instant, keeps a series per remaining stream label set instead of folding
+  every series into one.** `sum without (pod) (max_over_time({app="payment-service"}
+  | logfmt amount | unwrap amount [5m]))` answered one series on query_range and on an
+  instant query, where Loki 3.7.7 keeps every label but `pod` and folds each
+  group (`pkg/logql/evaluator.go:451-466`). The last stats pipe now keeps the
+  stream, VictoriaLogs answers a row per stream from stats, and the proxy drops
+  the excluded labels and folds the rows with the operator of the aggregation
+  (`sum`, `max`, `min`, `count` or `avg`, when it is the outermost aggregation of
+  the query). A `topk`, `bottomk`, `sort`, `stddev` or `stdvar` around it is
+  applied to the folded series, and so is a comparison around it and, on
+  query_range, a `label_replace` (on an instant query the label is not added:
+  `semantics/label-replace-instant-vector`); a binary expression folds each
+  operand on its own. `topk`, `bottomk`, `stddev` and `stdvar` with a `without`
+  clause, and a `without` under another aggregation, are answered as before.
+  `max_query_series` now applies to the folded answer, as on Loki
+  (`pkg/logql/engine.go:536-548`): four streams folded into two series fit a
+  limit of two, a plain client over the limit gets Loki's 400, and Logs
+  Drilldown gets the busiest series with Loki's warning; the per-stream rows read
+  before the fold are bounded by `-manual-range-metric-row-limit` (when it is not
+  below the tenant's effective `max_query_series`), whose error names that flag
+  and answers 400. For this fold only, a parser that names the keys it extracts
+  (`| logfmt amount`, `| logfmt amount, route`) makes the series of a stream the
+  stream plus those labels, as on Loki, so the unique values in a line no longer
+  split it: a `sum without (pod)` over `max_over_time` of such a pipeline was a
+  sum of per-line maxima. A bare `| logfmt` or `| json`, which extracts every key
+  the proxy does not expose, and a key that is also a label of the stream
+  selector (Loki names it `<key>_extracted`) keep the previous grouping. Every
+  other query, with or without an extraction list, is translated and answered
+  as before; what differs from Loki there is registered with
+  `semantics/unwrap-extraction-list-series-identity`,
+  `semantics/unwrap-sum-without-collapses` (fixed, #739) and the rows it names.
+
 ## [2.17.0] - 2026-10-17
 
 ### Performance

bench/ab/results/2026-10-17-unwrapwithout-1h.json

@@ -0,0 +1,333 @@
+{
+ "label": "unwrapwithout-1h",
+ "date": "2026-10-17",
+ "set": "unwrap-sum-without",
+ "description": "`sum without` over a range function with unwrap (PR for semantics/unwrap-sum-without-collapses): Loki keeps a series per remaining label set; the last stats pipe had no grouping and VictoriaLogs answered one row, so the proxy answered one series. W1-W4 are range queries over a logfmt stream group (max_over_time, sum_over_time, rate, max without), parsed with `| logfmt amount` so a Loki series is the stream (a bare `| logfmt` makes every payment line a series of its own on Loki: thousands of series, and Loki cannot answer within the timeout; semantics/unwrap-per-series-identity-raw-rows), W5-W6 are instant, W7 wraps the aggregation in topk, W8 counts a stream-label function whose series limit now applies to the folded answer, W9 is the label_replace instant control that must not move. A sum without over an unwrap with no parser is left out: no field of the generator's data is numeric without a parser, so every build answers an empty result. Run with the control set.",
+ "baseline": "base",
+ "candidate": "pr",
+ "reference": "loki",
+ "runs": 5,
+ "valid": true,
+ "restart_before": "0",
+ "restart_after": "0",
+ "noise": 0.3,
+ "timing": "warm",
+ "verdicts": {
+  "same": 9
+ },
+ "result_differs_from_reference": 5,
+ "result_differs_preexisting": 5,
+ "stale_answers": {
+  "base": 0,
+  "pr": 0,
+  "loki": 0
+ },
+ "rows": [
+  {
+   "shape": "W1 sum without pod, max_over_time",
+   "range": "1h",
+   "p50": {
+    "base": 0.015,
+    "pr": 0.015,
+    "loki": 0.017
+   },
+   "p95": {
+    "base": 0.023,
+    "pr": 0.016,
+    "loki": 0.017
+   },
+   "cold": {
+    "base": 0.042,
+    "pr": 0.024,
+    "loki": 0.079
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 39.36,
+    "pr": 39.36,
+    "loki": 39.36
+   },
+   "parity_vs_reference": "differs",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "differs",
+   "verdict": "same",
+   "speedup": 1.0
+  },
+  {
+   "shape": "W2 sum without pod, sum_over_time",
+   "range": "1h",
+   "p50": {
+    "base": 0.015,
+    "pr": 0.017,
+    "loki": 0.009
+   },
+   "p95": {
+    "base": 0.017,
+    "pr": 0.017,
+    "loki": 0.011
+   },
+   "cold": {
+    "base": 0.019,
+    "pr": 0.016,
+    "loki": 0.059
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 14.6,
+    "pr": 14.6,
+    "loki": 14.6
+   },
+   "parity_vs_reference": "differs",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "differs",
+   "verdict": "same",
+   "speedup": 0.88
+  },
+  {
+   "shape": "W3 sum without pod, rate",
+   "range": "1h",
+   "p50": {
+    "base": 0.054,
+    "pr": 0.05,
+    "loki": 0.01
+   },
+   "p95": {
+    "base": 0.058,
+    "pr": 0.057,
+    "loki": 0.013
+   },
+   "cold": {
+    "base": 0.06,
+    "pr": 0.06,
+    "loki": 0.031
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 13.76,
+    "pr": 13.76,
+    "loki": 13.76
+   },
+   "parity_vs_reference": "same",
+   "baseline_parity_vs_reference": "same",
+   "parity_vs_baseline": "same",
+   "verdict": "same",
+   "speedup": 1.08
+  },
+  {
+   "shape": "W4 max without pod, max_over_time",
+   "range": "1h",
+   "p50": {
+    "base": 0.015,
+    "pr": 0.016,
+    "loki": 0.009
+   },
+   "p95": {
+    "base": 0.016,
+    "pr": 0.016,
+    "loki": 0.009
+   },
+   "cold": {
+    "base": 0.012,
+    "pr": 0.016,
+    "loki": 0.032
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 13.2,
+    "pr": 17.41,
+    "loki": 17.41
+   },
+   "parity_vs_reference": "differs",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "differs",
+   "verdict": "same",
+   "speedup": 0.94
+  },
+  {
+   "shape": "W5 sum without pod, max_over_time, instant",
+   "range": "instant",
+   "p50": {
+    "base": 0.005,
+    "pr": 0.004,
+    "loki": 0.009
+   },
+   "p95": {
+    "base": 0.006,
+    "pr": 0.005,
+    "loki": 0.01
+   },
+   "cold": {
+    "base": 0.004,
+    "pr": 0.004,
+    "loki": 0.009
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 17.41,
+    "pr": 17.41,
+    "loki": 17.41
+   },
+   "parity_vs_reference": "same",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "differs",
+   "verdict": "same",
+   "speedup": 1.25
+  },
+  {
+   "shape": "W6 sum without pod, rate, instant",
+   "range": "instant",
+   "p50": {
+    "base": 0.004,
+    "pr": 0.004,
+    "loki": 0.007
+   },
+   "p95": {
+    "base": 0.005,
+    "pr": 0.006,
+    "loki": 0.008
+   },
+   "cold": {
+    "base": 0.004,
+    "pr": 0.004,
+    "loki": 0.008
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 2.99,
+    "pr": 2.99,
+    "loki": 2.99
+   },
+   "parity_vs_reference": "same",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "differs",
+   "verdict": "same",
+   "speedup": 1.0
+  },
+  {
+   "shape": "W7 topk 5 of sum without pod",
+   "range": "1h",
+   "p50": {
+    "base": 0.015,
+    "pr": 0.015,
+    "loki": 0.009
+   },
+   "p95": {
+    "base": 0.016,
+    "pr": 0.018,
+    "loki": 0.01
+   },
+   "cold": {
+    "base": 0.014,
+    "pr": 0.013,
+    "loki": 0.036
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 13.36,
+    "pr": 17.12,
+    "loki": 17.12
+   },
+   "parity_vs_reference": "differs",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "differs",
+   "verdict": "same",
+   "speedup": 1.0
+  },
+  {
+   "shape": "W8 sum without pod, count_over_time (production)",
+   "range": "1h",
+   "p50": {
+    "base": 0.35,
+    "pr": 0.35,
+    "loki": 40.002
+   },
+   "p95": {
+    "base": 0.418,
+    "pr": 0.398,
+    "loki": 40.002
+   },
+   "cold": {
+    "base": 0.388,
+    "pr": 0.387,
+    "loki": 40.002
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "0"
+   },
+   "vl_cpu_max": {
+    "base": 17.12,
+    "pr": 27.16,
+    "loki": 27.16
+   },
+   "parity_vs_reference": "differs",
+   "baseline_parity_vs_reference": "differs",
+   "parity_vs_baseline": "same",
+   "verdict": "same",
+   "speedup": 1.0
+  },
+  {
+   "shape": "W9 label_replace control, instant",
+   "range": "instant",
+   "p50": {
+    "base": 0.003,
+    "pr": 0.003,
+    "loki": 0.038
+   },
+   "p95": {
+    "base": 0.004,
+    "pr": 0.003,
+    "loki": 0.044
+   },
+   "cold": {
+    "base": 0.003,
+    "pr": 0.002,
+    "loki": 0.034
+   },
+   "status": {
+    "base": "200",
+    "pr": "200",
+    "loki": "200"
+   },
+   "vl_cpu_max": {
+    "base": 1.89,
+    "pr": 1.89,
+    "loki": 1.89
+   },
+   "parity_vs_reference": "same",
+   "baseline_parity_vs_reference": "same",
+   "parity_vs_baseline": "same",
+   "verdict": "same",
+   "speedup": 1.0
+  }
+ ]
+}