feat(asap-tools): Alibaba and Google trace replay builders

asap-tools/experiments/datasets/alibaba/ALIBABA_MAPPING.md

@@ -0,0 +1,80 @@
+# Alibaba trace → random query workloads (handoff)
+
+**Goal:** realistic, strong end-to-end results on Alibaba data for the eval section.
+
+The motivation section (Prometheus, PromSketch, PilotDB; cost vs accuracy) should ideally use the same data, label mapping and similar queries, so the two line up. They don't need to match exactly.
+
+## Data the end-to-end runs feed
+
+**Source:** Alibaba cluster-trace-microservices-v2022, MSMetrics, pod `cpu_utilization` (0–1).
+- Columns: `timestamp, msname, msinstanceid, nodeid, cpu_utilization, memory_utilization`.
+- 60 s step, about 470k pods per step, 28k distinct msname.
+- Only about the first hour is local: `benchmarks/metrics_observability/data/alibaba-cluster-data/cluster-trace-microservices-v2022/0d0_0d1/data/MSMetrics/`.
+  - That is `MSMetricsUpdate_0.csv` plus `MSMetrics_1.tar.gz`.
+  - Rows are **not** sorted by time.
+
+**Subset:** fit to our synthetic label shape, so the existing workloads, plans and configs apply unchanged.
+- 4 region × 5 service × 10 job × 50 label_0 = 10,000 series.
+- `job` = one microservice (msname).
+  - Seed 862 draws them from the 1,785 msnames with ≥50 pods reporting at all 60 steps.
+- `label_0` = one of its pods (msinstanceid; the first 50 by pod number).
+- `region` and `service` are synthetic groupings of microservices.
+  - The trace has neither level.
+  - `nodeid` cuts across msname, so it doesn't nest.
+- Label values: region `r`, service `r.s`, job `r.s.j`, label_0 `r.s.j.k`.
+
+**Values:** real pod CPU.
+
+| Statistic | Value |
+|---|---|
+| min | 0.003 |
+| median | 0.12 |
+| p95 | 0.40 |
+| p99 | 0.58 |
+| max | 0.997 |
+
+- No zeros.
+- The busiest microservice is about 5× the median one.
+- The top 10 of 200 microservices hold 15% of total CPU.
+- Region totals are 20–26k.
+
+**Time mapping:** a quick-result compromise.
+- Replayed at 10 samples/s per series, the same as the synthetic base set.
+- One trace minute becomes one 100 ms sample, so the 60-step hour loops every 6 s.
+- Spread and skew across pods and services are real; each series' pattern over time repeats.
+- A realistic run needs more trace (`fetchData.sh`) and a smaller speedup.
+
+**Files** (generated, not tracked):
+  - `replay.csv`: `offset_ms,region,service,job,label_0,value`; 600k rows, sorted by offset.
+  - `replay_mapping.csv`: our labels → msname/msinstanceid.
+  - `replay_table.json`: a paced-feeder workload table with `values.replay`.
+- Generator: `asap-tools/experiments/datasets/alibaba/make_replay.py`.
+
+## Query workloads
+
+Seed 862, 10 random workloads; manifests are in `asap-tools/experiments/datasets/fake_metric/random_workloads/seed-862/`.
+
+Each end-to-end run uses a workload's 7 `common-promql` queries:
+- **Metric and groupings:** metric `data`, grouped by `(region)`, `(region, service)` or `(region, service, job)`.
+- **Sums (2):** `sum by (G) (sum_over_time(data[W]))`.
+
+  | Window W | Repeated every |
+  |---|---|
+  | 1m | 1–2 s |
+  | 5m | 5–10 s |
+  | 10m | 10–20 s |
+  | 20m | 30–60 s |
+
+- **Quantiles (3):** 100 s lookback, repeated every 1 s.
+  - PromQL can't express a pooled quantile, so it uses `avg by (G) (quantile_over_time(phi, data[100s]))`.
+  - Every system, ClickHouse and LiteOlly included, currently computes this same `avg` of per-series quantiles, so all answers agree.
+  - The intended pooled `quantile by (G)` over 100 s is being added (LiteOlly support merged in #887–890).
+- **Top-k (2):** `topk(k, …)` over counts, with K capped at 1,000 keys.
+
+On Alibaba data these read naturally: "a quantile of pod CPU per microservice or region over the last 100 s", and "total CPU per service over 5–20 min".
+
+## End-to-end status (2026-10-15)
+
+- **Synthetic base:** 5 systems × 10 workloads, all valid.
+  - LiteOlly vs worst baseline: 6.2× less CPU than raw ClickHouse, and 11.6× lower p95 than Prometheus.
+- **Alibaba:** pilot w01 × 5 systems is finishing. The full 10 workloads follow if it is valid.

asap-tools/experiments/datasets/alibaba/make_replay.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
-"""A replay table for paced-feeder from Alibaba v2022 MSMetrics: pod CPU
-utilization in the random workloads' label shape (4 region x 5 service x 10
+"""A replay table for paced-feeder from Alibaba v2022 per-pod metrics (MSMetrics
+CPU by default, or an MSRTMCR call rate with --column) in the random workloads'
+label shape (4 region x 5 service x 10
 job x 50 label_0), so their plans apply unchanged.
 
 Each job is one microservice and each label_0 one of its pods; region and
@@ -17,29 +18,40 @@
 
 Usage: python3 make_replay.py MSMetricsUpdate_0.csv MSMetricsUpdate_1.csv \
     --out replay [--table eval_table.json]
+       python3 make_replay.py MSRTMCR_*.tar.gz --column providerrpc_mcr --out calls
 """
 
 import argparse
 import csv
+import io
 import json
 import random
+import tarfile
 from collections import defaultdict
 
 REGIONS, SERVICES, JOBS, PODS = 4, 5, 10, 50
 
 
 def rows(paths):
+    """The CSV rows of `paths`: plain CSVs, or tarballs of them."""
     for path in paths:
-        with open(path, newline="") as f:
-            for r in csv.DictReader(f):
-                yield r
+        if path.endswith(".tar.gz"):
+            with tarfile.open(path) as tar:
+                for member in tar:
+                    if member.isfile():
+                        text = io.TextIOWrapper(tar.extractfile(member), newline="")
+                        yield from csv.DictReader(text)
+        else:
+            with open(path, newline="") as f:
+                yield from csv.DictReader(f)
 
 
 def main():
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("csv", nargs="+")
     parser.add_argument("--out", required=True)
     parser.add_argument("--seed", type=int, default=862)
+    parser.add_argument("--column", default="cpu_utilization", help="the value column")
     parser.add_argument("--table", help="workload table whose `data` gets the replay")
     args = parser.parse_args()
 
@@ -72,7 +84,7 @@ def main():
 
     start = min(int(s) for s in steps)
     out = sorted(
-        (int(r["timestamp"]) - start, *labels[r["msinstanceid"]], r["cpu_utilization"])
+        (int(r["timestamp"]) - start, *labels[r["msinstanceid"]], r[args.column])
         for r in rows(args.csv)
         if r["msinstanceid"] in labels
     )
@@ -90,7 +102,7 @@ def write_table(table_path, out_path, rows, steps):
     paced-feeder's series order (mixed radix, label_0 fastest)."""
     by_series = {}
     for offset, *_, label_0, value in rows:  # sorted by offset
-        by_series.setdefault(label_0, []).append(round(float(value), 6))
+        by_series.setdefault(label_0, []).append(round(float(value or 0), 6))
     values = []
     for r in range(REGIONS):
         for s in range(SERVICES):