eval: export ASAPQuery random workload manifests by milindsrivastava1997 · Pull Request #226 · ProjectASAP/sketch-bench

milindsrivastava1997 · 2026-10-14T23:40:00Z

Adds --random-manifest to scripts/export_autosketch_eval_table.py. It writes the eval table for an ASAPQuery random workload manifest (asap-tools random_workloads.py, ProjectASAP/ASAPQuery#879), so autosketch_vs_asap can plan it and ASAPQuery's --planner saved can convert the plan.

Only the manifest's common-promql queries are exported (the five-system arm). Each is a schema RQE on metric data, using the manifest's labels:

Sums: on one stream at their own grouping, so a fine sum can serve a coarse one (roll-up).
Quantiles: avg by (G) of per-series quantiles, so the RQE is per series: 1,000 samples per 100s window at 10/s.
Top-k: grouped heaps over 500 series fall below the curves' first 1,000 keys, so they're read at 1,000 (which overstates the error) and noted in assumptions.

The e2e block holds the manifest's PromQL with one RQE per query.

Testing:

python3 scripts/test_export_autosketch_eval_table.py: 33 pass, including a new test for the export.
Workload 01 of seed 862: all 7 RQEs served, none dropped. The cheapest asap/cpu plan rolls the coarse sum up from the fine one, and the plan converts in ASAPQuery (archived in ASAPQuery#879).

🤖 Generated with Claude Code

--random-manifest writes the eval table of an ASAPQuery random workload (asap-tools random_workloads.py, ProjectASAP/ASAPQuery#879): its common-promql queries as schema RQEs on metric `data`, so a fine sum can serve a coarse one, quantiles per series (avg of per-series quantiles), and top-k heaps below the curves' 1,000 keys read there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

milindsrivastava1997

Code review findings (via Claude Code).

milindsrivastava1997 · 2026-10-15T01:27:19Z

+            **DEFAULT_TARGETS,
+        }
+        qs.append(q)
+        e2e.append({"promql": query["promql"], "interval_secs": int(t),


These e2e queries carry no expr, while #225 makes expr the standard field on every query that isn't a rate. #228 adds it, so this is fine if 228 lands with this PR; if 226 lands alone, random-manifest tables have no trees.

Leaving expr to #228, which is stacked on this PR; they should land together.

milindsrivastava1997 · 2026-10-15T01:27:19Z

+    by_name = {label["name"]: label for label in labels}
+    for name in grouping:
+        parent = by_name[name].get("child_of")
+        assert parent is None or parent in grouping, (grouping, name)


This asserts that a child label's parent is in the grouping instead of adding the parent, so a valid child-only grouping crashes the export. E.g. sum by (label_0) (...), where label_0 is a child of job, raises AssertionError. closed_labels() already adds parents.

Fixed in 5f1d374: the export now uses closed_labels(), so sum by (label_0) counts 2 × 500 groups (test added).

milindsrivastava1997 · 2026-10-15T01:27:19Z

+TOPK_MIN_CURVE_KEYS = 1_000
+
+
+def random_groups(labels, grouping):


random_groups() copies the fan-out logic of closed_labels()/grouping_cardinality() only because those take a SCHEMAS metric name. The copies have already drifted (this one asserts where the original adds parents). Letting closed_labels take a list of labels would remove the copy and fix the crash above.

Fixed in 5f1d374: closed_labels() takes a label list, labels_cardinality() wraps it for any labels, and random_groups() is gone.

milindsrivastava1997 · 2026-10-15T01:27:19Z

+            "alpha_rank": alpha, "alpha_memory": alpha, "tail_class": "",
+            "families": SYNTHETIC_FAMILIES["topk" if keys else capability],
+            "capability": "topk" if keys else capability,
+            "stream": f"data/{RANDOM_STREAMS[capability]}", "metric": "data",


Any common-promql capability other than sum, quantile or top-k stops the whole export with a bare KeyError here (and in SYNTHETIC_FAMILIES), and no table is written. Better to skip such queries with a message, or reject them with a clear error.

Fixed in 5f1d374: queries of other capabilities are left out with a stderr line, and the table is still written (test added).

milindsrivastava1997 · 2026-10-15T01:27:19Z

+            continue
+        [rqe] = query["rqes"]
+        capability = rqe["capability"]
+        grouping = names if capability == "quantile" else rqe["grouping_labels"]


Every quantile is assumed to be avg by (G) of per-series quantiles; the query's PromQL and grouping_labels are never checked. A cross-series quantile by (job) (0.9, data) would be exported as a per-series RQE with min_N = per_sec * s, so its planned sketch size and error are silently wrong.

Fixed in 5f1d374: a quantile's PromQL must match avg by (<grouping_labels>) (quantile_over_time(phi, data[W])), the only shape random_workloads.py writes; anything else is refused (test added).

milindsrivastava1997 · 2026-10-15T01:27:19Z

+        query_id, keys, assumptions = query["id"], None, []
+        if capability == "topk_by_value":
+            query_id = f"topk{rqe['k']}_{query_id}"
+            keys = series // groups


This reads a top-k's grouping_labels as the heap's grouping (keys = series // groups). #225 says ASAPQuery's QueryRequirements.grouping_labels for a top-k is the per-series labels it returns. Under that reading, topk(4, data) has grouping ['job', 'label_0'], giving 1,000 heaps of 1 key (clamped, with an assumption) instead of one heap over 1,000 keys. Which reading is right?

In the manifest, a top-k's grouping_labels is its topk by grouping, the heaps' buckets: random_workloads.py writes only topk by (G), with G as grouping_labels. So keys per heap = series / groups is right here. #225's reading (the per-series labels returned) is ASAPQuery's QueryRequirements, a different field; #228's expr maps between the two (grouping per-series, topk_by G).

milindsrivastava1997 · 2026-10-15T01:27:19Z

+    labels = manifest["labels"]
+    names = [label["name"] for label in labels]
+    series = random_groups(labels, names)
+    per_sec = manifest["samples_per_sec_per_series"]


samples_per_sec_per_series is copied without checking that it's an integer that divides 1000. The runner reads it with as_u64() and asserts 1000 % per_sec == 0, so a rate of 3/s panics only after the table is written, and a rate stored as 10.0 makes as_u64() return None and the runner silently falls back to SCRAPE_MS.

Fixed in 5f1d374: the rate must be an int dividing 1000, else ValueError before anything is written (3 and 10.0 tested).

milindsrivastava1997 · 2026-10-15T01:27:19Z

+        kind = "values" if capability == "quantile" else "keys"
+        q = {
+            "dataset": dataset, "query_id": query_id, "kind": kind,
+            "range": f"{s:g}s", "range_s": s, "step_s": int(t), "theta": theta, "K": keys,


f"{s:g}s" keeps only 6 significant digits, and this string is part of the RQE id. E.g. lookback_ms 1_234_567 gives 1234.57s while range_s is 1234.567, and lookbacks that differ below 6 digits can get the same id, with no uniqueness check on the table.

Fixed in 5f1d374: the range is {ms // 1000}s for whole seconds, else {ms}ms, and duplicate RQE ids are refused. Workload 01's ids are unchanged.

milindsrivastava1997 · 2026-10-15T01:27:19Z

+            "covered_min_N": n_items, "max_covered_share": 1 / groups,
+            "covered_max_N": n_items,
+            "assumptions": assumptions,
+            **DEFAULT_TARGETS,


Each manifest RQE carries an accuracy_sla (0.95 today), but this always applies DEFAULT_TARGETS. Nothing breaks yet, but a manifest with a stricter SLA would get sketches sized for the wrong accuracy and nothing would record it. Either read it or assert it matches the defaults.

Fixed in 5f1d374: an accuracy_sla other than 0.95 (DEFAULT_TARGETS, the runner's p95) is refused.

milindsrivastava1997 · 2026-10-15T01:27:19Z

        print(f"Done. {len(tables)} `synthetic` tables and {len(plan)} planned runs in "
              f"{args.out}", file=sys.stderr)
        return 0
+    if args.random_manifest:


--random-manifest is checked after --synthetic, so --synthetic --random-manifest m.json writes the synthetic grid and ignores the manifest. --samples-per-sec and --keep-duplicate-queries are accepted with it but have no effect. A mutually exclusive group, or a parser error, would make that visible.

Fixed in 5f1d374: --random-manifest with --synthetic, --summary, --samples-per-sec or --keep-duplicate-queries is a parser error.

- closed_labels() takes a label list, so random manifests reuse it: a child-only grouping counts its parents' groups instead of failing. - Other capabilities are left out (logged); a quantile that isn't an avg of per-series quantiles, an accuracy_sla other than 0.95, and a rate that doesn't divide 1000 are refused. - RQE ids carry the exact range; duplicate ids are refused. - --random-manifest refuses the flags it would ignore. Workload 01's table is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Per-series quantiles that differ only in their avg's grouping read the same per-series RQE; ASAPQuery's --planner saved refuses one leaf naming two RQEs. They now share the first query's id. Workloads 03, 04, 06, 08 and 09 of seed 862 hit this; workload 01's table is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

milindsrivastava1997

Code review: 10 findings, posted inline. The first three are correctness bugs in random_manifest_table.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+        ms, t = rqe["lookback_ms"], rqe["interval_ms"] / 1000
+        assert t == int(t), f"{query['id']}: interval {t} s is not whole seconds"
+        # Queries reading one RQE (per-series quantiles that differ only in
+        # their avg's grouping) share its id.


Correctness: the dedup key uses rqe.get("phi"), not the phi in the PromQL. If the RQE has no phi field (the test manifest's q3 has none), quantile_over_time(0.5, data[100s]) and quantile_over_time(0.99, data[100s]) get the same key, and the p99 is served from the p50's RQE. The regex already captures phi; put it in the key.

Holding: the quantile branch is being reworked (pooled quantiles grouped by G instead of avg of per-series), which changes this check; will fix with that rework.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+            e2e.append({"promql": query["promql"], "interval_secs": int(t),
+                        "rqes": [shared[key]]})
+            continue
+        groups = labels_cardinality(labels, grouping)


Correctness: RQEs aren't checked against the curves' first N (SCHEMA_MIN_COVERED_N = 1e3), as schema_queries() does. A per-series quantile_over_time(0.9, data[10s]) at 10/s covers only 100 samples, and an instant top-k over 2 jobs covers 500; both are exported, and the planner has to extrapolate to size them.

Fixed in ba319e0: RQEs under the curves' first N (1e3 items per group per window) are left out with a log line, and their e2e query with them (test added).

milindsrivastava1997 · 2026-10-15T02:22:24Z

+                             f"{RANDOM_ACCURACY_SLA} (DEFAULT_TARGETS) is exported")
+        grouping = rqe["grouping_labels"]
+        if capability == "quantile":
+            m = AVG_OF_SERIES_QUANTILES.match(query["promql"])


Correctness: the PromQL window data[...] is never compared with lookback_ms, and its phi is never compared with the RQE's phi. A data[5m] query with lookback_ms 100000 passes, so the RQE is sized for 100 s while the e2e query reads 5 m.

Holding: the quantile branch is being reworked (pooled quantiles grouped by G instead of avg of per-series), which changes this check; will fix with that rework. #228's full PromQL match may cover it too.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+        grouping = rqe["grouping_labels"]
+        if capability == "quantile":
+            m = AVG_OF_SERIES_QUANTILES.match(query["promql"])
+            if not m or m[1].split(", ") != grouping:


The grouping check compares m[1].split(", ") as an ordered list, so avg by (job,label_0) or avg by (label_0, job) raises a ValueError even though the shape is supported. Strip each label and compare sets.

Holding: the quantile branch is being reworked (pooled quantiles grouped by G instead of avg of per-series), which changes this check; will fix with that rework.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+                                 f"{query['promql']!r}")
+            grouping = names
+        ms, t = rqe["lookback_ms"], rqe["interval_ms"] / 1000
+        assert t == int(t), f"{query['id']}: interval {t} s is not whole seconds"


A bare assert is skipped under python -O. With interval_ms 500, int(t) truncates the interval to 0 and a zero query interval is exported. Raise a ValueError like the other input checks.

Fixed in ba319e0: a non-whole-second interval (e.g. 500 ms) is a ValueError (test added).

milindsrivastava1997 · 2026-10-15T02:22:24Z

+        [rqe] = query["rqes"]
+        capability = rqe["capability"]
+        if capability not in RANDOM_STREAMS:
+            print(f"left out {query['id']}: no RQE for capability {capability!r}",


No e2e dedup on (promql, interval_secs), unlike e2e_block(). Two manifest queries with the same PromQL and interval produce two identical e2e entries, so the query runs twice and its latency is counted twice.

Fixed in ba319e0: a repeated (PromQL, interval) is one e2e entry (test added).

milindsrivastava1997 · 2026-10-15T02:22:24Z

+    parser.add_argument("--random-manifest",
+                        help="write the eval table of an ASAPQuery random workload manifest "
+                             f"(asap-tools random_workloads.py), its {RANDOM_SUPPORT} queries, "
+                             "to --out instead")


--synthetic, --summary, --samples-per-sec and --keep-duplicate-queries are refused with --random-manifest, but --curves, --saturation, --curves-big, --saturation-big and --merge-curves are dropped without a warning, and inputs records only the manifest hash. Refuse those too.

Fixed in ba319e0: --curves, --curves-big, --saturation, --saturation-big and --merge-curves are refused with --random-manifest too.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+    labels = manifest["labels"]
+    names = [label["name"] for label in labels]
+    series = labels_cardinality(labels, names)
+    per_sec = manifest["samples_per_sec_per_series"]


isinstance(per_sec, int) accepts a bool: true passes and is exported as rate 1. This divisibility check is also written separately in main()'s --samples-per-sec check; a shared helper would keep the two from drifting.

Fixed in ba319e0: divides_1000() rejects bools and is shared with --samples-per-sec's check.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+        }
+        qs.append(q)
+        shared[key] = synthetic_entry(q)["id"]
+        e2e.append({"promql": query["promql"], "interval_secs": int(t),


Cleanup: synthetic_entry(q) runs inside the loop only to get the id, then again for every q after the loop. Keep the entry built in the loop. ids.count() makes the duplicate-id check O(n²); collections.Counter does it in one pass.

Fixed in ba319e0: each entry is built once in the loop; the duplicate-id check uses collections.Counter.

milindsrivastava1997 · 2026-10-15T02:22:24Z

+        with open(args.random_manifest) as f:
+            manifest = json.load(f)
+        inputs = {os.path.basename(args.random_manifest): sha256(args.random_manifest)}
+        table = random_manifest_table(manifest, args.revision, inputs)


Cleanup: this copies build()'s table-writing block (makedirs, json.dump with indent=1 and sort_keys, trailing newline). Write the table once after both branches so the two can't drift.

Fixed in ba319e0: run() writes the table once after the --random-manifest/build() branch.

- RQEs below the curves' first N (1e3 items per group) are left out (logged), as schema_queries() does. - A sub-second interval is a ValueError, not an assert; a bool rate is refused; divides_1000() is shared with --samples-per-sec. - A repeated query is one e2e entry. - The curve flags are refused with --random-manifest. - Each RQE entry is built once; one table writer for both modes. Seed 862's ten tables are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

milindsrivastava1997

Automated code review: 10 findings (7 correctness, 3 cleanup).

milindsrivastava1997 · 2026-10-15T02:30:35Z

+            if entry not in e2e:
+                e2e.append(entry)
+            continue
+        groups = labels_cardinality(labels, grouping)


The SCHEMA_MIN_COVERED_N check also drops exact sum RQEs, and their e2e queries, even though the exact-sum family reads no accuracy curve. For example, sum by (job,label_0) (sum_over_time(data[60s])) at 10/s has n_items=600 < 1000, so it is removed from both rqes and e2e. Instant sums with a one-scrape lookback are always dropped. Suggest limiting the check to non-exact families.

Fixed in e62a15e: the check now applies only to sketch families; exact accumulators read no curve (test added).

milindsrivastava1997 · 2026-10-15T02:30:35Z

+        if interval_ms <= 0 or interval_ms % 1000:
+            raise ValueError(f"{query['id']}: interval {interval_ms} ms is not whole "
+                             "seconds")
+        t = interval_ms // 1000


lookback_ms is never checked against the scrape period (1000/samples_per_sec ms). At 10/s, a lookback of 1_234_567 ms (as in test_ranges_in_rqe_ids_are_exact) or 150 ms exports fine, then from_table() in autosketch_vs_asap.rs panics on lookback_ms.is_multiple_of(scrape_ms). The exporter should reject this input, as it already does for interval_ms.

Fixed in e62a15e: a lookback that isn't a multiple of the scrape period is a ValueError (test added).

milindsrivastava1997 · 2026-10-15T02:30:35Z

+    for query in manifest["queries"]:
+        if query["support"] != RANDOM_SUPPORT:
+            continue
+        [rqe] = query["rqes"]


[rqe] = query['rqes'] assumes exactly one RQE per common-promql query. A binary or ratio query (e.g. quantile/quantile) or an empty rqes list aborts with ValueError: too many values to unpack and no query id. Suggest skipping with a log line or raising a clear error that names the query.

Fixed in e62a15e: a common-promql query with other than one RQE is a ValueError naming the query (test added).

milindsrivastava1997 · 2026-10-15T02:30:35Z

+RANDOM_STREAMS = {"sum": "value", "quantile": "dist", "topk_by_value": "value"}
+# The fewest distinct keys the top-k curves measure; a heap over fewer keys
+# is read there, which overstates its error.
+TOPK_MIN_CURVE_KEYS = 1_000


The AVG_OF_SERIES_QUANTILES regex requires avg by (...) with a non-empty, ', '-joined label list. avg(quantile_over_time(0.9, data[100s])) with grouping_labels=[], or avg by (job,label_0) with no space, raises 'a quantile must be avg by ...' and the whole export fails.

Holding for the quantile rework (pooled quantiles grouped by G), which replaces this check.

milindsrivastava1997 · 2026-10-15T02:30:35Z

+        if capability == "quantile":
+            m = AVG_OF_SERIES_QUANTILES.match(query["promql"])
+            if not m or m[1].split(", ") != grouping:
+                raise ValueError(f"{query['id']}: a quantile must be avg by "


This check validates only the avg grouping. It never checks that the PromQL's phi or range matches the RQE's phi/lookback_ms, yet RQE sharing is keyed on the RQE fields alone. PromQL quantile_over_time(0.99, data[5m]) with an RQE of phi 0.9 and lookback 100000 exports without error and may share another query's RQE id. The e2e query then reads a different window and quantile than the planner sized.

Holding for the quantile rework, as with the earlier phi/window comment.

milindsrivastava1997 · 2026-10-15T02:30:35Z

+        s = ms / 1000
+        n_items = rate * s / groups
+        # Accuracy is read off the curves, which start at this N.
+        if n_items < SCHEMA_MIN_COVERED_N:


For top-k heaps with fewer than 1,000 keys, K is raised to 1,000 while min_N, groups and theta stay the same. Top-k by (job, label_0) has 1 key per heap but is exported with K=1000, so the plan's cost and error for that RQE are off by orders of magnitude. Only an assumptions string records this.

Intentional: the 500-key heaps of seed 862 read at 1,000 keys is conservative (fewer keys, fewer collisions). Leaving the clamp as is; tiny heaps don't occur in these workloads.

milindsrivastava1997 · 2026-10-15T02:30:35Z

+# --random-manifest: the five-system queries of an ASAPQuery random workload
+# (asap-tools random_workloads.py), and the streams their capabilities read.
+RANDOM_SUPPORT = "common-promql"
+RANDOM_STREAMS = {"sum": "value", "quantile": "dist", "topk_by_value": "value"}


sum and topk_by_value both map to the stream data/value, but autosketch_vs_asap builds one MetricFacts per stream and assumes 'A stream is one value of its metric, so one capability'. With both a sum and a topk at the same grouping, the stream's facts (including hydra_dataset) take the first RQE's capability, and data_shape is merged across capabilities.

Harmless today (metric data has no Hydra dataset, and exact sums give no data shape), but agreed it's cleaner to split: deferring to the quantile rework, which regenerates the archived tables anyway.

milindsrivastava1997 · 2026-10-15T02:30:36Z

+            print(f"left out {query['id']}: {n_items:g} items per group per window, "
+                  f"below the curves' first N ({SCHEMA_MIN_COVERED_N:g})", file=sys.stderr)
+            continue
+        query_id, keys, assumptions = query["id"], None, []


Cleanup: the TOPK_MIN_CURVE_KEYS clamp and the below-curves skip are special cases in the exporter, and each path behaves differently: schema_queries raises, while this path logs and drops. Handling N or K below the first curve point once, in the curve lookup, would remove both copies and keep them from drifting apart.

Agreed in principle, but it's a curve-lookup refactor in the optimizer, out of scope here. Deferring.

milindsrivastava1997 · 2026-10-15T02:30:36Z

+        rqes.append(synthetic_entry(q))
+        shared[key] = rqes[-1]["id"]
+        e2e.append({"promql": query["promql"], "interval_secs": t, "rqes": [shared[key]]})
+    duplicates = sorted(i for i, n in collections.Counter(r["id"] for r in rqes).items()


Cleanup: this duplicate-RQE-id check repeats build()'s existing assert len(ids) == len(set(ids)) and can't be reached through the shared path. One shared helper over rqes would do the job in both places.

Fixed in e62a15e: check_unique_ids() is shared by build() and the random-manifest path.

milindsrivastava1997 · 2026-10-15T02:30:36Z

+        # Queries reading one RQE (per-series quantiles that differ only in
+        # their avg's grouping) share its id.
+        key = (capability, tuple(grouping), ms, t, rqe.get("phi"), rqe.get("k"))
+        if key in shared:


Nit: entry not in e2e is a linear scan over a list of dicts, so the dedup is quadratic. The shared dict already holds the key for an O(1) check. A set keyed on (promql, interval_secs), or reusing e2e_block's unique, would avoid the scan.

Fixed in e62a15e: the dedup uses a set of (promql, interval_secs).

- The curves' first-N check skips exact families, which read no curve. - A lookback that isn't whole scrapes, or a query reading other than one RQE, is a ValueError naming the query. - check_unique_ids() is shared with build(); the e2e dedup is a set. Seed 862's ten tables are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…for-random-manifest-tables Takes #226's shared-RQE e2e entries and gives every entry, including one reading a shared RQE, its tree (e2e_query()). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011gKTCAEz3sJ1Hzde16xxU1

* eval: RQE expression trees in e2e tables Every e2e query gets `expr`, its RQE expression tree: a leaf per RQE it reads (id, metric or table, capability, lookback_ms, grouping, phi/k/value), and a `/` node for template 10. Rate queries, and schema queries merged across several RQEs (--keep-duplicate-queries), get none. Regenerate the e2e-rate10 tables: only `expr` is added, so the RQE ids and plans are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011gKTCAEz3sJ1Hzde16xxU1 * eval: RQE expression trees for random-manifest tables --random-manifest e2e queries carry `expr` (#227, ProjectASAP/ASAPQuery#886): avg by (G) of per-series quantile_over_time is an avg over a per-series quantile leaf; topk by (G) of sum_over_time a per-series top-k leaf with `topk_by` G (left out for a global top-k); sum by (G) of sum_over_time and spatial quantile by (G) a leaf at G; SQL cardinality a leaf on its table with `value`. Counter-gated top-k and rate have none. Leaf ids are the table's RQE ids; values come from the manifest's RQE. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011gKTCAEz3sJ1Hzde16xxU1 * eval: leaf capabilities as a set, renaming only topk Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011gKTCAEz3sJ1Hzde16xxU1 * eval: random-manifest trees from RQE fields, not PromQL random_expr() builds each tree from the manifest RQE's capability, grouping_labels, phi and k, as asap-tools' rqe_query() renders them: quantile is avg by G of a per-series quantile leaf, topk_by_value a per-series top-k leaf ranked within G, sum a leaf at G. No PromQL regexes, so no list-order label comparison; the spatial quantile shape (read as a group quantile from a per-series RQE) and the unreachable SQL cardinality branch go. A test builds seed-862 workload-01's table and checks every tree names its query's RQE. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011gKTCAEz3sJ1Hzde16xxU1 --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

#225's rqe_leaf() conflicts with #228's, which adds topk_by; kept #228's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings phase2-ver3 up to main at a3753f2 (#226). The repository squash-merges pull requests, so main is unchanged and this branch lands as one commit. Conflicts and how they were resolved: - aqpbm-datagen: the branch moved the crate's tests out of lib.rs into the modules they test and replaced mix_connected_bits with splitmix64::finalize; main added child_of, fan_out, scale_by and scale_range (#197), the Pareto distribution and their tests in lib.rs. The moved tests stay in their modules. lib.rs keeps only the 17 tests and helpers main added, all of them, and imports the shared helpers from test_support, which now fills the four new ColumnSpec fields. main's scale_by hash calls splitmix64::finalize, the same function as mix_connected_bits. 50 tests, as on main, none duplicated. - aqpbm-planeval: the branch removed the old in-process evaluator, so main's edit to src/handle.rs (#183) is dropped with it and the branch's Cargo.toml is kept. The harness fills the new ColumnSpec fields. - Cargo.lock: main's lockfile, with the branch's dependencies added. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This was referenced Oct 14, 2026

feat(asap-tools): add accelerated random E2E workload artifacts ProjectASAP/ASAPQuery#879

Merged

eval: RQE expression trees for random-manifest tables #227

Closed

eval: RQE expression trees for random-manifest tables #228

Merged

milindsrivastava1997 commented Oct 15, 2026

View reviewed changes

milindsrivastava1997 and others added 2 commits October 14, 2026 21:31

milindsrivastava1997 commented Oct 15, 2026

View reviewed changes

milindsrivastava1997 and others added 2 commits October 15, 2026 00:22

Merge origin/main into eval/random-manifest

9b89a47

#225's rqe_leaf() conflicts with #228's, which adds topk_by; kept #228's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

milindsrivastava1997 merged commit a3753f2 into main Oct 15, 2026
2 checks passed

This was referenced Oct 15, 2026

eval: eval table from an ASAPQuery random-workload manifest #223

Closed

Tracking: RQE frontend for the paper eval (sketch-bench side) #224

Closed

		TOPK_MIN_CURVE_KEYS = 1_000


		def random_groups(labels, grouping):

Conversation

milindsrivastava1997 commented Oct 14, 2026

Uh oh!

milindsrivastava1997 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

milindsrivastava1997 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment