[zzylol]

Being rerun. The stack below now runs the sketch, exact and raw pipelines side by side (#25), and experiments.py has pipelines instead of inline-cost. The results below are from the earlier setup; this PR will be updated with the pipelines and accuracy-size reruns.

Part 10 of the telemetry-functions stack: the measured results of experiments 1 (inline cost) and 2 (accuracy versus size), generated by report_experiments.py from experiments.py output (#30). Every stage is a pinned asap-collector container on 56-core x86 servers, with 10-second windows. Full tables and CSVs are in docs/results/experiments.md.

Inline cost: branch CPU with and without each function

Function	Added branch CPU per point	Branch CPU at 200,000 points/s per source
none (forwarding only)	(about 0.3 µs in total)	0.06 cores
KLL latency quantiles	2.7–4.1 µs	0.60 cores
Count-Min + heap, top-k	3.7–5.3 µs	0.81 cores
HLL over 4 attribute keys	9–11 µs (about 2.3 µs per key)	one core reached at about 100,000 points/s

• The global stages (merge, detector) receive a few KB to 180 KB per window and use under 0.01 cores.
• The exact baseline, which receives every raw point, needs a full core at about 95,000 points/s for cardinality and about 400,000 points/s for latency quantiles.
• Answers arrive 2–3 s after a window ends (these runs used a 2-second merge grace).
• These per-point costs put a branch below the ~1M points/s per core reference. The next step is a profile of the per-point path; decoding OTAP into observations and the per-attribute split are the likely costs, not yet measured.

Accuracy versus size (one window per setting)

Setting	Error	Merge bytes per window
KLL k = 50 / 100 / 200 / 400	p99 median error 13.3% / 7.0% / 7.5% / 2.0% (routes ≥ 20,000 points)	39 / 58 / 96 / 181 KB
Count-Min 256 / 512 / 1,024 / 2,048 columns, 100,000 keys	recall@10 1.0 at every width; top-10 count error 6.0% / 2.8% / 1.3% / 0.46%	7 / 11 / 12 / 19 KB

Single windows, so neighbouring settings can cross (KLL k = 100 vs 200).

[zzylol]

Taken out of the stack and converted to draft. Review found the results come from tooling and a setup that the stack replaces: the forwarding-only raw branch is gone (store and query is now raw-sink-a/b -> loki-raw + query job) and merges now emit as soon as both sources have sent a window. The PRs above now sit directly on main. This PR will be rerun on main after they merge, with the inline-cost baseline redefined and the review points (merge/detector CPU columns, delivered rate from full windows, generator vs exact bottleneck, split Count-Min/KLL tables) addressed.