feat: series-aggregation SDK benchmark (--series-per-sketch)

[zzylol]

Summary

• Adds --series-per-sketch N flag to e2esdkbench so workers aggregate N series into one sketch before export (using a coarser group.id attribute instead of per-series series.id). When N ≥ total series or N=0, all series collapse into a single sketch.
• Adds SeriesPerSketch field to the summary JSON/struct so results carry the grouping parameter.
• Output filenames include _grpN_ suffix when N≠1 (e.g. ddsketch_grp100_50000mps_summary.json) preserving backward compatibility with existing bench_sdk_e2e.sh runs.
• New script otel_collector_benchmark/bench_series_agg.sh sweeps over configurable --groups values against a single baseline run, always using the no-op collector (nopcol). Produces per-run timeseries CSVs, summary JSONs, collector resource CSVs, and a combined aggregate_summary.csv with a bw_reduction_vs_baseline_pct column.

Benchmark Results

Full e2e run: 50,000 MPS · 1,000 series · 60s duration · nopcol collector

Bandwidth (avg KB/s) and reduction vs raw baseline

Sketch	grp1 (no agg)	grp10	grp100	grp1000
baseline (raw)	2,586 KB/s — 0%	—	—	—
ddsketch	2,042 KB/s — 24.8%	231 KB/s — 91.5%	24 KB/s — 99.1%	3.2 KB/s — 99.9%
kll	1,127 KB/s — 56.4%	128 KB/s — 95.0%	14.9 KB/s — 99.4%	2.4 KB/s — 99.9%
hll	15,843 KB/s — −513%	1,586 KB/s — 38.6%	160 KB/s — 93.8%	17 KB/s — 99.3%

SDK CPU % at each group size

Sketch	grp1	grp10	grp100	grp1000
baseline (raw)	97.5%	—	—	—
ddsketch	47.3%	33.7%	32.2%	17.4%
kll	35.7%	26.5%	27.9%	10.5%
hll	39.2%	22.0%	21.0%	10.6%

Avg heap (MB) at each group size

Sketch	grp1	grp10	grp100	grp1000
baseline (raw)	22 MB	—	—	—
ddsketch	22.6 MB	3.5 MB	2.6 MB	2.3 MB
kll	14.1 MB	3.3 MB	2.5 MB	2.5 MB
hll	92.4 MB	9.2 MB	2.8 MB	2.5 MB

Key observations

• KLL grp1 already saves 56% bandwidth vs raw with no grouping — smaller sketch encoding than ddsketch at this cardinality.
• HLL grp1 is counterproductive (+513% BW, 92 MB heap) — HLL registers are large relative to a raw sample at 1,000 series; aggregation only pays off at grp10+.
• Grouping just 10 series per sketch cuts bandwidth by 88–95% across all sketch types.
• All sketches converge to near-zero bandwidth at grp1000, with SDK CPU dropping to ~10%.

Test plan

• go build ./cmd/e2esdkbench passes
• Full benchmark run completed for all sketch types (ddsketch, kll, hll) at 50k MPS, groups 1/10/100/1000
• Confirm existing bench_sdk_e2e.sh behaviour unchanged (no --series-per-sketch flag → default 1 → original filenames)

🤖 Generated with Claude Code