[milindsrivastava1997]

Adds rqe-optimizer/examples/workload_scenarios.rs: sweeps that show how the MILP's optimal deployment moves as the query workload changes. Each sweep steps one knob of a base workload; each step is solved with the MILP and printed as the deployments it runs, the queries each one serves, and per-query merges, latency and achieved accuracy vs SLA.

cargo run --release -p rqe-optimizer --example workload_scenarios -- \
  --saturation-dir <study dir with out_grid_1e7_cost/, out_1e9/ and optimizer_cost/> \
  [--sweep lookback-interval|mixed-grouping|mixed-type|latency-sla|accuracy-sla] \
  [--w-cpu X --w-mem Y] [--allow-undeployable-families]

Sweeps

Sweep	Workload	Knob
lookback-interval	10 p99 queries on 10 distinct metrics	(lookback, interval) ∈ {5m,1h,6h,1d,7d} × {15s,1m,5m}, shared by all 10
mixed-grouping	1h p99 by (service) and by (service, endpoint)	each alone, then both
mixed-type	1h rate, p99, distinct count, top-k	each alone, then together
latency-sla	p99 over 1h/6h/1d of one metric	p99_1h latency SLA ∈ {none, 5, 2, 1, 0.5, 0.2} ms
accuracy-sla	p99 over 1h/6h/1d; distinct count over 1h/6h/1d	1h query's SLA: rank error ∈ {0.05…0.001}; relative error ∈ {0.1…0.01}

Results on out_10_6_26_1342 (w_cpu=1, w_mem=0), before rebasing onto #173/#174

That export no longer parses (rows now need accuracy_metric). The example is rebased and builds against current main but has not been rerun; these results will be refreshed from a study_saturation.py --phase optimizer-cost run.

• lookback-interval: always KLL k=200; window x grows with lookback (1m at 5m → 35m at 7d), slide = interval, memory ∝ lookback/interval (7d @ 15s ≈ 129 GB for 10 queries).
• latency-sla: p99_1h/6h/1d share one KLL deployment throughout. Tightening p99_1h does not split it off; the shared window grows instead (15m → 30m at 1 ms → 1h at 0.5 ms), raising CPU 1.45e-3 → 3.18e-3. 0.2 ms is infeasible.
• accuracy-sla, distinct count: shared HLL lg_k=12 until distinct_1h's relative-error SLA reaches 0.01; then distinct_1h splits onto its own HLL lg_k=14 and 6h/1d stay on lg_k=12.
• accuracy-sla, p99: KLL k=200 gives rank error 0.0026 at 1h; at ≤ 0.002 p99_1h is unservable with this export (no k=800 row).
• mixed-type: top-k is unservable with this export (cms-heap cols=2048 is not in the saturation grid).

Not done

• Rerun every sweep on a current study directory (curves + optimizer_cost/) and refresh the results above.
• Cost-table vs saturation-curve mismatches still warn instead of panic; switch to a panic like small_problem now that both come from one study.
• After rqe-optimizer: let mergeable finer-group deployments serve coarser RAQEs #187 (finer groupings serve coarser RAQEs), mixed-grouping stops being a baseline.

🤖 Generated with Claude Code