rqe-optimizer: add AutoSketch's per-query configuration search as a baseline

[zzylol]

Before this PR

rqe-optimizer had only the batch planner, brute-force enumeration and the HiGHS MILP, with no AutoSketch baseline to compare it against. The only port of AutoSketch's search (NSDI '24, Algorithm 4) was in ASAPQuery-backend (#547, data_plane/examples/autosketch_comparison.rs). It was hardcoded to a CMS width/depth grid.

After this PR

rqe_optimizer::autosketch adds AutoSketch-Adapted: AutoSketch's configuration search, run once per RQE over the measured cost table. This is the AutoSketch baseline in the evaluation plan, ProjectASAP/ASAPQuery#777 (§3).

pub fn window_adapter(rqe: &Rqe) -> (Seconds, Seconds);
pub fn search(rqe: &Rqe, costs: &[AtomicCostEntry], seed: u64,
              accuracy: impl FnMut(&Rqe, &AtomicCostEntry) -> Option<f64>) -> RqeSearch;
pub fn plan(rqes: &[Rqe], costs: &[AtomicCostEntry], seed: u64,
            accuracy: impl FnMut(&Rqe, &AtomicCostEntry) -> Option<f64>)
    -> Result<AutoSketchPlan, Vec<String>>; // Err = IDs of unservable RQEs

• Search (search): Algorithm 4 for one RQE, over the measured configurations of every sketch variant that serves its capability.
• Objective: minimize memory per instance, with insert CPU breaking ties. This is AutoSketch's resource score SC with no ALU term.
• Accuracy: the caller passes the accuracy oracle; the evaluation passes the measured value at the query window's size. A configuration is feasible if its value meets the RQE's target, and a missing measurement is infeasible.
• Plan (plan): gives each RQE its own deployment, never shared, even when two RQEs pick the same one, so objectives::score charges ingest per RQE. The mapping is the identity. Latency is not considered.
• Window (window_adapter): one sliding sketch per query. The window is the lookback S; the slide is the repeat interval T, or gcd(S, T) when T does not divide S. A query reads one instance and never merges.
• Per-RQE output: the selected configuration, the probes (distinct configurations evaluated, in order) and search wall time. The evaluation charges each probe's benchmark time to AutoSketch's planning time.

How it maps to the paper

Paper (Algorithm 4; App. G, H)	This PR
LHS initialization (Alg. 5): distinct values on every dimension	Same, per variant, over the measured parameter values. A variant with one parameter (KLL k, DDSketch alpha, HLL lg_k) gets one random seed, because the paper's rule would sample every value, which is exhaustive. The cheapest configuration is seeded only if no LHS point was measured.
Evaluate, then generate neighbors toward fewer resources if feasible and more if not	Same
CALC_NEIGHBOR (Alg. 6): rows ±1, columns ×2 or ÷2	Move one parameter to its adjacent measured value; direction is judged by measured memory, so alpha (smaller means more memory) works
EXAMINE: (1) skip evaluated configs, (2) skip those over the hardware limit, (3) skip those no cheaper than a known-feasible config	(1) and (3) as in the paper. (2) is dropped: software has no stage/page limits.
Stopping: a path from a feasible start stops shrinking at the first failure; one from an infeasible start stops at the first success	Same, per path
No feasible config: Algorithm 4 returns the most accurate; §5.2 says it throws an error	Follows §5.2: the RQE is reported as unservable
The compiler maps an operator to one sketch, and the search tunes its parameters	Extension: searches every variant for the capability and keeps the cheapest, so it chooses from the same sketches as the MILP

Validation

• cargo test -p rqe-optimizer: 22 passed, including these 10 for this module:
  • picks the smallest feasible configuration;
  • probes fewer configurations than exhaustive search on an 8×8 grid, never the same one twice;
  • two identical RQEs get two deployments, and score charges ingest twice;
  • adapted windows (T | S and the gcd case) pass candidates::is_eligible;
  • an unservable RQE is reported by ID;
  • a missing measurement is infeasible;
  • on a single axis the search stops at the feasibility boundary from either side;
  • it chooses the cheaper of two variants;
  • a higher-is-better metric is treated as a floor;
  • LHS samples take distinct values on every axis.
• cargo clippy -p rqe-optimizer --all-targets -- -D warnings and cargo fmt --check -p rqe-optimizer are clean.

🤖 Generated with Claude Code