@@ -7,10 +7,13 @@
 use std::path::PathBuf;
 
 use asap_planner::optimizer::{
-    load_optional_selected_atomic_cost_table, run_greedy_pipeline, AtomicCostTable, LabelSetFacts,
+    build_milp_workload, load_flat_atomic_cost_table, load_optional_selected_atomic_cost_table,
+    load_workload_facts, run_greedy_pipeline, solve_milp, AtomicCostTable, LabelSetFacts,
+    LabelSetFactsError,
 };
 use asap_planner::ControllerConfig;
 use clap::Parser;
+use rqe_optimizer::milp::Objective;
 
 #[derive(Parser, Debug)]
 #[command(
@@ -27,26 +30,57 @@ struct Args {
     #[arg(long = "data-ingestion-interval-ms", value_parser = clap::value_parser!(u64).range(1..))]
     data_ingestion_interval_ms: u64,
 
-    /// YAML label-set facts: `series_count` per (metric, spatial filter) and
-    /// `cardinality` per (metric, spatial filter, grouping labels).
-    #[arg(long = "label-set-facts")]
-    label_set_facts: PathBuf,
-
-    /// Path to the versioned atomic-cost document sketch-bench's `atomic-costs`
-    /// subcommand exports. Requires --atomic-cost-workload to select exactly
-    /// one measured workload profile. Omitted: every
-    /// benchmarked-family candidate (CMS/HLL/KLL) is dropped, since there is
-    /// no data to cost it at — only trivial accumulators and EXACT remain
-    /// selectable.
-    #[arg(long = "atomic-costs")]
+    /// Greedy only. YAML label-set facts: `series_count` per (metric, spatial
+    /// filter) and `cardinality` per (metric, spatial filter, grouping labels).
+    #[arg(
+        long = "label-set-facts",
+        required_unless_present = "milp",
+        conflicts_with = "milp"
+    )]
+    label_set_facts: Option<PathBuf>,
+
+    /// Greedy: the versioned atomic-cost document sketch-bench's
+    /// `atomic-costs` subcommand exports; requires --atomic-cost-workload.
+    /// Omitted: every benchmarked-family candidate (CMS/HLL/KLL) is dropped,
+    /// leaving only trivial accumulators and EXACT.
+    /// MILP: the flat cost table `export_rqe_optimizer_costs.sh` writes
+    /// (`rqe_atomic_costs.json`); required.
+    #[arg(long = "atomic-costs", required_if_eq("milp", "true"))]
     atomic_costs: Option<PathBuf>,
 
-    /// JSON `profiles[].workload` value copied from the sketch-bench atomic-cost
-    /// document. This makes the empirical workload profile explicit and avoids
-    /// mixing costs from different traces or time windows.
-    #[arg(long = "atomic-cost-workload", requires = "atomic_costs")]
+    /// Greedy only. JSON `profiles[].workload` value copied from the
+    /// sketch-bench atomic-cost document. This makes the empirical workload
+    /// profile explicit and avoids mixing costs from different traces or time
+    /// windows.
+    #[arg(
+        long = "atomic-cost-workload",
+        requires = "atomic_costs",
+        conflicts_with = "milp"
+    )]
     atomic_cost_workload: Option<PathBuf>,
 
+    /// Plan with sketch-bench's rqe-optimizer MILP and print the plan; writes
+    /// no configs yet.
+    #[arg(long)]
+    milp: bool,
+
+    /// MILP only. YAML workload facts: per metric, `cardinality` per label
+    /// set, including the set of all its labels (the series count).
+    #[arg(
+        long = "workload-facts",
+        required_if_eq("milp", "true"),
+        requires = "milp"
+    )]
+    workload_facts: Option<PathBuf>,
+
+    /// MILP only. Objective weight on CPU-sec/sec. Default: rqe-optimizer's.
+    #[arg(long = "w-cpu", requires = "milp", value_parser = parse_weight)]
+    w_cpu: Option<f64>,
+
+    /// MILP only. Objective weight on memory GiB. Default: rqe-optimizer's.
+    #[arg(long = "w-mem", requires = "milp", value_parser = parse_weight)]
+    w_mem: Option<f64>,
+
     #[arg(short, long, action = clap::ArgAction::Count)]
     verbose: u8,
 }
@@ -64,7 +98,14 @@ fn main() -> anyhow::Result<()> {
 
     let yaml_str = std::fs::read_to_string(&args.input_config)?;
     let config: ControllerConfig = serde_yaml::from_str(&yaml_str)?;
-    let facts = LabelSetFacts::from_path(&args.label_set_facts)?;
+    if args.milp {
+        return run_milp(&args, &config);
+    }
+    let facts = LabelSetFacts::from_path(
+        args.label_set_facts
+            .as_deref()
+            .expect("clap requires --label-set-facts without --milp"),
+    )?;
 
     let atomic_cost_table = match load_optional_selected_atomic_cost_table(
         args.atomic_costs.as_deref(),
@@ -108,3 +149,105 @@ fn main() -> anyhow::Result<()> {
 
     Ok(())
 }
+
+/// Objective weights must be finite and non-negative: a negative weight
+/// rewards cost, and NaN poisons every coefficient.
+fn parse_weight(s: &str) -> Result<f64, String> {
+    match s.parse::<f64>() {
+        Ok(w) if w.is_finite() && w >= 0.0 => Ok(w),
+        Ok(w) => Err(format!("must be finite and >= 0, got {w}")),
+        Err(e) => Err(e.to_string()),
+    }
+}
+
+fn run_milp(args: &Args, config: &ControllerConfig) -> anyhow::Result<()> {
+    config.warn_default_slas();
+    let Some(hints) = config.metrics.as_deref() else {
+        return Err(LabelSetFactsError::MissingMetricHints.into());
+    };
+    let facts = load_workload_facts(
+        args.workload_facts
+            .as_deref()
+            .expect("clap requires --workload-facts with --milp"),
+        hints,
+        args.data_ingestion_interval_ms,
+    )?;
+    let costs = load_flat_atomic_cost_table(
+        args.atomic_costs
+            .as_deref()
+            .expect("clap requires --atomic-costs with --milp"),
+    )?;
+    let Objective::AUCCost { w_cpu, w_mem } = Objective::default();
+    let (w_cpu, w_mem) = (args.w_cpu.unwrap_or(w_cpu), args.w_mem.unwrap_or(w_mem));
+    // All-zero weights make every plan cost 0, so the solver's pick is arbitrary.
+    anyhow::ensure!(
+        w_cpu > 0.0 || w_mem > 0.0,
+        "--w-cpu and --w-mem are both 0; at least one must be positive"
+    );
+    let objective = Objective::AUCCost { w_cpu, w_mem };
+    tracing::debug!(?objective, cost_rows = costs.len(), "milp: inputs loaded");
+
+    let workload = build_milp_workload(config, &facts, args.data_ingestion_interval_ms)?;
+    let (deployments, solution) = solve_milp(&workload, &facts, &costs, objective)?;
+
+    let mut active: Vec<usize> = solution.mapping.clone();
+    active.sort();
+    active.dedup();
+    println!("=== Deployments: {} ===", active.len());
+    for &d in &active {
+        let dep = &deployments[d];
+        println!(
+            "  [{d}] {:?} {} config={} metric={} grouping={:?} window={}ms slide={}ms",
+            dep.capability,
+            dep.config.sketch,
+            dep.config.sketch_config,
+            dep.metric,
+            dep.grouping_labels,
+            dep.window_ms,
+            dep.slide_ms,
+        );
+    }
+    println!("\n=== Raqes: {} ===", workload.raqes.len());
+    for ((raqe, &d), latency_ms) in workload
+        .raqes
+        .iter()
+        .zip(&solution.mapping)
+        .zip(&solution.plan_cost.query_latency_ms)
+    {
+        println!("  {} -> [{d}] latency={latency_ms:.3e}ms", raqe.id);
+    }
+    let cost = &solution.plan_cost;
+    println!(
+        "\nobjective={:.6e} cpu={:.6e} cpu-sec/sec memory={:.3} MB",
+        objective.value(cost),
+        cost.cpu_secs_per_sec(),
+        cost.memory_bytes() / 1e6,
+    );
+    for (phase, c) in [
+        ("ingest", &cost.ingest),
+        ("merge", &cost.merge),
+        ("query", &cost.query),
+        ("storage", &cost.storage),
+    ] {
+        println!(
+            "  {phase}: cpu={:.6e} cpu-sec/sec memory={:.3} MB",
+            c.cpu_secs_per_sec,
+            c.memory_bytes / 1e6
+        );
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::parse_weight;
+
+    #[test]
+    fn weights_must_be_finite_and_non_negative() {
+        assert_eq!(parse_weight("0.5"), Ok(0.5));
+        assert_eq!(parse_weight("0"), Ok(0.0));
+        for bad in ["-1", "NaN", "inf", "x"] {
+            assert!(parse_weight(bad).is_err(), "{bad}");
+        }
+    }
+}
@@ -67,7 +67,8 @@ pub fn extract_aqes(
     metric_schema: &PromQLSchema,
     scrape_interval_ms: u64,
 ) -> Result<Vec<OptimizerItem>, OptimizerError> {
-    let mut acc: HashMap<OptimizerItemKey, (QueryRequirements, Vec<String>, f64)> = HashMap::new();
+    let mut acc: HashMap<OptimizerItemKey, (QueryRequirements, Vec<String>, usize)> =
+        HashMap::new();
 
     for rqe in rqes {
         if rqe.t_repeat_ms == 0 {
@@ -82,13 +83,11 @@ pub fn extract_aqes(
             match extract_requirements(&leaf, metric_schema, scrape_interval_ms) {
                 Ok(req) => {
                     let key = OptimizerItemKey::from_rqe(&req, rqe);
-                    let entry = acc.entry(key).or_insert_with(|| (req, Vec::new(), 0.0));
+                    let entry = acc.entry(key).or_insert_with(|| (req, Vec::new(), 0));
                     if !entry.1.contains(&leaf) {
                         entry.1.push(leaf);
                     }
-                    // query_frequency_hz must stay in Hz (queries per real second)
-                    // regardless of t_repeat_ms's internal unit — 1000.0 / ms, not 1.0 / ms.
-                    entry.2 += 1000.0 / rqe.t_repeat_ms as f64;
+                    entry.2 += 1;
                 }
                 Err(reason) => {
                     return Err(OptimizerError::UnsupportedLeaf {
@@ -104,10 +103,12 @@ pub fn extract_aqes(
     Ok(acc
         .into_iter()
         .map(
-            |(key, (requirements, query_strings, query_frequency_hz))| OptimizerItem {
+            |(key, (requirements, query_strings, occurrences))| OptimizerItem {
                 requirements,
                 query_strings,
-                query_frequency_hz,
+                // Hz (queries per real second): 1000.0 / ms, not 1.0 / ms.
+                query_frequency_hz: occurrences as f64 * 1000.0 / key.t_repeat_ms as f64,
+                occurrences,
                 t_repeat_ms: key.t_repeat_ms,
                 accuracy_sla: f64::from_bits(key.accuracy_sla_bits),
                 latency_sla_ms: key.latency_sla_ms_bits.map(f64::from_bits),