solverforge-scoring 0.8.3

// LoadBalance collector for computing unfairness (sqrt of variance).

use std::collections::HashMap;
use std::hash::Hash;
use std::marker::PhantomData;

use super::{Accumulator, UniCollector};

/* Result of load balancing - tracks loads per item and computes unfairness.

Unfairness is the square root of sum of squared deviations from mean.
Lower values indicate fairer distribution. Zero means perfectly balanced.
*/
#[derive(Debug, Clone)]
pub struct LoadBalance<K> {
    loads: HashMap<K, i64>,
    // Unfairness as integer (use with of_soft() for scoring)
    unfairness: i64,
}

impl<K> LoadBalance<K> {
    // Returns map of balanced items to their total load.
    pub fn loads(&self) -> &HashMap<K, i64> {
        &self.loads
    }

    // Returns unfairness as i64 for use with `of_soft()`.
    // This is the raw value - `of_soft()` handles decimal scaling.
    #[inline]
    pub fn unfairness(&self) -> i64 {
        self.unfairness
    }
}

/* Creates a load balance collector.

# Example

```
use solverforge_scoring::stream::collector::{load_balance, UniCollector, Accumulator};

struct Shift { employee_id: usize }

let collector = load_balance(
|s: &Shift| s.employee_id,
|_s: &Shift| 1i64,  // Each shift counts as 1
);

let mut acc = collector.create_accumulator();
acc.accumulate(&collector.extract(&Shift { employee_id: 0 }));
acc.accumulate(&collector.extract(&Shift { employee_id: 0 }));
acc.accumulate(&collector.extract(&Shift { employee_id: 1 }));

let result = acc.finish();
// Employee 0 has 2, Employee 1 has 1 → unfairness = sqrt(0.5) ≈ 1
assert_eq!(result.unfairness(), 1);
```
*/
pub fn load_balance<A, K, F, M>(key_fn: F, metric_fn: M) -> LoadBalanceCollector<A, K, F, M>
where
    K: Clone + Eq + Hash + Send + Sync,
    F: Fn(&A) -> K + Send + Sync,
    M: Fn(&A) -> i64 + Send + Sync,
{
    LoadBalanceCollector {
        key_fn,
        metric_fn,
        _phantom: PhantomData,
    }
}

// Collector for computing load balance unfairness.
pub struct LoadBalanceCollector<A, K, F, M> {
    key_fn: F,
    metric_fn: M,
    _phantom: PhantomData<fn(&A) -> K>,
}

impl<A, K, F, M> UniCollector<A> for LoadBalanceCollector<A, K, F, M>
where
    A: Send + Sync,
    K: Clone + Eq + Hash + Send + Sync,
    F: Fn(&A) -> K + Send + Sync,
    M: Fn(&A) -> i64 + Send + Sync,
{
    type Value = (K, i64);
    type Result = LoadBalance<K>;
    type Accumulator = LoadBalanceAccumulator<K>;

    #[inline]
    fn extract(&self, entity: &A) -> Self::Value {
        ((self.key_fn)(entity), (self.metric_fn)(entity))
    }

    fn create_accumulator(&self) -> Self::Accumulator {
        LoadBalanceAccumulator::new()
    }
}

// Accumulator for load balance with incremental variance computation (O(1) updates).
pub struct LoadBalanceAccumulator<K> {
    // Count of items per balanced key (for duplicate tracking)
    item_counts: HashMap<K, usize>,
    // Cumulative load per balanced key
    loads: HashMap<K, i64>,
    // Sum of all loads
    sum: i64,
    // Integral part of squared deviation
    squared_deviation_integral: i64,
    // Fractional numerator for incremental variance
    squared_deviation_fraction_numerator: i64,
}

impl<K: Clone + Eq + Hash> LoadBalanceAccumulator<K> {
    fn new() -> Self {
        Self {
            item_counts: HashMap::new(),
            loads: HashMap::new(),
            sum: 0,
            squared_deviation_integral: 0,
            squared_deviation_fraction_numerator: 0,
        }
    }

    fn add_to_metric(&mut self, key: &K, diff: i64) {
        let old_value = *self.loads.get(key).unwrap_or(&0);
        let new_value = old_value + diff;

        if old_value != new_value {
            self.loads.insert(key.clone(), new_value);
            self.update_squared_deviation(old_value, new_value);
            self.sum += diff;
        }
    }

    fn reset_metric(&mut self, key: &K) {
        if let Some(old_value) = self.loads.remove(key) {
            if old_value != 0 {
                self.update_squared_deviation(old_value, 0);
                self.sum -= old_value;
            }
        }
    }

    // Incremental variance update formula.
    fn update_squared_deviation(&mut self, old_value: i64, new_value: i64) {
        // Term 1: x_new² - x_old²
        let term1 = new_value * new_value - old_value * old_value;

        // Term 2: 2 * (sum_others) * (sum_old - sum_new)
        let sum_others = 2 * (self.sum - old_value);
        let new_sum = self.sum - old_value + new_value;
        let sum_diff = self.sum - new_sum;

        // Term 3: sum_new² - sum_old²
        let term3 = new_sum * new_sum - self.sum * self.sum;

        // Term 4: 2 * (old*sum_old - new*sum_new)
        let term4 = 2 * (old_value * self.sum - new_value * new_sum);

        let fraction_delta = sum_others * sum_diff + term3 + term4;

        self.squared_deviation_integral += term1;
        self.squared_deviation_fraction_numerator += fraction_delta;
    }

    // Returns unfairness as i64 (no pre-scaling).
    // Use with `of_soft()` which handles the decimal scaling.
    fn compute_unfairness(&self) -> i64 {
        let n = self.item_counts.len();
        match n {
            0 => 0,
            1 => {
                // For n=1: sqrt(fraction + integral)
                let tmp = self.squared_deviation_fraction_numerator as f64
                    + self.squared_deviation_integral as f64;
                tmp.sqrt().round() as i64
            }
            _ => {
                // For n>1: sqrt(fraction/n + integral)
                let tmp = (self.squared_deviation_fraction_numerator as f64 / n as f64)
                    + self.squared_deviation_integral as f64;
                tmp.sqrt().round() as i64
            }
        }
    }
}

impl<K: Clone + Eq + Hash + Send + Sync> Accumulator<(K, i64), LoadBalance<K>>
    for LoadBalanceAccumulator<K>
{
    #[inline]
    fn accumulate(&mut self, value: &(K, i64)) {
        let (key, metric) = value;
        if *metric == 0 {
            return; // Skip zero-metric entries (e.g., unassigned shifts)
        }
        let count = self.item_counts.entry(key.clone()).or_insert(0);
        *count += 1;
        self.add_to_metric(key, *metric);
    }

    #[inline]
    fn retract(&mut self, value: &(K, i64)) {
        let (key, metric) = value;
        if *metric == 0 {
            return; // Skip zero-metric entries
        }
        if let Some(count) = self.item_counts.get_mut(key) {
            if *count > 0 {
                *count -= 1;
                if *count == 0 {
                    self.item_counts.remove(key);
                    self.reset_metric(key);
                } else {
                    self.add_to_metric(key, -*metric);
                }
            }
        }
    }

    fn finish(&self) -> LoadBalance<K> {
        LoadBalance {
            loads: self.loads.clone(),
            unfairness: self.compute_unfairness(),
        }
    }

    fn reset(&mut self) {
        self.item_counts.clear();
        self.loads.clear();
        self.sum = 0;
        self.squared_deviation_integral = 0;
        self.squared_deviation_fraction_numerator = 0;
    }
}