Struct ScalingPolicy 

pub struct ScalingPolicy {
    pub fill_ratio_threshold: f64,
    pub push_latency_threshold_ns: u64,
    pub flush_latency_threshold_us: u64,
    pub min_shards: u16,
    pub max_shards: u16,
    pub cooldown: Duration,
    pub scale_down_delay: Duration,
    pub underutilized_threshold: f64,
    pub metrics_window: Duration,
    pub auto_scale: bool,
}

Policy configuration for dynamic shard scaling.

Fields

fill_ratio_threshold: f64

Fill ratio threshold to trigger scale-up (0.0 - 1.0). Default: 0.7 (70%)

push_latency_threshold_ns: u64

Push latency threshold in nanoseconds to trigger scale-up. Default: 5ns (after this, we’re seeing contention)

flush_latency_threshold_us: u64

Batch flush latency threshold in microseconds to trigger scale-up. Default: 1000μs (1ms)

min_shards: u16

Minimum number of shards (floor for scaling down).

max_shards: u16

Maximum number of shards (ceiling for scaling up).

cooldown: Duration

Cooldown period between scaling operations. Default: 1 second

scale_down_delay: Duration

How long a shard must be underutilized before scaling down. Default: 10 seconds

underutilized_threshold: f64

Fill ratio below which a shard is considered underutilized. Default: 0.1 (10%)

metrics_window: Duration

Metrics collection window. Default: 100ms

auto_scale: bool

Enable automatic scaling (if false, scaling is manual only).

Implementations

impl ScalingPolicy

pub fn default_for_cpus(cpus: u16) -> Self

Create a default scaling policy based on CPU count. Scales from 1 shard up to the number of physical cores.

pub fn high_throughput() -> Self

Create a policy optimized for high-throughput GPU workloads. Uses more aggressive scaling with higher max shard count.

max_shards is capped at u16::MAX (65 535) because shard ids are 16-bit. On hosts with more than 32 767 CPUs the “2× CPU count” target saturates rather than wraps — this is the intended behavior (pre-fix this was just an implicit saturating_mul artifact; the cap is now documented and the saturation is explicit).

pub fn conservative() -> Self

Create a conservative policy for stable workloads.

pub fn normalize(self) -> Self

Normalize the policy by auto-adjusting conflicting values.

This allows users to set either min_shards or max_shards independently without worrying about the other. If max_shards < min_shards, max_shards is adjusted to equal min_shards.

pub fn validate(&self) -> Result<(), ConfigError>

Validate the policy.

is_finite() guards reject NaN and ±∞ explicitly before the range check runs. NaN thresholds slip past raw <= / > comparisons (every comparison against f64::NaN returns false), so a config deserialized from 0.0/0.0-style arithmetic or an unfortunate environment-templated string would “validate” successfully and then sit inert at runtime — mapper.rs:560 does m.fill_ratio > self.policy.fill_ratio_threshold, which is always false against NaN, so the scaler would never fire (mirror hazard for scale-down).

Trait Implementations

impl Clone for ScalingPolicy

fn clone(&self) -> ScalingPolicy

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ScalingPolicy

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ScalingPolicy

fn default() -> Self

Returns the “default value” for a type.