Struct TrafficStats 

pub struct TrafficStats {

    pub duration_s: f64,
    pub num_req: usize,
    pub avg_isl: f64,
    pub avg_osl: f64,
    pub avg_ttft_ms: f64,
    pub avg_itl_ms: f64,
    pub shape_count: usize,
    pub ttft_count: usize,
    pub itl_count: usize,
    pub avg_accept_length: Option<f64>,
    pub avg_kv_hit_rate: f64,
    pub hit_rate_count: usize,
    pub accept_length_forward_count: usize,
}

Accumulated traffic statistics returned by [TrafficAccumulator::drain].

IMPORTANT: When fields here are added or renamed, update the PyO3 binding in lib/bindings/python/rust/llm/replay.rs (drain_traffic method) so the exported JSON dict matches. The Python adapter in replay_adapter.py reads these keys by name.

Fields

duration_s: f64

num_req: usize

Requests offered to the replay runtime during the window. This matches the live planner’s requests_started_total demand signal.

avg_isl: f64

avg_osl: f64

avg_ttft_ms: f64

avg_itl_ms: f64

shape_count: usize

Completed, non-rejected requests behind avg_isl and avg_osl.

ttft_count: usize

Completed requests behind avg_ttft_ms.

itl_count: usize

Completed requests behind avg_itl_ms.

avg_accept_length: Option<f64>

Mean visible tokens produced per decode request-forward, including the base token. None means the window had no decode forwards.

avg_kv_hit_rate: f64

Mean prefix-cache hit rate (0.0-1.0) across router admissions in the window, computed as mean(overlap_blocks / isl_blocks) over admitted requests (i.e. the arithmetic mean of per-request ratios). Matches the semantics of the real router’s dynamo_component_router_kv_hit_rate Prometheus histogram, which observes one overlap/isl sample per request; the PromQL query sum(increase(_sum)) / sum(increase(_count)) returns the arithmetic mean of those samples, independent of per-request ISL size.

hit_rate_count: usize

Number of samples behind avg_kv_hit_rate (its denominator: router admissions with isl_blocks > 0). Carried so a consumer that merges several drained windows can reconstruct the exact sample-weighted mean rather than approximating it with num_req.

accept_length_forward_count: usize

Number of decode request-forwards behind avg_accept_length (its denominator). Same purpose as hit_rate_count for cross-window merges.

Trait Implementations

impl Clone for TrafficStats

fn clone(&self) -> TrafficStats

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TrafficStats

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

Formats the value using the given formatter.