Struct BrokerReport 

pub struct BrokerReport {

    pub steady_state_policy: SteadyStatePolicy,
    pub monitoring_stopped_after_convergence: bool,
    pub controller_samples: u64,
    pub controller_elapsed: Duration,
    pub controller_cpu_nanos: Option<u64>,
    pub controller_cpu_accounting_failures: usize,
    pub steady_probe_interval: Duration,
    pub opening_bracket: OpeningBracketReport,
    pub consumer_trajectory: Vec<usize>,
    pub producer_trajectory: Vec<usize>,
    pub actual_consumer_trajectory: Vec<usize>,
    pub actual_producer_trajectory: Vec<usize>,
    pub trajectory_elapsed: Vec<Duration>,
    pub cap_trajectory: Vec<CapObservation>,
    pub throughput_trace: Vec<f64>,
    pub throughput_trace_dropped: usize,
    pub trace_samples_dropped: usize,
    pub moves: usize,
    pub reverts: usize,
    pub nonlinear_probes: usize,
    pub nonlinear_probe_improvements: usize,
    pub nonlinear_override: bool,
    pub inconclusive_ratifications: usize,
    pub resurveys: usize,
    pub rejected_windows: usize,
    pub flow_transient_windows: usize,
    pub source_bound_samples: usize,
    pub inelastic_samples: usize,
    pub pressure_vetoed_shrinks: usize,
    pub time_to_converge: Option<Duration>,
    pub final_consumer_threads: usize,
    pub final_producer_limit: usize,
    pub final_consumer_live: usize,
    pub final_producer_active: usize,
    pub peak_controlled_slots: usize,
    pub peak_auxiliary_threads: usize,
    pub producer_measurement: Option<ProducerMeasurementStats>,
    pub rejections: Vec<Rejection>,
    pub final_epoch: u64,
    pub final_phase: BrokerPhase,
    pub current_drift: Duration,
    pub terminal_error: Option<String>,
    pub final_model: Option<Model>,
}

What the broker did, for logging and for tests.

Fields

steady_state_policy: SteadyStatePolicy

Configured behavior after the first stable split.

monitoring_stopped_after_convergence: bool

True when the controller terminated itself at convergence instead of continuing to sample until the application called finish.

controller_samples: u64

Number of controller wakeups, including startup warm-up windows.

controller_elapsed: Duration

Wall time for which the controller remained alive.

controller_cpu_nanos: Option<u64>

Controller thread CPU time across its complete lifetime. This is measured with only one thread-clock read at start and one at exit.

controller_cpu_accounting_failures: usize

Controller thread-clock reads that failed.

steady_probe_interval: Duration

Effective cadence between responsive steady-state probes.

opening_bracket: OpeningBracketReport

Cost and result of the optional startup-only opening bracket.

consumer_trajectory: Vec<usize>

Consumer threads at each decision, in order.

producer_trajectory: Vec<usize>

Producer limit at each decision, in order.

actual_consumer_trajectory: Vec<usize>

Actual live/active occupancy observed at every recorded sample.

actual_producer_trajectory: Vec<usize>

trajectory_elapsed: Vec<Duration>

Controller elapsed time for each retained trajectory point.

cap_trajectory: Vec<CapObservation>

Times at which the inferred producer ceiling or its evidence changed.

throughput_trace: Vec<f64>

Real pipeline-throughput windows retained while a decision was open. Steady-state probes are excluded, so this remains bounded startup or resurvey evidence rather than recurring telemetry overhead.

throughput_trace_dropped: usize

trace_samples_dropped: usize

Oldest samples discarded after reaching trace_capacity.

moves: usize

Splits applied, including reverts.

reverts: usize

Moves undone because throughput got worse.

nonlinear_probes: usize

Guarded local candidates attempted after the opening rejected or could not distinguish the model’s differing answer.

nonlinear_probe_improvements: usize

Probe targets retained because their throughput confidence interval was strictly better than the previous split.

nonlinear_override: bool

Whether the final split is held by measured opening-bracket evidence rather than the allocation-independent cost model.

inconclusive_ratifications: usize

Ratifications whose uncertainty interval could not establish a regression or a clear non-regression.

resurveys: usize

Times a settled split was re-opened by a detected regime change.

rejected_windows: usize

Windows discarded as carrying too little work to estimate a rate from.

flow_transient_windows: usize

Otherwise-usable windows rejected because producer buffers filled or drained materially, so the two stages did not process equivalent work.

source_bound_samples: usize

Samples where the run was source-bound rather than decode-bound.

inelastic_samples: usize

Samples where the producer could not use more concurrency at all.

pressure_vetoed_shrinks: usize

Model-requested producer shrinks vetoed because runnable producer work was still queued behind the current limit.

time_to_converge: Option<Duration>

Time from the end of warm-up until the split stopped changing.

Measured to the last actual change of allocation, not to the moment the controller last stopped deliberating. Those differ: a settled split can be re-opened by drift, re-solved, and left exactly where it was. That is a decision being reconsidered, not an allocation moving, and reporting it as a fresh convergence is misleading in the direction that matters — it makes a stable run look unstable.

Measured on a 2.3 billion read Flex run at -t 64: the split reached 25 decode slots about two seconds in and never moved again for the remaining 111 seconds, while a single harmless resurvey made this field read 87 seconds.

final_consumer_threads: usize

Final split.

final_producer_limit: usize

final_consumer_live: usize

Actual controlled occupancy observed when the controller stopped. Under freeze-after-convergence this is earlier than application shutdown.

final_producer_active: usize

peak_controlled_slots: usize

Largest observed sum of live consumer workers and active producer slots.

peak_auxiliary_threads: usize

Peak live threads explicitly outside the controlled slot budget: the controller itself plus producer-reported coordinators/optional samplers.

producer_measurement: Option<ProducerMeasurementStats>

Producer measurement diagnostics. Native cumulative counters report zero samples and observation time; sampled adapters also report their cadence and sampler CPU.

rejections: Vec<Rejection>

Rejected targets and the workload epoch in which they failed.

final_epoch: u64

final_phase: BrokerPhase

current_drift: Duration

terminal_error: Option<String>

final_model: Option<Model>

Last solved model, for diagnosing a split that looks wrong.

Implementations

impl BrokerReport

pub fn converged(&self) -> bool

Whether the broker settled and stayed settled.

Trait Implementations

impl Clone for BrokerReport

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BrokerReport

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

Formats the value using the given formatter.

impl Default for BrokerReport

fn default() -> Self

Returns the “default value” for a type.

impl Serialize for BrokerReport

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.