LogBus capacity run

0.0s

Speed Rep

Throughput

events per second (EPS)

Offered Achieved

Backlog

events written by the feeder but not yet committed by the sink

Memory

physical memory held by the process — resident set size — sampled about once a second

Latency

log scale · one value per dwell, not a continuous trace

Measured p50–p99 Predicted from queue depth

Phase

climb

Verdict

—

Offered

—EPS

Achieved

—EPS

Backlog

—events

Latency

p50—

p99—

est—

Blocked

engine—

harness—

CPU

—%

Memory

—MB

Show the numbers behind the chart

t (s)	Phase	Offered	Achieved	Backlog	Memory (MB)	CPU (%)	p50 (ms)	p99 (ms)	Predicted (ms)	Duty	Feeder blk	Verdict

Source: load/raw/2026-08-07T01-54-56-587Z.jsonl. 10 repetitions, zero events lost across 35,414,488. Achieved rate is differenced from the sink's cumulative tx-events counter; backlog is feed_written − sink_tx; memory is the engine's own /api/snapshot report of its resident set size, roughly one sample per second — so the peak is a single sample in a fast-moving drain and the true maximum may be higher.

Verdict, engine duty and feeder blocked are dwell-level, not per-sample. Each is one figure the controller computed over a whole dwell and it is repeated on every sample inside it — the readout holds its value flat across the dwell rather than tracing it, exactly as the percentiles do. Engine duty is the fraction of the dwell the engine spent saturated, and it is what the trip rule tests. Feeder blocked is the fraction the feeder itself spent unable to write: when it is high the ceiling being measured is the harness, not the engine. CPU is a per-sample reading from the same snapshot as memory, and on a reconstructed point it is carried from the nearest real sample — a spikier series than memory, so read it there as an indication rather than a measurement.

Hollow markers (†) are reconstructed, not sampled. A dwell the controller cuts short the instant it trips can be shorter than the once-a-second stats cadence and leave no sample behind — and those are exactly the dwells that broke the engine. Their endpoints are rebuilt from the dwell's own segment record, which carries the offered and achieved rates, the backlog it entered and left with, and its percentiles. Only memory has no counterpart there and is carried from the nearest real sample.

The dashed latency line is not a measurement. It is backlog ÷ achieved rate — Little's Law — the latency the queue depth implies. Where the solid and dashed lines agree, the engine is keeping up and its queues explain everything. Where the solid line sits far below the dashed one, the histogram is under-reporting: it only counts events that finished, so a queue that is still building is invisible to it. Latency percentiles are a lagging indicator of saturation; backlog is a leading one.