cplan / continuous batching
The first token
is the timetable.
End-to-end SLO attainment can look healthy while every user stares at an empty screen. This board puts the first token on the clock. Set your mean-TTFT budget, pick a load, and watch static batching miss the gate that continuous batching clears.
The decision
At a fixed first-token budget, which scheduler keeps the platform inside it as arrival rate rises?
simulated sweep
Arrival rate
Departures / mean TTFT
—
Continuous —
Static —
The gate, in milliseconds
dots = mean TTFT · line = your budget
The scheduling read
—
—
—
continuous loads on time
—
static loads on time
—
TTFT cut at selected load
—
static p99 at selected load
Source receipts/receipt.json
Model scheduler simulator / 4s end-to-end SLO
Budget and SLO floor are your inputs
Open receipt ↗