Control-Interval Matching

Marketplace pricing, autoscaling, dispatch radius and incentive systems are all feedback loops: measure a condition, act, observe the result. They oscillate for the reasons any control system oscillates — delayed feedback plus excessive gain.

The rule: the control interval must be at least as long as the actuation delay. If a driver takes ten minutes to reposition, adjusting the price every thirty seconds is issuing corrections for an error that has not yet had a chance to change, and each correction compounds the last.

Why it matters

An oscillating marketplace is worse for everyone than a slightly mispriced stable one, and the damage is not only economic. Oscillation teaches both sides of the market to game it — riders learn to wait five minutes, drivers learn to wait for the peak — and once behaviour adapts, the instability becomes strategic rather than mechanical and is far harder to remove.

Implementation patterns

• Rate-limit the change, not just the level. Capping how fast a multiplier can move per interval is the single most effective damper and the one most often missing.
• Add hysteresis — different thresholds for increasing and decreasing — so the system does not chatter at a boundary.
• Smooth and deduplicate the input. A user who cancels and re-requests three times is one unit of demand and three signals; raw counts make noise look like a trend.
• Choose the spatial or logical granularity deliberately. Cells too small give noisy estimates and create boundary cliffs that users exploit; too large dilute the signal until the control cannot act.
• Separate loops that operate on different clocks. Incentives act over hours and should target a predicted condition; stacking an hourly loop on a per-minute one gives two coupled oscillators.
• Watch for loops that feed each other. Widening a dispatch radius during a shortage pulls supply from adjacent regions, so the shortage propagates rather than resolving — a coupling that neither loop's owner can see alone.

Industry example

Ride-hailing platforms such as Ola face this in its purest form: a regional driver shortage raises price, the price suppresses demand immediately and attracts supply slowly, and the mismatch in those two response times is the source of the oscillation. The same structure governs surge pricing in food delivery, dynamic dispatch in logistics, and — outside marketplaces entirely — aggressive autoscaling on a metric that scaling itself changes.

Failure scenarios

• Control interval shorter than actuation delay, producing sustained oscillation.
• Retries counted as demand, inflating the signal.
• Coupled loops in pricing, dispatch and incentives that individually look stable.
• Boundary cliffs between regions that users arbitrage.
• Autoscaling on a metric that scaling changes, such as scaling on queue depth while adding consumers changes the queue depth — the classic self-referential loop.

Trade-offs

Damping a loop makes it slower to respond to genuine change, which means real shortages persist longer before being corrected. That is a real cost and it is measurable in lost transactions.

The judgement is that a stable, slightly-behind controller outperforms a fast oscillating one, because the oscillation's costs are hidden — user distrust, gaming behaviour, and second-order effects on adjacent regions — while the slow controller's costs are visible and bounded. When speed genuinely matters, buy it by shortening the actuation delay rather than the control interval: better prediction, pre-positioned supply, or standing capacity.

Interview question

"Your surge multiplier is oscillating between 1.0 and 2.5 every few minutes in one district. Walk me through the loop, tell me which parameter you change first, and explain why increasing the update frequency would make it worse."