PuduLangResilience.Domain.Backoff

type: module path: "@root/src/PuduLangResilience/Domain/Backoff.pudu" fidelity: Active grammar: "[[grammar/pudu]]" depth_score: 0.85 depth_status: DEEP tags: [module, deep] aliases: [PuduLangResilience.Domain.Backoff]

Purpose

The pure delay before each retry: constant, linear, or exponential growth, a ceiling, and jitter.

Interface

Signatures

export type Shape = Constant | Linear | Exponential

export type Plan = { shape: Shape, delay: Int, maxDelay: Option[Int], useJitter: Bool }

export fn next(plan: &Plan, attempt: Int, state: Float, draw: Float) -> (Int, Float)

export fn steady(plan: &Plan, attempt: Int) -> Int

Linkage

• Requires: [[src/PuduLangResilience/Utils/Numeric]], Std.Math, Std.Math.Float, Std.Option.
• Consumed by: [[src/PuduLangResilience/Retry]].

Algorithm

1. A zero base answers zero. Constant: the base. Linear: (attempt + 1) * base. Exponential: the base doubled attempt times. All saturate.
2. Constant and linear jitter move the delay by up to a quarter either way: d + d/2 * draw - d/4.
3. Exponential jitter follows the decorrelated curve 2^t * tanh(sqrt(4t)) at t = attempt + draw, less the previous value, times the base divided by 1.4; a curve that stops being finite saturates.
4. The ceiling caps the result.

Negative Logic (Prohibited Paths)

• No delay is negative or overflows.

Edge Cases

• At draw 0.5 constant jitter keeps the delay exactly.

Depth

DEPTH 0.85 (DEEP). Tested by the suite mirroring this module under test/.

Grill Log

• Q: Why the decorrelated curve for exponential jitter? A: Its medians stay near whole multiples of the base while spreading clients that failed together. _Rejected:_ full jitter (median halves the intended wait).

Referenced by

[[src/PuduLangResilience/Domain/_MOC]] · [[src/PuduLangResilience/Retry]] · [[src/PuduLangResilience/Utils/Numeric]]