# Feedback age: identical observations, opposite routing outcomes

This is original Hesela hypothetical arithmetic, not Tars, C3, a device
measurement or a reproduction of either paper. Generated 2026-11-21.

## Reproduce

Run Node.js 24 (standard library only):

```sh
node model.mjs
node --test model.test.mjs
```

To regenerate: `node model.mjs --write`. The test suite independently sums
per-job service clocks and counts the levels of the minimum queue size.

## Contract

One common simulation clock, in milliseconds. At decision time 102, replica A
has a zero-work snapshot sampled at 0 and received at 2; B has a zero-work
snapshot sampled at 100 and received at 102. Both transport delays are 2 ms.
Client holding times are 100 and 0 ms; total snapshot ages are 102 and 2 ms.
These exact timestamps are synthetic, not a recommendation to subtract
unsynchronized host clocks.

Servers stay empty until a hidden batch from other clients arrives at time 102,
immediately before the focal request. Batch admission precedes service.
Each replica has one serial FIFO worker; every job takes exactly 5 ms.
The work-ahead count includes all admitted jobs ahead of the focal request.
No partially served job exists at this boundary. Completion is (work+1)*5 ms.
Focal-request network transport, failures and consistency costs are omitted.

The age-blind policy minimizes observed completion, ties to A. The illustrative
age rule excludes observations older than 20 ms and uses the same ranking;
if all are excluded it chooses the youngest, then the same score/id tie rule.
This is an arbitrary example threshold, not Tars's algorithm or a recommendation.
The non-deployable oracle sees the actual work. Each state is a separate decision,
not a continuing workload or rate-controller simulation.

## Exact result and negative control

In worlds (A=8,B=0) and (A=0,B=8), the visible observations are identical.
The age rule picks B in both: it changes 45 to 5 ms in the first world and
5 to 45 ms in the second. Thus those observations cannot determine the
best server for both worlds, regardless of the chosen deterministic rule.

The full grid has 81 pairs, 0..8 jobs ahead at each server. Equal enumeration
weight is a declared demonstration, not an empirical workload distribution.
The age rule wins 36 comparisons, loses 36 and ties 9. Both policies sum
to 2025 ms (25 ms mean); the oracle sums to 1425 ms (1425/81 ms mean).
Either policy has 600/81 ms mean regret relative to that oracle.
There is no sampling uncertainty or confidence interval: every declared state
is enumerated. Other workload weights can change the comparison.

## Interpretation and boundaries

Feedback age is missing context for routing, not proof that an old observation
is wrong or that a newer observation is better. A fresh empty snapshot can
be overtaken by an immediate burst. This construction imposes no arrival bound,
so it deliberately cannot infer queue bounds from age alone. No empirical
latency improvement, optimal TTL, or universal policy ranking follows.

The research motivation is Tars, arXiv:1702.08172v1, sections III-A and IV-A.
The model does not execute or redistribute its simulator. Source metadata
and exact locators are in sources.json. Public interpretation:
https://hesela.com/analyses/storage-tail-slowdowns-raid-fanout/#feedback-age

Code and generated synthetic results: original Hesela work, MIT license.
Paper rights remain with their respective owners.
