What does Single-Host Hot Migration and the Ten-Second Contract conclude, in numbers, about “Blue-Green Is Not Zero Downtime: the Interruption Window Set by Health Checks, Failure Thresholds and Connection Drain”?

Answer: Chapter 1 shows that a common probe configuration needs 35.0 seconds at worst to notice a dead backend, that a 10 second drain still cuts 9.5699 long connections in expectation, and that not one of four hundred fixed-seed switches failed zero requests.

Scope of the numbers: Examples use fictional names, parameters are illustrative, and all calculations come from the book's programs, which connect to no server. Marked boxes separate published evidence, runnable Python calculations (22 programs), classical sources and the author's hypotheses.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. printed values that match the answer: 10, 35.0

detection delay D after a silent failure (seconds)
confighktbestmeanworst P(D<=10) lost req
slow30.035.065.080.095.00.00016000
common10.035.025.030.035.00.0006000
brisk5.032.012.014.517.00.0002900
mid4.023.07.09.011.00.7501800
tight2.031.05.06.07.01.0001200
tighter2.021.03.04.05.01.000800
aggressive1.020.51.52.02.51.000400
rate lam = 200 requests per second, budget 10.0 s
worst case is never below (k-1) h + t: zero needs h -> 0
Show the full program
# Chapter 1. p11: detection delay of a health check, D = U h + (k-1) h + t.
# Model: a backend dies at a uniformly random instant. A probe is sent
# every h seconds; a probe to a dead backend fails only after its
# timeout t; the router marks the backend down after k consecutive
# failed probes. With U uniform on [0, 1) (the phase of the next probe)
# the detection delay is D = U h + (k - 1) h + t, so
#   best (k-1) h + t,  mean (k - 1/2) h + t,  worst k h + t.
# Requests that reach the dead backend during D fail; at a rate of
# lam requests per second that is lam E[D] failed requests on average.
# The probe settings and the rate lam = 200 per second are illustrative.

CONFIGS = (
    ("slow", 30.0, 3, 5.0),
    ("common", 10.0, 3, 5.0),
    ("brisk", 5.0, 3, 2.0),
    ("mid", 4.0, 2, 3.0),
    ("tight", 2.0, 3, 1.0),
    ("tighter", 2.0, 2, 1.0),
    ("aggressive", 1.0, 2, 0.5),
)
LAM = 200.0
BUDGET = 10.0


def delays(h, k, t):
    best = (k - 1) * h + t
    return best, best + 0.5 * h, best + h


def p_within(h, k, t, budget):
    best = (k - 1) * h + t
    return min(1.0, max(0.0, (budget - best) / h))


print("detection delay D after a silent failure (seconds)")
print("%-11s %5s %2s %5s %7s %7s %7s %8s %8s" % (
    "config", "h", "k", "t", "best", "mean", "worst",
    "P(D<=10)", "lost req"))
for name, h, k, t in CONFIGS:
    b, m, w = delays(h, k, t)
    print("%-11s %5.1f %2d %5.1f %7.1f %7.1f %7.1f %8.3f %8.0f" % (
        name, h, k, t, b, m, w, p_within(h, k, t, BUDGET), LAM * m))
print("rate lam = %.0f requests per second, budget %.1f s" % (
    LAM, BUDGET))
print("worst case is never below (k-1) h + t: zero needs h -> 0")

Single source and citation

Fan, Y. (2026). Zero-Downtime Hot Migration and Rollback on Multi-Container Single Hosts: Lossless Schema Evolution behind a Shared Tunnel and a Ten-Second Stop-the-Bleeding Contract (Version 1.0.0), Chapter 1: Blue-Green Is Not Zero Downtime: the Interruption Window Set by Health Checks, Failure Thresholds and Connection Drain. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-269 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026269,
  author    = {Fan, Yafu},
  title     = {Zero-Downtime Hot Migration and Rollback on Multi-Container Single Hosts: Lossless Schema Evolution behind a Shared Tunnel and a Ten-Second Stop-the-Bleeding Contract},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book