{
  "slug": "fail-stutter",
  "term": "Fail-stutter",
  "aliases": [
    "fail-stutter fault model",
    "fail-stutter fault tolerance"
  ],
  "definition": "Fail-stutter is a fault-model proposal that extends fail-stop behavior with performance faults in components that still operate.",
  "note": "The 2001 proposal separates correctness failures from operation below a chosen performance specification. It covers hardware and software; performance can differ persistently across components or vary over time. It is not limited to repeated pauses. The authors explicitly leave full formalization and evaluation open. [arpaci-hotos01]\n\nFail-slow describes degraded operation; gray failure additionally requires an application to observe trouble that the system's failure observer misses. These are related concepts, not interchangeable aliases. [gunawi-fast18] [huang-hotos17]\n\nWorked example, not a device benchmark: four mirrored pairs offer logical write rates of 100, 100, 100 and 10 MB/s. Dividing a 3,100 MB payload equally takes 77.5 s (40 MB/s overall); assigning payload in proportion to those constant rates takes 10 s (310 MB/s). Both policies write the same logical bytes with mirrored copies. This idealized allocation calculation follows section 3.2; it assumes divisible work, no shared bottleneck or placement overhead, and freedom to choose placement. Existing RAID layouts need not permit it. [arpaci-hotos01] Executable arithmetic and tests accompany the worked example. [hesela-model]",
  "related": [
    "fail-slow",
    "gray-failure",
    "fail-stop"
  ],
  "citations": [
    {
      "id": "arpaci-hotos01",
      "title": "R. H. Arpaci-Dusseau and A. C. Arpaci-Dusseau. Fail-Stutter Fault Tolerance. HotOS VIII, 2001, sections 2-3 and 5. DOI: 10.1109/HOTOS.2001.990058.",
      "url": "https://research.cs.wisc.edu/wind/Publications/fail-stutter.pdf",
      "accessed": "2026-11-20"
    },
    {
      "id": "gunawi-fast18",
      "title": "Gunawi et al. Fail-Slow at Scale. FAST 2018, sections 1-3 and Table 3; 101 reports, 112 root-cause occurrences.",
      "url": "https://www.usenix.org/system/files/conference/fast18/fast18-gunawi.pdf",
      "accessed": "2026-11-20"
    },
    {
      "id": "huang-hotos17",
      "title": "Huang et al. Gray Failure: The Achilles' Heel of Cloud-Scale Systems. HotOS 2017, section 3.2. DOI: 10.1145/3102980.3103005.",
      "url": "https://www.microsoft.com/en-us/research/wp-content/uploads/2017/06/paper-1.pdf",
      "accessed": "2026-11-20"
    },
    {
      "id": "hesela-observers",
      "title": "Hesela: Fail-slow NVMe SSDs: what SMART and heartbeats miss",
      "url": "https://hesela.com/analyses/nvme-fail-slow-unseen-by-smart/",
      "accessed": "2026-11-20"
    },
    {
      "id": "hesela-model",
      "title": "Hesela: exact-rational fail-stutter placement example, assumptions and reproducible tests (not hardware measurements).",
      "url": "https://hesela.dev/research/fail-stutter/README.txt",
      "accessed": "2026-11-20"
    }
  ],
  "status": "unknown",
  "advantages": [
    {
      "text": "Separating performance from correctness permits a design to keep using slower components when its correctness contract allows.",
      "citationIds": [
        "arpaci-hotos01"
      ]
    }
  ],
  "limitations": [
    {
      "text": "The performance specification and the boundary between slow and failed are design choices, not a universal latency threshold.",
      "citationIds": [
        "arpaci-hotos01"
      ]
    }
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-20",
  "provenance": [
    "Source synthesis of the cited papers; the arithmetic is an idealized constructed example, not an implementation or performance measurement.",
    "The year identifies the HotOS VIII publication, not the first observation of performance faults."
  ]
}
