{
  "slug": "eventually-perfect-failure-detector",
  "term": "Eventually perfect failure detector",
  "aliases": [
    "diamond-P failure detector",
    "eventual perfect failure detector"
  ],
  "definition": "An eventually perfect failure detector eventually and permanently suspects every crashed process at every correct process and eventually stops false suspicions among correct processes.",
  "note": "Class diamond-P combines strong completeness with eventual strong accuracy. After some finite time, correct processes are never suspected by correct processes; that time need not be known to an observer. Correct means never crashing in the modeled crash-stop run. [ct96]\n\nUnlike a perfect detector, this class permits initial false suspicions. Its contract does not prescribe a bounded warm-up duration, a numerical detection deadline or a measured false-alarm rate. A finite trace with no false alarms cannot establish a guarantee about all future behavior. [ct96]\n\nChandra and Toueg construct an eventually perfect detector using heartbeat messages and increasing timeouts under eventual bounds on relative process speeds and message transmission. The algorithm and those assumptions matter; increasing a timeout or selecting a phi threshold alone is not a proof that an implementation belongs to the class. [ct96]\n\nPartial synchrony separates unknown bounds from bounds that hold after an unknown stabilization time. Neither formulation turns eventual accuracy into an operational completion deadline. [dls88]",
  "related": [
    "failure-detector",
    "perfect-failure-detector",
    "partial-synchrony",
    "phi-accrual-detector"
  ],
  "citations": [
    {
      "id": "ct96",
      "title": "Chandra and Toueg. Unreliable Failure Detectors for Reliable Distributed Systems. JACM 43(2), 225-267, 1996; sections 2.1-2.4 and 9.",
      "url": "https://www.cs.utexas.edu/~lorenzo/corsi/cs380d/papers/p225-chandra.pdf",
      "accessed": "2026-11-16"
    },
    {
      "id": "dls88",
      "title": "Dwork, Lynch and Stockmeyer. Consensus in the Presence of Partial Synchrony. JACM 35(2), 288-323, 1988; sections 1.2-1.3 and 2.3.",
      "url": "https://groups.csail.mit.edu/tds/papers/Lynch/jacm88.pdf",
      "accessed": "2026-11-16"
    }
  ],
  "provenance": [
    "CT1996 sections 2.1-2.4 and Figure 1 define diamond-P; section 9 and Theorem 9.1 give the partially synchronous implementation and assumptions.",
    "DLS1988 sections 1.2-1.3 and 2.3 distinguish timing models. The finite-trace qualification is an inference from the eventual, all-future-time property; no implementation was benchmarked."
  ],
  "status": "unknown",
  "advantages": [],
  "limitations": [
    {
      "text": "These crash-stop process properties do not certify storage health, establish Byzantine fault tolerance, cover recovery after restart or enforce fencing against a stale writer.",
      "citationIds": [
        "ct96"
      ]
    }
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-16"
}
