{
  "slug": "piggybacking-code",
  "term": "Piggybacking code",
  "aliases": [
    "Piggyback codes",
    "Piggybacking framework"
  ],
  "definition": "A piggybacking code couples code instances by adding functions of earlier messages to later instances while preserving the base code's recovery sets.",
  "note": "The dependency order matters: recover the first message, remove its contribution from the next instance, then decode that instance. With an MDS base code, any k of n nodes still recover all messages. Invertible transformations within individual nodes also preserve those recovery sets. The functions need separate design to reduce repair traffic; the recovery argument alone promises no saving. [framework]\n\nHesela's GF(3) illustration stores (a0,b0), (a1,b1), and (a0+a1,b0+b1+g(a0,a1)) on three nodes. Arithmetic is modulo 3. Each of four tested functions, including a nonlinear product, preserves all 81 messages for every two-node survivor pair: 972 round trips in total. This is a small original recovery example, not an efficient repair construction or a reproduction of a paper's performance results. [model]\n\nAn unsafe cyclic variant stores (a0+a1+b0,b0+b1+a0) on the third node instead. Losing the first node leaves only 27 distinct observations for 81 possible messages. For example, messages (0,0,0,0) and (1,0,2,0) become indistinguishable. Arbitrary cross-instance mixing is therefore not covered by the triangular recovery guarantee. This counterexample does not say every cyclic construction fails. [model]\n\nBandwidth claims also need a helper-set contract. Hulett and Wootters distinguish recovery from any eligible d helpers from recovery using some selected d helpers; their low-substriping impossibility results concern scalar MDS base codes and linear repair. Those restrictions cannot be dropped when comparing piggybacking with general regenerating codes. [limits]",
  "status": "research",
  "advantages": [],
  "limitations": [],
  "related": [
    "hitchhiker",
    "subpacketization",
    "maximum-distance-separable-code",
    "repair-read-amplification",
    "optimal-access-repair"
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-22",
  "citations": [
    {
      "id": "framework",
      "title": "Rashmi, Shah and Ramchandran. A Piggybacking Design Framework for Read-and Download-efficient Distributed Storage Codes, section II and Appendix.",
      "url": "https://www.cs.cmu.edu/~nihars/publications/Piggybacking.pdf",
      "accessed": "2026-11-22"
    },
    {
      "id": "limits",
      "title": "Hulett and Wootters. Limitations on the Achievable Repair Bandwidth of Piggybacking Codes with Low Substriping (2017), arXiv:1707.02337v1, sections 3, 5 and 6.",
      "url": "https://arxiv.org/abs/1707.02337v1",
      "accessed": "2026-11-22"
    },
    {
      "id": "model",
      "title": "Hesela: exhaustive GF(3) recovery checks and a cyclic-coupling collision.",
      "url": "https://hesela.dev/research/piggybacking/model.test.js",
      "accessed": "2026-11-22"
    },
    {
      "id": "reading",
      "title": "Hesela: local repair, Hitchhiker read plans and physical-read boundaries.",
      "url": "https://hesela.com/analyses/lrc-repair-costs-locality/#hitchhiker-read-plans",
      "accessed": "2026-11-22"
    }
  ],
  "provenance": [
    "Primary mathematical sources reviewed beyond their abstracts. This record describes a construction framework, not a default HDFS policy.",
    "The independently authored finite model is exhaustively enumerated, without sampled data or estimated uncertainty. It checks recovery, not latency, bandwidth savings, disk reads or field reliability.",
    "Run node --test model.test.js beside model.js; both files are available under /research/piggybacking/. The code introduces no third-party dependencies."
  ]
}
