{
  "slug": "finesse",
  "term": "Finesse",
  "aliases": [
    "Finesse resemblance detection"
  ],
  "definition": "Finesse selects candidate chunks for delta compression using features extracted from equal-sized subchunks and grouped by local rank.",
  "note": "Mechanism: Each subchunk contributes its largest rolling Rabin fingerprint. Features are split into contiguous groups, sorted within each group, then combined across groups at matching ranks into super-features. This removes the repeated fingerprint transformations of N-transform matching. [paper]\n\nEvaluation: Across six backup datasets, feature computation averaged 3.5x and 3.2x faster on the two tested processors; complete prototype throughput improved by 41%-85%. These are different measurements, not interchangeable speedups. The setup used 8 KB expected chunks, twelve features, three super-features, first-match base selection and a 400 MB base cache. [paper]\n\nBoundary: Very different chunk sizes can disrupt the corresponding subchunks and lose matches. A resemblance candidate still needs base retrieval and delta encoding; matching a super-feature does not establish byte equality. [paper]",
  "related": [
    "resemblance-detection",
    "delta-compression",
    "deduplication",
    "rolling-hash",
    "destor"
  ],
  "citations": [
    {
      "id": "paper",
      "title": "Zhang et al. Finesse: Fine-Grained Feature Locality based Fast Resemblance Detection for Post-Deduplication Delta Compression. FAST 2019, pp. 121-128.",
      "url": "https://www.usenix.org/system/files/fast19-zhang.pdf",
      "accessed": "2026-11-22"
    }
  ],
  "status": "research",
  "advantages": [],
  "limitations": [],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-22",
  "provenance": [
    "Primary-source review of sections 3-5, Algorithms 1-2 and Figure 3; reported prototype results were not independently reproduced. [paper]"
  ]
}
