{
  "slug": "cachesack",
  "term": "CacheSack",
  "aliases": [],
  "definition": "CacheSack is a flash-cache admission optimizer that assigns workload categories policy mixtures using estimated disk-read cost, flash writes and cache occupancy.",
  "note": "Its four choices are NeverAdmit, AdmitOnSecondMiss, AdmitOnMiss and AdmitOnWrite; the last also admits read misses. Policy mixtures are optimized under a capacity bound using an approximate retention-time model. [paper]\n\nA lower-tier RAM buffer can serve a flash miss. Counting flash hits alone therefore misses the disk-read objective. [paper]\n\nPolicy fractions describe fractions of blocks in a category, not per-request admission probabilities. The reported five-minute training windows can use older access history retained by the ghost cache. The paper gives a sufficient condition: ghost-cache TTL exceeds the optimal modeled retention time, rather than every candidate horizon. Relative cost weights are not disclosed, limiting independent reconstruction of the production cost objective. [paper]",
  "related": [
    "cache-admission",
    "flashield",
    "request-hit-ratio",
    "ghost-cache-entry"
  ],
  "citations": [
    {
      "id": "paper",
      "title": "Yang et al. CacheSack: Admission Optimization for Google Datacenter Flash Caches. ATC 2022, pp. 1021-1036; sections 3.4, 5, 6.3-6.5, 7.1 and Appendix A.",
      "url": "https://www.usenix.org/system/files/atc22-yang-tzu-wei.pdf",
      "accessed": "2026-11-12"
    },
    {
      "id": "hesela",
      "title": "Hesela. Layered-cache read accounting.",
      "url": "https://hesela.com/analyses/sieve-cache-eviction-traces/#cachesack-accounting",
      "accessed": "2026-11-12"
    }
  ],
  "status": "research",
  "advantages": [],
  "limitations": [],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-13",
  "provenance": [
    "Source review; the linked counterexample is constructed accounting, not a CacheSack execution or reproduction of the production evaluation."
  ]
}