{
  "slug": "hydra-remote-memory",
  "term": "Hydra remote memory",
  "aliases": [
    "Hydra (FAST 2022)"
  ],
  "definition": "Hydra is a research resilience layer that erasure-codes individual remote-memory pages and controls fragment placement.",
  "note": "Completion: Each page has k data and r parity fragments on distinct remote machines. Failure-recovery writes wait for all k+r; EC-only completes after k without the same resilience guarantee. Reads can finish with the first k of k+delta requested fragments. [paper]\n\nPlacement: CodingSets confines choices to disjoint extended groups, then selects less-loaded members. Under the paper's failure model, it trades fewer loss events for larger events, not a reduction in expected data lost. That invariance is not a guarantee for arbitrary rack-correlated failures; see the CodingSets topology control. [paper]\n\nFailure scope: Remote unreachability is distinct from silent corruption. The paper defines separate corruption modes and excludes local-network failure without disk fallback. Remote redundancy is not persistent storage. [paper]\n\nArtifact mismatch: The README illustrates 8+2, while the pinned header activates NDATAS=4 and NDISKS=NDATAS+2 with IS_PAGE_SIZE=4096. Configuration examples must not be reported as compiled defaults. [artifact] [header]\n\nReproduction boundary: The README targets historical Ubuntu, kernel and RDMA stacks. The paper's persistent-memory comparison uses DRAM-based emulation, not measured Optane hardware. No Hydra deployment was run here. [artifact] [paper]\n\nHesela arithmetic: For a fresh 4+2 page version with only four acknowledged fragments, losing one of those four leaves only three. With all six acknowledged, any two fragment losses leave four. The downloadable model enumerates every loss subset and keeps pending writes unavailable. This is a finite illustrative model, not an observed failure probability. [model] [results]",
  "status": "research",
  "advantages": [],
  "limitations": [],
  "related": [
    "carbink",
    "infiniswap",
    "codingsets"
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-15",
  "citations": [
    {
      "id": "paper",
      "title": "Lee et al. Hydra: Resilient and Highly Available Remote Memory. FAST 2022, pp. 181-198; sections 3.3, 4.1, 5 and 7.5.",
      "url": "https://www.usenix.org/system/files/fast22-lee.pdf",
      "accessed": "2026-11-15"
    },
    {
      "id": "artifact",
      "title": "Hydra authors: README example and historical prerequisites; pinned revision",
      "url": "https://github.com/SymbioticLab/Hydra/blob/9918789f1faa5169307b1d4ae7a1c86303dc3563/README.md",
      "accessed": "2026-11-15"
    },
    {
      "id": "header",
      "title": "Hydra authors: active EC configuration and page size; pinned header",
      "url": "https://github.com/SymbioticLab/Hydra/blob/9918789f1faa5169307b1d4ae7a1c86303dc3563/resilience_manager/infiniswap.h",
      "accessed": "2026-11-15"
    },
    {
      "id": "model",
      "title": "Hesela: executable acknowledged-fragment enumeration with explicit assumptions",
      "url": "https://hesela.dev/research/hydra-completion.mjs",
      "accessed": "2026-11-15"
    },
    {
      "id": "results",
      "title": "Hesela: generated 4+2 completion matrix, counts and denominators",
      "url": "https://hesela.dev/research/hydra-completion.json",
      "accessed": "2026-11-15"
    }
  ],
  "provenance": [
    "Primary-source synthesis and an original finite counting model, not a native Hydra experiment, benchmark replication or measured reliability estimate.",
    "The remote-memory system is unrelated to the Hydra filesystem fuzzer; the slug disambiguates the two projects."
  ]
}
