{
  "status": "research",
  "advantages": [],
  "limitations": [
    {
      "text": "The reported latency is not a disk-persistence measurement; RPC duplicate suppression alone does not establish durable application-level retry handling.",
      "citationIds": [
        "paper"
      ]
    }
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-16",
  "slug": "erpc",
  "term": "eRPC",
  "aliases": [
    "Efficient RPC"
  ],
  "definition": "eRPC is a datacenter RPC library that supplies asynchronous request handling, loss recovery and congestion control over unreliable packet transports.",
  "note": "The paper uses UDP on lossy Ethernet and unreliable datagrams on InfiniBand, rather than one-sided remote-memory operations. Dispatch threads must run their event loops; long handlers can use worker threads. Client retransmission uses go-back-N, while the server avoids executing the same request handler twice. Suspected node failure produces error callbacks. [paper]\n\nTable 6 reports a 5.5 microsecond client-side median and 6.3 microsecond 99th percentile for replicated PUTs: three replicas, one client, 16-byte keys, 64-byte values and logs plus data in DRAM. Leader-side median latency is separately 3.1 microseconds. These observation points are not interchangeable. Competing systems were compared through published results on different hardware, not one common testbed. [paper]\n\nAt the pinned artifact revision, the SMR example supports PUTs and one client, lacks log compaction and membership changes, and documents a maximum-RPC-size constraint during leader replacement. These are example-application limits, not limits of the RPC library. [smr-notes]\n\nThe optional PmemLog constructor sets the volatile entry count to zero, clears persistent metadata and appends a dummy entry. Its append path persists entry data before the entry count. This source inspection demonstrates initialization and append ordering, not restart recovery; the presence of PMEM code does not turn the paper's DRAM result into a persistent-storage benchmark. [pmem-source]",
  "related": [
    "raft",
    "rdma-storage"
  ],
  "citations": [
    {
      "id": "paper",
      "title": "Kalia, Kaminsky and Andersen. Datacenter RPCs can be General and Fast. NSDI 2019, pp. 1-16; sections 3, 5.3, 7.1, Table 6 and Appendix B.",
      "url": "https://www.usenix.org/system/files/nsdi19-kalia.pdf",
      "accessed": "2026-11-16"
    },
    {
      "id": "smr-notes",
      "title": "eRPC SMR example: implementation constraints and leader-change notes, revision de83dab3eab4a0fb19bfc4881c11d4a6b89ff17d",
      "url": "https://github.com/erpc-io/eRPC/blob/de83dab3eab4a0fb19bfc4881c11d4a6b89ff17d/apps/smr/README.md",
      "accessed": "2026-11-16"
    },
    {
      "id": "pmem-source",
      "title": "eRPC PmemLog: constructor and append, revision de83dab3eab4a0fb19bfc4881c11d4a6b89ff17d",
      "url": "https://github.com/erpc-io/eRPC/blob/de83dab3eab4a0fb19bfc4881c11d4a6b89ff17d/apps/smr/pmem_log.h",
      "accessed": "2026-11-16"
    }
  ],
  "provenance": [
    "Literature and pinned-source review. No network benchmark, crash-recovery experiment or independent reproduction was performed. The inspected revision is not asserted to be the paper's evaluated artifact."
  ]
}
