RIFL
RIFL is an RPC duplicate-suppression framework that couples durable results with request identity, data migration and safe metadata reclamation.
Completion: The operation and its saved reply must become durable atomically. Retries keep the same client identifier and sequence number, and return the original reply without repeating the mutation. A completion record travels with a designated object; recovery rebuilds its lookup before serving retries. [1]
Lifetime: Client acknowledgements retire a completed prefix. An early unfinished request can retain later results. Lease expiry permits reclamation but must prevent an old request from being admitted as new; the earlier outcome can remain uncertain. [1]
Code boundary: The pinned implementation uses an inclusive ackId: results at or below it are reclaimable, unlike the paper's exclusive firstIncomplete boundary. checkDuplicate distinguishes retained in-progress/completed entries from new requests; only the new-request branches validate the lease. Acknowledged retries raise StaleRpcException. [2]
Lease validation: A local expiry estimate triggers coordinator validation, not immediate rejection. The pinned validator rejects a lease when the coordinator returns a zero lease identifier. [3]
Relations
- related
- Retry deduplication, Retry rendezvous, CURP
Limitations
- RIFL requires underlying durable storage and retry support. Losing client state can leave an operation's outcome uncertain; a protected inner RPC does not automatically protect an outer workflow. [1]
Citations
- Lee et al.: Implementing Linearizability at Large Scale and Low Latency. SOSP 2015, pp. 71-86; DOI 10.1145/2815400.2815416; sections 2-5, 9-10.
- RAMCloud UnackedRpcResults.cc: checkDuplicate and Client::processAck, commit b1866af
- RAMCloud ClientLeaseValidator.h: validate, commit b1866af
Sources & caveats
- Full-paper and pinned-source review, not a performance reproduction or formal verification. The later repository commit is not identified as the paper's evaluated artifact.