Online Commutative Processing
Online Commutative Processing delivers snapshot-visible records to an order-independent scan, including old values supplied before their storage is reused.
Contract: each record in a declared scan range is processed once, using the version visible at query start. Callback order can differ from key order. KVell+ applies this model to analytical queries alongside transactions. [1]
Retention: point ranges permit repeated reads but retain versions. Mixed conventional snapshot queries can also postpone reclamation. Pending commits still need old values for recovery. [1]
Reuse: delayed propagation can reuse a block read already needed for an update. [1] A deleted-version marker prevents a later scan from falling back to an older, incorrect value. [2]
Artifact: the pinned KVell+ prototype supports one scan range and no point ranges. Its asynchronous API requires the caller to drain an injector queue. A write callback is followed by a separate transaction commit. [3]
Relations
- related
- KVell
Limitations
- The author README warns that crash recovery may not work and concurrency bugs may remain; its API examples do not establish production readiness. [3]
Citations
- Lepers et al. Kvell+: Snapshot Isolation without Snapshots. OSDI 2020, pp. 425-441, sections 2 and 4.
- KVell+ at bd570b32: old-version lookup and deleted-version index entries.
- KVell+ at bd570b32: README, API and prototype limitations.
Further reading
Sources & caveats
- Full 17-page author paper and README, gc.c and in-memory-index.c at bd570b32 reviewed. No engine execution or performance replication.
- The USENIX proceedings date the paper to 2020; the author README bibliography says 2021. The conference record determines the date here.