Beyond 5s Resolver v3: 2.20× more conflict history under the same 6 GiB budget

I have completed the Resolver milestone of Beyond 5s, an experimental FoundationDB fork that replaces the Resolver’s fixed time-based conflict-history window with a memory-budgeted structure.

Final v3 uses a temporal B+ tree backed by a shared block pool for nodes and key bytes. It combines recursive partitioning of sorted conflict ranges, one mixed descent for reads and write destinations, and cooperative software prefetching for both read and write paths, while preserving serial mutation order.

Results:

  • Retention: under the same 6 GiB history budget, v3 retains 582.9–584.7 seconds of the measured trace, versus FoundationDB’s calibrated 265-second window: 2.20× more history.
  • Intel: the 5,400-second trace takes 0.77× FoundationDB’s time. Most short workloads are also faster; one range workload is 1.12× slower.
  • AMD: results are workload-dependent. Several paths are approximately level, while three are slower, by up to 1.32×. The AMD long trace validates decisions and retention but gives no timing ratio.
  • Validation: every comparison verifies identical conflict decisions and replay hashes. The final snapshot includes 232 tests, among them a regression test for a stale saved-rank defect found after the original v3.8 tag. The fix validates the destination leaf’s generation and mutation epoch before a saved rank is reused.

This completes the Resolver milestone. Transactions longer than five seconds end to end also need Storage Server history and Commit Proxy routing history, followed by minRV tracking, which the Resolver itself does not require: it retains conflict history by budget and rejects commits whose read version is below its boundary.

Article and figures:

Design, decisions, validation evidence, and the link to the code and final release:

I would particularly value feedback on:

  1. whether memory-budgeted conflict history fits FoundationDB’s architectural direction;
  2. which conflict-range workloads from real deployments would be more representative benchmarks than conflictset, skiplisttest and the custom point/range workload, especially range-heavy ones with high abort rates; and
  3. which deterministic-simulation scenarios or invariants maintainers would require before considering an upstream proposal.