prisma-risk / tsoracle / 26381315822

95%

Build Type

push

github

Committed by web-flow

Commit Message

perf(paxos-toolkit): cache next/compacted log indices instead of scanning per append (#391)

RocksdbStorage::next_log_idx ran a reverse CF iterator plus a compacted_idx
point-get on every append, and append_entry/append_entries re-read compacted_idx
a second time after the batch. That was scan + two point-reads per decided entry,
on every node, on the critical replication path. OmniPaxos's append API carries no
index, so the store had to rediscover the tail each time; openraft never scans
because its entries carry their own index.

Cache next_idx and compacted_idx on the struct, seed them once at open_in (the
only remaining scan), and maintain them in lockstep with every mutating write:
increment on append, (from+count).max(compacted) on append_on_prefix, reset to
compacted when trim empties the log, and next_idx.max(idx) when set_compacted_idx
raises the floor. The cache is updated only after the write commits, so a failed
write leaves it matching disk, and open_in re-derives both from persisted state on
reopen, preserving crash-recovery semantics. append_entry/append_entries/get_log_len
now touch zero RocksDB reads.

The maintenance rules reproduce the source-of-truth formula MemStorage embodies, so
the cross-impl conformance harness remains the contract. Adds two cross-impl
conformance scenarios (append_on_prefix shrink, unpaired full trim) and a
RocksdbStorage-only guard for the compacted-floor case.

Closes #336

Coverage Stats

54 of 57 new or added lines in 1 file covered. (94.74%)

5 existing lines in 3 files now uncovered.

12605 of 13291 relevant lines covered (94.84%)

366509.35 hits per line