Synchronization performance
$id-3572255392439745 title: Defer incremental synchronization time bound to issue 1607 date: 2026/09/13 source: @ottojung kind: accepted-tradeoff
Journal 3 currently imposes no end-to-end asymptotic running-time requirement on synchronization when a receiver already has part of the relevant journal history. In particular, the current specification work does not need to prove O(N), o(N), affected-closure-proportional, change-sensitive, or another end-to-end running-time bound before the replay-log synchronization semantics are considered correct.
GitHub issue #1607, "Make incremental synchronization cost depend on changes, not graph size", owns the future change-sensitive non-validation synchronization performance contract: materialized-view construction/cutover, identifier/index maintenance, graph traversal, normalization work, and other lifecycle processing which may still be graph-sized even when only a small Journal suffix is new.
Historical validation is not deferred by this intent. Its complexity contract is owned by $id-6845129073418625: synchronization/reset historical validation is O(C) in newly admitted/affected state and must not rescan unrelated retained history.
Until #1607 is resolved and this intent is superseded or updated, correctness-oriented Journal 3 specifications may permit whole-replica non-validation work even when only a small journal suffix is new.
This deferral concerns non-validation running time only. It does not weaken the historical-validation bounds in $id-6845129073418625, synchronization correctness, convergence, replay correctness, or independently stated streaming requirements.
$id-4924739474925738 title: Streamable journal synchronization date: 2026/09/13 source: @ottojung kind: requirement
Journal synchronization and replay processing must be streamable over journal records without requiring the complete journal history, complete newly transferred suffix, or complete changed-node set to be materialized in RAM as one collection.
Implementations may retain bounded iterator/runtime buffers and whatever local derived state is independently required by IncrementalGraph processing, but the journal transport/replay layer itself must support incremental consumption of records.
This requirement does not imply the deferred end-to-end running-time bound from $id-3572255392439745. Whole-replica non-validation graph work may still be permitted for correctness; historical validation remains separately bounded by $id-6845129073418625.