- Rotated filename timestamp is now ISO basic at second precision
(20260902T143000Z); the exact cutoff ts remains inside the files.
- Kanta(retention=N) accepts a plain int as number of days.
- tests/test_rotation.py covers both framers: history split, idempotent
reopen, no-op cases, aged-out reduction to a single snapshot, internal
snapshot validation, extension-agnostic rotated naming, and the
disabled default.
- Fix plan_rotation to seed the cutoff state from the replay base
snapshot (previously asserted when the base already covered the cutoff).
- docs: snapshots carry the schema version in effect at their position,
keeping replay/migration behavior identical to the unrotated file.
New kanta.rotation module plans rotation from in-memory bytes: walks
snapshots backwards to a replay base predating the cutoff, validates
replayed state against every snapshot in range, and rebuilds the file as
leading cutoff snapshot + re-framed retained changes (+ final snapshot
when enough changes survived). KantaImpl.open executes the plan under
the exclusive lock before replay: copy2 aside to {stem}@{ts}.kantadb,
in-place locked rewrite with fdatasync, then trim the rotated copy to
the dropped-history prefix. Enabled via Kanta(retention=timedelta).
Rotation rewrites the database file while holding the exclusive lock;
flock follows the open file description across ftruncate, and on Windows
in-place rewrite avoids share-mode rename restrictions. Also update
rotation doc timestamp format to ISO basic with microseconds.
The producer now emits only plain key assignment (full-value replacement
for scalars, lists and type changes, including dict-over-non-dict) and
$delete for removed keys. $-prefixed keys are still escaped as $$;
values are stored verbatim.
The consumer remains compatible with jsondiff-produced diffs ($replace,
positional $insert/$delete, per-index list diffs). A dict diff against
a list state is treated as jsondiff list-ops only when every key is a
command or integer position, otherwise as our wholesale replacement.
Unescaping is narrowed to the $$ prefix so verbatim single-$ values
pass through untouched.
compute_diff now emits a simple subset of jsondiff's marshaled syntax:
plain key assignment for adds and full-value changes (lists replaced
wholesale), $delete for removed object keys, and $replace only when a
dict replaces a non-dict. Removed keys are always $delete, even when the
object becomes empty (jsondiff emitted $replace: {}).
apply_diff (single implementation in serialization.base, used by both
replay and patch_state) additionally accepts jsondiff-produced patches:
positional $insert/$delete and per-index nested diffs on lists, plus
$ escaping of user keys and values per the marshal format.
jsondiff moves to the dev dependency group; it is used only by tests
that verify patch compatibility in both directions.
LeoVasanko