loro-crdt 1.13.6 → 1.13.8

package/CHANGELOG.md

@@ -1,5 +1,43 @@
 # Changelog
 
+## 1.13.8
+
+### Patch Changes
+
+- ac1feb6: Reduce peak memory across large snapshot import, `toJSON`, update import, and
+  snapshot export by preserving lazy state, bounding the decompressed SSTable
+  block cache by bytes, and preallocating the snapshot output exactly. Snapshot
+  import now validates every embedded SSTable block checksum before accepting
+  external bytes.
+
+  Full snapshot export no longer walks the alive-container graph: store entries
+  for referenced containers are created when the creating op or imported diff is
+  applied, so a cold export after a large import is a flush plus KV export
+  instead of re-reading every container from compressed blocks. As part of this,
+  full export round-trips imported state bytes faithfully (inconsistent
+  alive-container parent metadata is now rejected by shallow export, which still
+  walks, rather than by full export), and an ensured-but-empty mergeable child is
+  no longer materialized into the KV store by a full export — importers resolve
+  it from its parent map marker. First `toJSON` after import also avoids decoding
+  each lazy container twice.
+
+## 1.13.7
+
+### Patch Changes
+
+- 3986805: Speed up snapshot import. When decoding a Loro snapshot, the redundant per-block SSTable validation (eager block-metadata decode and per-block checksums) is now skipped, because the whole snapshot body is already protected by the document-level checksum verified during decoding. This removes a second hash pass over the data (roughly halving B4 snapshot import time) while preserving integrity guarantees.
+
+  This fast path is internal to Loro's snapshot decoding. The public `MemKvStore::import_all` still verifies every block's checksum; a separate `import_all_unchecked` opts into the unchecked path and is only used where an outer checksum already guarantees integrity.
+
+- 9b6330d: Speed up local text editing (~35% faster on the B4 editing trace). Three hot-path
+  changes: the lock-order debug instrumentation is now compiled out of release
+  builds (it ran on every per-op lock acquisition); the visible-op count is bumped
+  incrementally for local ops instead of recomputing it from the version vectors
+  (which also allocated) on every op; and a couple of per-op allocations on the
+  text insert/delete path were removed (lazy error-context formatting and inline
+  storage for entity ranges).
+- c37e76a: Fix a checkout hang after snapshot import. Change-store blocks decoded from a snapshot recorded a wrong end lamport (`lamport_range.1` was set to the start lamport of the block's last change instead of its end). When a change was split across multiple blocks and a lamport-based lookup engaged the binary search path in `ChangeStore::get_change_by_lamport_lte` — e.g. the movable-list diff calculator resolving historical positions during `checkout` — the degenerate range made the search loop forever. The binary search now also caps its steps and falls back to a scan if it ever fails to converge; the fallback considers both unflushed in-memory blocks and the underlying kv store, so lamport lookups (e.g. `getChangeAtLamport`) stay correct for local changes that have not been flushed yet.
+
 ## 1.13.6
 
 ### Patch Changes