summaryrefslogtreecommitdiff
path: root/src/third_party/wiredtiger/src/btree/bt_split.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/third_party/wiredtiger/src/btree/bt_split.c')
-rw-r--r--src/third_party/wiredtiger/src/btree/bt_split.c11
1 files changed, 9 insertions, 2 deletions
diff --git a/src/third_party/wiredtiger/src/btree/bt_split.c b/src/third_party/wiredtiger/src/btree/bt_split.c
index 2c6f8449e3c..1fd13c47fe0 100644
--- a/src/third_party/wiredtiger/src/btree/bt_split.c
+++ b/src/third_party/wiredtiger/src/btree/bt_split.c
@@ -1271,8 +1271,15 @@ __split_parent_climb(WT_SESSION_IMPL *session, WT_PAGE *page)
* to a different part of the tree where it will be written; in other words, in one part of the
* tree we'll skip the newly created insert split chunk, but we'll write it upon finding it in a
* different part of the tree.
- */
- if (__wt_btree_syncing_by_other_session(session)) {
+ *
+ * Historically we allowed checkpoint itself to trigger an internal split here. That wasn't
+ * correct, since if that split climbs the tree above the immediate parent the checkpoint walk
+ * will potentially miss some internal pages. This is wrong as checkpoint needs to reconcile the
+ * entire internal tree structure. Non checkpoint cursor traversal doesn't care the internal
+ * tree structure as they just want to get the next leaf page correctly. Therefore, it is OK to
+ * split concurrently to cursor operations.
+ */
+ if (WT_BTREE_SYNCING(S2BT(session))) {
__split_internal_unlock(session, page);
return (0);
}