Fix "failed to re-find parent key" btree VACUUM failure by tweaking

_bt_pagedel to recover from the failure: just search the whole parent level
if searching to the right fails.  This does nothing for the underlying problem
that index keys became out-of-order in the grandparent level.  However, we
believe that there is no other consequence worse than slightly inefficient
searching, so this narrow patch seems like the safest solution for the back
branches.

diff --git a/src/backend/access/nbtree/README b/src/backend/access/nbtree/README
index 80b6d173ab4b5917075216224322486dcc3d23ce..7fe4ea02dbf8e2b460fc8dbdb0dea78f186e3b13 100644 (file)
--- a/src/backend/access/nbtree/README
+++ b/src/backend/access/nbtree/README
@@ -189,9 +189,24 @@ the half-dead page's level and its parent's level may be a little out of
 whack: key space that appears to belong to the half-dead page's parent on the
 parent level may really belong to its right sibling.  We can tolerate this,
 however, because insertions and deletions on upper tree levels are always
-done by reference to child page numbers, not keys.  The only cost is that
-searches may sometimes descend to the half-dead page and then have to move
-right, rather than going directly to the sibling page.
+done by reference to child page numbers, not keys.  Searches may sometimes
+descend to the half-dead page and then have to move right, rather than going
+directly to the sibling page, but this is no different from the behavior
+during a split.
+
+A special case that arises from using half-dead pages for rightmost children
+is that it's possible for the grandparent level's sequence of keys to become
+out-of-order.  This occurs when there are a large number of insertions into
+the key space that's been implicitly transferred to the right sibling of the
+half-dead page's parent.  If the right sibling itself splits, the split
+bounding key (which could be less than the high key of the parent page) is
+inserted into the grandparent level to the right of the parent page.  This
+is pretty ugly, but it causes no serious damage.  Searches, again, may descend
+a bit to the left of the optimal path but will be able to recover.  The only
+problem is that when it comes time to delete the half-dead page, _bt_pagedel's
+normal strategy for finding the target page's parent can fail: the search for
+the page's high key may well descend to the right of the parent.  In this case
+we recover by searching from the left end of the parent level.
 
 A deleted page cannot be reclaimed immediately, since there may be other
 processes waiting to reference it (ie, search processes that just left the

diff --git a/src/backend/access/nbtree/nbtinsert.c b/src/backend/access/nbtree/nbtinsert.c
index 184be5765f2c4fadc7b6850cd3bdf986006431b9..9389413c76cafebc95888f4e2630bc8a32737064 100644 (file)
--- a/src/backend/access/nbtree/nbtinsert.c
+++ b/src/backend/access/nbtree/nbtinsert.c
@@ -1253,8 +1253,8 @@ _bt_insert_parent(Relation rel,
 
                /* Check for error only after writing children */
                if (pbuf == InvalidBuffer)
-                       elog(ERROR, "failed to re-find parent key in \"%s\"",
-                                RelationGetRelationName(rel));
+                       elog(ERROR, "failed to re-find parent key in \"%s\" for split pages %u/%u",
+                                RelationGetRelationName(rel), bknum, rbknum);
 
                /* Recursively update the parent */
                newres = _bt_insertonpg(rel, pbuf, stack->bts_parent,

diff --git a/src/backend/access/nbtree/nbtpage.c b/src/backend/access/nbtree/nbtpage.c
index 79710e8985487633d9c8171634b4c924525625b4..b3deae803f10bc93e4d38c9ed91a3229843db849 100644 (file)
--- a/src/backend/access/nbtree/nbtpage.c
+++ b/src/backend/access/nbtree/nbtpage.c
@@ -897,16 +897,28 @@ _bt_pagedel(Relation rel, Buffer buf, bool vacuum_full)
        rbuf = _bt_getbuf(rel, rightsib, BT_WRITE);
 
        /*
-        * Next find and write-lock the current parent of the target page.
-        * This is essentially the same as the corresponding step of
-        * splitting.
+        * Next find and write-lock the current parent of the target page. This is
+        * essentially the same as the corresponding step of splitting.  However,
+        * it's possible for the search to fail (for reasons explained in README).
+        * If that happens, we recover by searching the whole parent level, which
+        * is a tad inefficient but doesn't happen often enough to be a problem.
         */
        ItemPointerSet(&(stack->bts_btitem.bti_itup.t_tid),
                                   target, P_HIKEY);
        pbuf = _bt_getstackbuf(rel, stack, BT_WRITE);
        if (pbuf == InvalidBuffer)
-               elog(ERROR, "failed to re-find parent key in \"%s\"",
-                        RelationGetRelationName(rel));
+       {
+               /* Find the leftmost page in the parent level */
+               pbuf = _bt_get_endpoint(rel, opaque->btpo.level + 1, false);
+               stack->bts_blkno = BufferGetBlockNumber(pbuf);
+               stack->bts_offset = InvalidOffsetNumber;
+               _bt_relbuf(rel, pbuf);
+               /* and repeat search from there */
+               pbuf = _bt_getstackbuf(rel, stack, BT_WRITE);
+               if (pbuf == InvalidBuffer)
+                       elog(ERROR, "failed to re-find parent key in \"%s\" for deletion target page %u",
+                                RelationGetRelationName(rel), target);
+       }
        parent = stack->bts_blkno;
        poffset = stack->bts_offset;