[Tarantool-patches] [PATCH] applier: fix tx boundary check for half-applied txns

[Tarantool-patches] [PATCH] applier: fix tx boundary check for half-applied txns

[Serge Petrenko]

In case there are 2 "new" instances, running tarantool 2.2+,
master and replica, and one "old" instance, running an earlier tarantool
version, in a full-mesh cluster, it may happen that the "new" replica
receives part of a tx from an "old" instance, and the remaining part
from a "new" instance.

Since "new" instances preserve tx boundaries, "new" replica would skip
the tx remains assuming it has already applied the full tx if it has
applied the first tx row. This leads to gaps in "new" replica's WAL and
to skipping the remaining part of the tx forever.

Fix this behaviour to apply the full tx even if it's beginning is
already applied in mixed clusters.

Closes #5125
---

https://github.com/tarantool/tarantool/issues/5125
https://github.com/tarantool/tarantool/tree/sp/gh-5125-applier-tx-boundaries

 src/box/applier.cc | 23 +++++++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)

diff --git a/src/box/applier.cc b/src/box/applier.cc
index df48b4796..6ca4cca94 100644
--- a/src/box/applier.cc
+++ b/src/box/applier.cc
@@ -737,6 +737,7 @@ applier_apply_tx(struct stailq *rows)
 	struct xrow_header *first_row = &stailq_first_entry(rows,
 					struct applier_tx_row, next)->row;
 	struct xrow_header *last_row;
+	last_row = &stailq_last_entry(rows, struct applier_tx_row, next)->row;
 	struct replica *replica = replica_by_id(first_row->replica_id);
 	/*
 	 * In a full mesh topology, the same set of changes
@@ -748,9 +749,28 @@ applier_apply_tx(struct stailq *rows)
 			       &replicaset.applier.order_latch);
 	latch_lock(latch);
 	if (vclock_get(&replicaset.applier.vclock,
-		       first_row->replica_id) >= first_row->lsn) {
+		       last_row->replica_id) >= last_row->lsn) {
 		latch_unlock(latch);
 		return 0;
+	} else if (vclock_get(&replicaset.applier.vclock,
+			      first_row->replica_id) >= first_row->lsn) {
+		/*
+		 * We've received part of the tx from an old
+		 * instance not knowing of tx boundaries.
+		 * Skip the already applied part.
+		 */
+		struct xrow_header *tmp;
+		while (true) {
+			tmp = &stailq_first_entry(rows,
+						  struct applier_tx_row,
+						  next)->row;
+			if (tmp->lsn <= vclock_get(&replicaset.applier.vclock,
+						   tmp->replica_id)) {
+				stailq_shift(rows);
+			} else {
+				break;
+			}
+		}
 	}
 
 	/**
@@ -835,7 +855,6 @@ applier_apply_tx(struct stailq *rows)
 	 * instances, which send every single tx row as a separate
 	 * transaction.
 	 */
-	last_row = &stailq_last_entry(rows, struct applier_tx_row, next)->row;
 	vclock_follow(&replicaset.applier.vclock, last_row->replica_id,
 		      last_row->lsn);
 	latch_unlock(latch);
-- 
2.24.3 (Apple Git-128)

Re: [Tarantool-patches] [PATCH] applier: fix tx boundary check for half-applied txns

[Vladislav Shpilevoy]

Thanks for the patch!

LGTM.

Re: [Tarantool-patches] [PATCH] applier: fix tx boundary check for half-applied txns

[Kirill Yukhin]

Hello,

On 05 июл 14:57, Serge Petrenko wrote:
> In case there are 2 "new" instances, running tarantool 2.2+,
> master and replica, and one "old" instance, running an earlier tarantool
> version, in a full-mesh cluster, it may happen that the "new" replica
> receives part of a tx from an "old" instance, and the remaining part
> from a "new" instance.
> 
> Since "new" instances preserve tx boundaries, "new" replica would skip
> the tx remains assuming it has already applied the full tx if it has
> applied the first tx row. This leads to gaps in "new" replica's WAL and
> to skipping the remaining part of the tx forever.
> 
> Fix this behaviour to apply the full tx even if it's beginning is
> already applied in mixed clusters.
> 
> Closes #5125
> ---
> 
> https://github.com/tarantool/tarantool/issues/5125
> https://github.com/tarantool/tarantool/tree/sp/gh-5125-applier-tx-boundaries

LGTM.
I've checked your patch into 2.3, 2.4 and master.

--
Regards, Kirill Yukhin