[pmg-devel] [PATCH v2 pmg-api] fix #5189: cluster: avoid sync errors for statistics and quarantine

[Friedrich Weber <f.weber@proxmox.com>]

After restoring a backup from a cluster on a fresh node with
statistics, and then creating a cluster, the following can happen
(node 1 being master and node 2 being a node): `ClusterInfo` on node
1 has no record about the last-synchronized `CStatistic` row id of
node 2. Thus, pmgmirror on node 1 initializes the record with -1 and
tries to synchronize *all* `CStatistic` rows with cid 2 from node 2.
But (some of) these rows may already exist on cid 1, because they
were part of the backup, so pmgmirror on node 1 triggers a Postgres
unique constraint violation, statistics synchronization on node 1
fails, and node 1 remains in the "synchronizing" state.

Fix this as follows: When a new node is added to a cluster, the master
now initializes its `ClusterInfo` record of the last-synchronized
`CStatistic` row id for that node cid with the maximum row id that
exists in the local `CStatistic` for that node cid, or with -1 if the
local `CStatistic` has no row for that node cid. This is valid because
the newly-added node copies the master's `CStatistic` table during
cluster join.

Do the same for the `CMailStore` table, where a similar sync error
could happen e.g. if the table has rows for both node cids, node 2 is
shut down and manually deleted from the cluster.conf, the maxcid is
manually reset to 1, and a fresh node is joined to the cluster and
gets assigned cid 2.

Signed-off-by: Friedrich Weber <f.weber@proxmox.com>
---

Notes:
    v1 -> v2:
    - initialize master `ClusterInfo` record for `CStatistic` once when
      joining a new node instead of initializing on the first cluster
      sync, as discussed with Stoiko off-list
    - do the same for `CMailStore`, where a similar sync error can happen
    
    As a side effect of this change, `ClusterInfo` on node 2 will now also
    have `lastid_CStatistic` and `lastid_CMailStore` rows for its own cid
    2 after join, because the database is copied over *after* the master
    runs `update_master_clusterinfo`, and `update_client_clusterinfo` on
    node 2 only deletes rows with the master cid 1. However, if I
    understand the cluster sync code correctly (and that's a big if),
    these additional rows should not have any effect, as node 2 will
    only ever read and update `ClusterInfo` rows for other cids, never
    for its own cid.
    
    v1: https://lists.proxmox.com/pipermail/pmg-devel/2024-January/002657.html

 src/PMG/DBTools.pm | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/src/PMG/DBTools.pm b/src/PMG/DBTools.pm
index 6112566..8770d06 100644
--- a/src/PMG/DBTools.pm
+++ b/src/PMG/DBTools.pm
@@ -1132,6 +1132,14 @@ sub update_master_clusterinfo {
 	$dbh->do ("INSERT INTO ClusterInfo (cid, name, ivalue) select $clientcid, 'lastmt_$table', " .
 		  "EXTRACT(EPOCH FROM now())::INTEGER");
     }
+
+    my @lastid_tables = ('CStatistic', 'CMailStore');
+
+    for my $table (@lastid_tables) {
+        $dbh->do("INSERT INTO ClusterInfo (cid, name, ivalue) " .
+	    "SELECT $clientcid, 'lastid_$table', COALESCE (max (rid), -1) FROM $table " .
+	    "WHERE cid = $clientcid");
+    }
 }
 
 sub update_client_clusterinfo {
-- 
2.39.2