From 1e283bee4cddf60b47148626ec0e9bfd6f7660b0 Mon Sep 17 00:00:00 2001 From: Jelte Fennema-Nio Date: Wed, 1 Nov 2023 11:46:01 +0100 Subject: [PATCH] Fix flaky multi_cluster_management (#7295) One of our most flaky and most anoying tests is multi_cluster_management. It usually fails like this: ```diff SELECT citus_disable_node('localhost', :worker_2_port); citus_disable_node -------------------- (1 row) SELECT public.wait_until_metadata_sync(60000); +WARNING: waiting for metadata sync timed out wait_until_metadata_sync -------------------------- (1 row) ``` This tries to address that by hardening wait_until_metadata_sync. I believe the reason for this warning is that there is a race condition in wait_until_metadata_sync. It's possible for the pre-check to fail, then have the maintenance daemon send a notification. And only then have the backend start to listen. I tried to fix it in two ways: 1. First run LISTEN, and only then read do the pre-check. 2. If we time out, check again just to make sure that we did not miss the notification somehow. And don't show a warning if all metadata is synced after the timeout. It's hard to know for sure that this fixes it because the test is not repeatable and I could not reproduce it locally. Let's just hope for the best. --------- Co-authored-by: Onur Tirtir --- src/backend/distributed/test/metadata_sync.c | 48 +++++++++++-------- .../expected/multi_cluster_management.out | 4 +- .../regress/sql/multi_cluster_management.sql | 4 +- 3 files changed, 33 insertions(+), 23 deletions(-) diff --git a/src/backend/distributed/test/metadata_sync.c b/src/backend/distributed/test/metadata_sync.c index 46d2303d6..8ad4b15f2 100644 --- a/src/backend/distributed/test/metadata_sync.c +++ b/src/backend/distributed/test/metadata_sync.c @@ -90,6 +90,28 @@ activate_node_snapshot(PG_FUNCTION_ARGS) } +/* + * IsMetadataSynced checks the workers to see if all workers with metadata are + * synced. + */ +static bool +IsMetadataSynced(void) +{ + List *workerList = ActivePrimaryNonCoordinatorNodeList(NoLock); + + WorkerNode *workerNode = NULL; + foreach_ptr(workerNode, workerList) + { + if (workerNode->hasMetadata && !workerNode->metadataSynced) + { + return false; + } + } + + return true; +} + + /* * wait_until_metadata_sync waits until the maintenance daemon does a metadata * sync, or times out. @@ -99,19 +121,10 @@ wait_until_metadata_sync(PG_FUNCTION_ARGS) { uint32 timeout = PG_GETARG_UINT32(0); - List *workerList = ActivePrimaryNonCoordinatorNodeList(NoLock); - bool waitNotifications = false; - - WorkerNode *workerNode = NULL; - foreach_ptr(workerNode, workerList) - { - /* if already has metadata, no need to do it again */ - if (workerNode->hasMetadata && !workerNode->metadataSynced) - { - waitNotifications = true; - break; - } - } + /* First we start listening. */ + MultiConnection *connection = GetNodeConnection(FORCE_NEW_CONNECTION, + LOCAL_HOST_NAME, PostPortNumber); + ExecuteCriticalRemoteCommand(connection, "LISTEN " METADATA_SYNC_CHANNEL); /* * If all the metadata nodes have already been synced, we should not wait. @@ -119,15 +132,12 @@ wait_until_metadata_sync(PG_FUNCTION_ARGS) * the notification and we'd wait unnecessarily here. Worse, the test outputs * might be inconsistent across executions due to the warning. */ - if (!waitNotifications) + if (IsMetadataSynced()) { + CloseConnection(connection); PG_RETURN_VOID(); } - MultiConnection *connection = GetNodeConnection(FORCE_NEW_CONNECTION, - LOCAL_HOST_NAME, PostPortNumber); - ExecuteCriticalRemoteCommand(connection, "LISTEN " METADATA_SYNC_CHANNEL); - int waitFlags = WL_SOCKET_READABLE | WL_TIMEOUT | WL_POSTMASTER_DEATH; int waitResult = WaitLatchOrSocket(NULL, waitFlags, PQsocket(connection->pgConn), timeout, 0); @@ -139,7 +149,7 @@ wait_until_metadata_sync(PG_FUNCTION_ARGS) { ClearResults(connection, true); } - else if (waitResult & WL_TIMEOUT) + else if (waitResult & WL_TIMEOUT && !IsMetadataSynced()) { elog(WARNING, "waiting for metadata sync timed out"); } diff --git a/src/test/regress/expected/multi_cluster_management.out b/src/test/regress/expected/multi_cluster_management.out index b92d8d136..3eb549ab5 100644 --- a/src/test/regress/expected/multi_cluster_management.out +++ b/src/test/regress/expected/multi_cluster_management.out @@ -90,7 +90,7 @@ SELECT citus_disable_node('localhost', :worker_2_port); (1 row) -SELECT public.wait_until_metadata_sync(60000); +SELECT public.wait_until_metadata_sync(20000); wait_until_metadata_sync --------------------------------------------------------------------- @@ -812,7 +812,7 @@ SELECT citus_disable_node('localhost', 9999); (1 row) -SELECT public.wait_until_metadata_sync(60000); +SELECT public.wait_until_metadata_sync(20000); wait_until_metadata_sync --------------------------------------------------------------------- diff --git a/src/test/regress/sql/multi_cluster_management.sql b/src/test/regress/sql/multi_cluster_management.sql index ab268939f..86fbd15b6 100644 --- a/src/test/regress/sql/multi_cluster_management.sql +++ b/src/test/regress/sql/multi_cluster_management.sql @@ -39,7 +39,7 @@ SELECT master_get_active_worker_nodes(); SELECT 1 FROM master_add_node('localhost', :worker_2_port); SELECT citus_disable_node('localhost', :worker_2_port); -SELECT public.wait_until_metadata_sync(60000); +SELECT public.wait_until_metadata_sync(20000); SELECT master_get_active_worker_nodes(); -- add some shard placements to the cluster @@ -328,7 +328,7 @@ SELECT 1 FROM master_add_inactive_node('localhost', 9996, groupid => :worker_2_g SELECT master_add_inactive_node('localhost', 9999, groupid => :worker_2_group, nodecluster => 'olap', noderole => 'secondary'); SELECT master_activate_node('localhost', 9999); SELECT citus_disable_node('localhost', 9999); -SELECT public.wait_until_metadata_sync(60000); +SELECT public.wait_until_metadata_sync(20000); SELECT master_remove_node('localhost', 9999); -- check that you can't manually add two primaries to a group