SERVER-50077 Set unsharded shard version on collection creation only on primary node on the shard op observer and clear the filtering metadata if the shard register update or the refresh fails on shardCollection

2 files changed, 27 insertions, 9 deletions

diff --git a/src/mongo/db/s/shard_server_op_observer.cpp b/src/mongo/db/s/shard_server_op_observer.cpp
index 4e5009c8748..f8e80658e21 100644
--- a/src/mongo/db/s/shard_server_op_observer.cpp
+++ b/src/mongo/db/s/shard_server_op_observer.cpp
@@ -468,7 +468,7 @@ void ShardServerOpObserver::onCreateCollection(OperationContext* opCtx,
     // sharded or unsharded and set it on the CSR. If this method is called with the metadata as
     // UNKNOWN, this means an internal collection creation, which can only be UNSHARDED
     auto* csr = CollectionShardingRuntime::get(opCtx, collectionName);
-    if (!csr->getCurrentMetadataIfKnown())
+    if (opCtx->writesAreReplicated() && !csr->getCurrentMetadataIfKnown())
         csr->setFilteringMetadata(opCtx, CollectionMetadata());
 }
 
diff --git a/src/mongo/db/s/shardsvr_shard_collection.cpp b/src/mongo/db/s/shardsvr_shard_collection.cpp
index b83b444f52f..95fb4a22143 100644
--- a/src/mongo/db/s/shardsvr_shard_collection.cpp
+++ b/src/mongo/db/s/shardsvr_shard_collection.cpp
@@ -500,7 +500,15 @@ void refreshAllShards(OperationContext* opCtx,
                       const NamespaceString& nss,
                       const ShardId& dbPrimaryShardId,
                       const std::vector<ChunkType>& initialChunks) {
-    forceShardFilteringMetadataRefresh(opCtx, nss);
+    // If the refresh fails, then the shard will end with a shardVersion UNSHARDED.
+    try {
+        forceShardFilteringMetadataRefresh(opCtx, nss);
+    } catch (const DBException&) {
+        UninterruptibleLockGuard noInterrupt(opCtx->lockState());
+        AutoGetCollection autoColl(opCtx, nss, MODE_IX);
+        CollectionShardingRuntime::get(opCtx, nss)->clearFilteringMetadata(opCtx);
+        throw;
+    }
 
     auto shardRegistry = Grid::get(opCtx)->shardRegistry();
 
@@ -544,13 +552,23 @@ UUID shardCollection(OperationContext* opCtx,
             const InitialSplitPolicy::ShardCollectionConfig& initialChunks) {
             // Insert chunk documents to config.chunks on the config server.
             writeFirstChunksToConfig(opCtx, initialChunks);
-
-            updateShardingCatalogEntryForCollection(opCtx,
-                                                    nss,
-                                                    targetState,
-                                                    initialChunks,
-                                                    *request.getCollation(),
-                                                    request.getUnique());
+            // If an error happens when contacting the config server, we don't know if the update
+            // succeded or not, which might cause the local shard version to differ from the config
+            // server, so we clear the metadata to allow another operation to refresh it.
+            try {
+                updateShardingCatalogEntryForCollection(opCtx,
+                                                        nss,
+                                                        targetState,
+                                                        initialChunks,
+                                                        *request.getCollation(),
+                                                        request.getUnique());
+
+            } catch (const DBException&) {
+                UninterruptibleLockGuard noInterrupt(opCtx->lockState());
+                AutoGetCollection autoColl(opCtx, nss, MODE_IX);
+                CollectionShardingRuntime::get(opCtx, nss)->clearFilteringMetadata(opCtx);
+                throw;
+            }
 
             refreshAllShards(opCtx, nss, dbPrimaryShardId, initialChunks.chunks);
         };