SERVER-39227 Don't throw StaleConfig from command response in DBClientBase when connected to mongos

jstests/noPassthrough/mongos_exhausts_stale_config_retries.js

@@ -0,0 +1,63 @@
+// Verifies mongos returns StaleConfig if it exhausts its allowed stale version retry attempts,
+// using the command read and write modes.
+//
+// @tags: [requires_sharding]
+(function() {
+    "use strict";
+
+    const dbName = "test";
+    const collName = "foo";
+    const ns = dbName + '.' + collName;
+
+    const st = new ShardingTest({shards: 2, config: 1});
+    const testDB = st.s.getDB(dbName);
+
+    // Only testing the command read and write modes.
+    assert(testDB.getMongo().readMode() === "commands");
+    assert(testDB.getMongo().writeMode() === "commands");
+
+    // Shard a collection with the only chunk on shard0.
+    assert.commandWorked(st.s.adminCommand({enableSharding: dbName}));
+    st.ensurePrimaryShard(dbName, st.shard0.shardName);
+    assert.commandWorked(st.s.adminCommand({shardCollection: ns, key: {_id: 1}}));
+
+    const sourcePrimary = st.rs0.getPrimary();
+    const recipientPrimary = st.rs1.getPrimary();
+
+    // Disable the best-effort recipient metadata refresh after migrations and move the chunk
+    // between shards so the recipient shard, shard1, is stale.
+    assert.commandWorked(sourcePrimary.adminCommand(
+        {configureFailPoint: "doNotRefreshRecipientAfterCommit", mode: "alwaysOn"}));
+    assert.commandWorked(
+        st.s.adminCommand({moveChunk: ns, find: {_id: 0}, to: st.shard1.shardName}));
+
+    // Disable metadata refreshes on the recipient shard so it will indefinitely return StaleConfig.
+    assert.commandWorked(recipientPrimary.adminCommand(
+        {configureFailPoint: "skipShardFilteringMetadataRefresh", mode: "alwaysOn"}));
+
+    // Test various read and write commands that are sent with shard versions and thus can return
+    // StaleConfig. Batch writes, i.e. insert/update/delete, return batch responses with ok:1 and
+    // NoProgressMade write errors when retries are exhausted, so they are excluded.
+    const kCommands = [
+        {aggregate: collName, pipeline: [], cursor: {}},
+        {count: collName},
+        {distinct: collName, query: {}, key: "_id"},
+        {find: collName},
+        {findAndModify: collName, query: {_id: 0}, update: {$set: {x: 1}}},
+    ];
+
+    kCommands.forEach((cmd) => {
+        // The recipient shard should return StaleConfig until mongos exhausts its retries and
+        // returns the final StaleConfig to the client.
+        assert.commandFailedWithCode(testDB.runCommand(cmd),
+                                     ErrorCodes.StaleConfig,
+                                     "expected to fail with StaleConfig, cmd: " + tojson(cmd));
+    });
+
+    assert.commandWorked(sourcePrimary.adminCommand(
+        {configureFailPoint: "doNotRefreshRecipientAfterCommit", mode: "off"}));
+    assert.commandWorked(recipientPrimary.adminCommand(
+        {configureFailPoint: "skipShardFilteringMetadataRefresh", mode: "off"}));
+
+    st.stop();
+})();

src/mongo/client/dbclient_base.cpp

@@ -156,9 +156,15 @@ rpc::UniqueReply DBClientBase::parseCommandReplyMessage(const std::string& host,
         uassertStatusOK(_metadataReader(opCtx, commandReply->getCommandReply(), host));
     }
 
-    auto status = getStatusFromCommandResult(commandReply->getCommandReply());
-    if (status == ErrorCodes::StaleConfig) {
-        uassertStatusOK(status.withContext("stale config in runCommand"));
+    // StaleConfig is thrown because clients acting as routers handle the exception at a higher
+    // level. Routing clients only expect StaleConfig from shards, so the exception should not be
+    // thrown when connected to a mongos, which allows StaleConfig to be returned to clients that
+    // connect to a mongos with DBClient, e.g. the shell.
+    if (!isMongos()) {
+        auto status = getStatusFromCommandResult(commandReply->getCommandReply());
+        if (status == ErrorCodes::StaleConfig) {
+            uassertStatusOK(status.withContext("stale config in runCommand"));
+        }
     }
 
     return rpc::UniqueReply(replyMsg, std::move(commandReply));