SERVER-20267 use the explain command path to answer OP_QUERY with $explain on mongos

7 files changed, 99 insertions, 29 deletions

diff --git a/jstests/core/explain_find.js b/jstests/core/explain_find.js
index fe6ac234986..2e2699ea05b 100644
--- a/jstests/core/explain_find.js
+++ b/jstests/core/explain_find.js
@@ -32,3 +32,12 @@ explain = db.runCommand({
 printjson(explain);
 assert.commandWorked(explain);
 assert.eq(2, explain.executionStats.nReturned);
+
+// Compatibility test for the $explain OP_QUERY flag. This can only run if find command is disabled.
+if (!db.getMongo().useReadCommands()) {
+    var explain = t.find({$query: {a: 4}, $explain: true}).limit(-1).next();
+    assert("queryPlanner" in explain);
+    assert("executionStats" in explain);
+    assert.eq(1, explain.executionStats.nReturned);
+    assert("allPlansExecution" in explain.executionStats);
+}
diff --git a/src/mongo/db/dbmessage.cpp b/src/mongo/db/dbmessage.cpp
index b1e6a245c28..eb50ecb063a 100644
--- a/src/mongo/db/dbmessage.cpp
+++ b/src/mongo/db/dbmessage.cpp
@@ -172,7 +172,7 @@ T DbMessage::readAndAdvance() {
 void replyToQuery(int queryResultFlags,
                   AbstractMessagingPort* p,
                   Message& requestMsg,
-                  void* data,
+                  const void* data,
                   int size,
                   int nReturned,
                   int startingFrom,
diff --git a/src/mongo/db/dbmessage.h b/src/mongo/db/dbmessage.h
index 73c30f09bac..aa74fd59ffb 100644
--- a/src/mongo/db/dbmessage.h
+++ b/src/mongo/db/dbmessage.h
@@ -327,7 +327,7 @@ struct DbResponse {
 void replyToQuery(int queryResultFlags,
                   AbstractMessagingPort* p,
                   Message& requestMsg,
-                  void* data,
+                  const void* data,
                   int size,
                   int nReturned,
                   int startingFrom = 0,
diff --git a/src/mongo/s/commands/cluster_find_cmd.cpp b/src/mongo/s/commands/cluster_find_cmd.cpp
index 68cd2c00953..937515d17c7 100644
--- a/src/mongo/s/commands/cluster_find_cmd.cpp
+++ b/src/mongo/s/commands/cluster_find_cmd.cpp
@@ -35,10 +35,7 @@
 #include "mongo/db/commands.h"
 #include "mongo/db/query/cursor_response.h"
 #include "mongo/db/stats/counters.h"
-#include "mongo/s/cluster_explain.h"
 #include "mongo/s/query/cluster_find.h"
-#include "mongo/s/strategy.h"
-#include "mongo/util/timer.h"
 
 namespace mongo {
 namespace {
@@ -116,31 +113,13 @@ public:
 
         // Parse the command BSON to a LiteParsedQuery.
         bool isExplain = true;
-        auto lpqStatus = LiteParsedQuery::makeFromFindCommand(std::move(nss), cmdObj, isExplain);
-        if (!lpqStatus.isOK()) {
-            return lpqStatus.getStatus();
+        auto lpq = LiteParsedQuery::makeFromFindCommand(std::move(nss), cmdObj, isExplain);
+        if (!lpq.isOK()) {
+            return lpq.getStatus();
         }
 
-        auto& lpq = lpqStatus.getValue();
-
-        BSONObjBuilder explainCmdBob;
-        int options = 0;
-        ClusterExplain::wrapAsExplain(
-            cmdObj, verbosity, serverSelectionMetadata, &explainCmdBob, &options);
-
-        // We will time how long it takes to run the commands on the shards.
-        Timer timer;
-
-        vector<Strategy::CommandResult> shardResults;
-        Strategy::commandOp(
-            txn, dbname, explainCmdBob.obj(), options, fullns, lpq->getFilter(), &shardResults);
-
-        long long millisElapsed = timer.millis();
-
-        const char* mongosStageName = ClusterExplain::getStageNameForReadOp(shardResults, cmdObj);
-
-        return ClusterExplain::buildExplainResult(
-            txn, shardResults, mongosStageName, millisElapsed, out);
+        return ClusterFind::runExplain(
+            txn, cmdObj, *lpq.getValue(), verbosity, serverSelectionMetadata, out);
     }
 
     bool run(OperationContext* txn,
diff --git a/src/mongo/s/query/cluster_find.cpp b/src/mongo/s/query/cluster_find.cpp
index 0d0fb418333..4f40dd3c795 100644
--- a/src/mongo/s/query/cluster_find.cpp
+++ b/src/mongo/s/query/cluster_find.cpp
@@ -42,9 +42,11 @@
 #include "mongo/db/query/canonical_query.h"
 #include "mongo/db/query/find_common.h"
 #include "mongo/db/query/getmore_request.h"
+#include "mongo/rpc/metadata/server_selection_metadata.h"
 #include "mongo/s/catalog/catalog_cache.h"
 #include "mongo/s/chunk_manager.h"
 #include "mongo/s/client/shard_registry.h"
+#include "mongo/s/cluster_explain.h"
 #include "mongo/s/config.h"
 #include "mongo/s/grid.h"
 #include "mongo/s/query/cluster_client_cursor_impl.h"
@@ -329,6 +331,37 @@ StatusWith<CursorResponse> ClusterFind::runGetMore(OperationContext* txn,
     return CursorResponse(request.nss, idToReturn, std::move(batch), startingFrom);
 }
 
+Status ClusterFind::runExplain(OperationContext* txn,
+                               const BSONObj& findCommand,
+                               const LiteParsedQuery& lpq,
+                               ExplainCommon::Verbosity verbosity,
+                               const rpc::ServerSelectionMetadata& serverSelectionMetadata,
+                               BSONObjBuilder* out) {
+    BSONObjBuilder explainCmdBob;
+    int options = 0;
+    ClusterExplain::wrapAsExplain(
+        findCommand, verbosity, serverSelectionMetadata, &explainCmdBob, &options);
+
+    // We will time how long it takes to run the commands on the shards.
+    Timer timer;
+
+    std::vector<Strategy::CommandResult> shardResults;
+    Strategy::commandOp(txn,
+                        lpq.nss().db().toString(),
+                        explainCmdBob.obj(),
+                        options,
+                        lpq.nss().toString(),
+                        lpq.getFilter(),
+                        &shardResults);
+
+    long long millisElapsed = timer.millis();
+
+    const char* mongosStageName = ClusterExplain::getStageNameForReadOp(shardResults, findCommand);
+
+    return ClusterExplain::buildExplainResult(
+        txn, shardResults, mongosStageName, millisElapsed, out);
+}
+
 StatusWith<ReadPreferenceSetting> ClusterFind::extractUnwrappedReadPref(const BSONObj& cmdObj,
                                                                         const bool isSlaveOk) {
     BSONElement queryOptionsElt;
diff --git a/src/mongo/s/query/cluster_find.h b/src/mongo/s/query/cluster_find.h
index ad878a11b37..7a7d3deac69 100644
--- a/src/mongo/s/query/cluster_find.h
+++ b/src/mongo/s/query/cluster_find.h
@@ -33,6 +33,7 @@
 #include "mongo/bson/bsonobj.h"
 #include "mongo/db/cursor_id.h"
 #include "mongo/db/query/cursor_response.h"
+#include "mongo/db/query/explain_common.h"
 
 namespace mongo {
 
@@ -43,6 +44,10 @@ class OperationContext;
 struct GetMoreRequest;
 struct ReadPreferenceSetting;
 
+namespace rpc {
+class ServerSelectionMetadata;
+}  // namespace rpc
+
 /**
  * Methods for running find and getMore operations across a sharded cluster.
  */
@@ -71,6 +76,21 @@ public:
                                                  const GetMoreRequest& request);
 
     /**
+     * Helper to run an explain of a find operation on the shards. Fills 'out' with the result of
+     * the of the explain command on success. On failure, returns a non-OK status and does not
+     * modify 'out'.
+     *
+     * Used both if mongos receives an explain command and if it receives an OP_QUERY find with the
+     * $explain modifier.
+     */
+    static Status runExplain(OperationContext* txn,
+                             const BSONObj& findCommand,
+                             const LiteParsedQuery& lpq,
+                             ExplainCommon::Verbosity verbosity,
+                             const rpc::ServerSelectionMetadata& serverSelectionMetadata,
+                             BSONObjBuilder* out);
+
+    /**
      * Extracts the read preference from 'cmdObj', or determines the read pref based on 'isSlaveOk'
      * if 'cmdObj' does not contain a read preference.
      *
diff --git a/src/mongo/s/strategy.cpp b/src/mongo/s/strategy.cpp
index 89bc8a0e2d4..357cd8843c0 100644
--- a/src/mongo/s/strategy.cpp
+++ b/src/mongo/s/strategy.cpp
@@ -49,6 +49,7 @@
 #include "mongo/db/query/lite_parsed_query.h"
 #include "mongo/db/query/getmore_request.h"
 #include "mongo/db/stats/counters.h"
+#include "mongo/rpc/metadata/server_selection_metadata.h"
 #include "mongo/s/bson_serializable.h"
 #include "mongo/s/catalog/catalog_cache.h"
 #include "mongo/s/client/shard_registry.h"
@@ -194,6 +195,34 @@ void Strategy::queryOp(OperationContext* txn, Request& request) {
         auto canonicalQuery = CanonicalQuery::canonicalize(q, WhereCallbackNoop());
         uassertStatusOK(canonicalQuery.getStatus());
 
+        // If the $explain flag was set, we must run the operation on the shards as an explain
+        // command rather than a find command.
+        if (canonicalQuery.getValue()->getParsed().isExplain()) {
+            const LiteParsedQuery& lpq = canonicalQuery.getValue()->getParsed();
+            BSONObj findCommand = lpq.asFindCommand();
+
+            // We default to allPlansExecution verbosity.
+            auto verbosity = ExplainCommon::EXEC_ALL_PLANS;
+
+            const bool secondaryOk = (readPreference.pref != ReadPreference::PrimaryOnly);
+            rpc::ServerSelectionMetadata metadata(secondaryOk, readPreference);
+
+            BSONObjBuilder explainBuilder;
+            uassertStatusOK(ClusterFind::runExplain(
+                txn, findCommand, lpq, verbosity, metadata, &explainBuilder));
+
+            BSONObj explainObj = explainBuilder.done();
+            replyToQuery(0,  // query result flags
+                         request.p(),
+                         request.m(),
+                         static_cast<const void*>(explainObj.objdata()),
+                         explainObj.objsize(),
+                         1,  // numResults
+                         0,  // startingFrom
+                         CursorId(0));
+            return;
+        }
+
         // Do the work to generate the first batch of results. This blocks waiting to get responses
         // from the shard(s).
         std::vector<BSONObj> batch;
@@ -205,11 +234,11 @@ void Strategy::queryOp(OperationContext* txn, Request& request) {
             ClusterFind::runQuery(txn, *canonicalQuery.getValue(), readPreference, &batch);
         uassertStatusOK(cursorId.getStatus());
 
-        // Build the response document.
         // TODO: this constant should be shared between mongos and mongod, and should
         // not be inside ShardedClientCursor.
         BufBuilder buffer(ShardedClientCursor::INIT_REPLY_BUFFER_SIZE);
 
+        // Fill out the response buffer.
         int numResults = 0;
         for (const auto& obj : batch) {
             buffer.appendBuf((void*)obj.objdata(), obj.objsize());