SERVER-56373 Add storeFindAndModifyImagesInOplog parameter and introduce 'needsRetryImage' field to oplog entries.

8 files changed, 109 insertions, 44 deletions

diff --git a/src/mongo/db/commands/dbcheck.cpp b/src/mongo/db/commands/dbcheck.cpp
index 6f867aa710a..c6ae019e61c 100644
--- a/src/mongo/db/commands/dbcheck.cpp
+++ b/src/mongo/db/commands/dbcheck.cpp
@@ -481,7 +481,8 @@ private:
                                                   {},
                                                   kUninitializedStmtId,
                                                   {},
-                                                  OplogSlot());
+                                                  OplogSlot(),
+                                                  {});
                 uow.commit();
                 return result;
             });
diff --git a/src/mongo/db/op_observer_impl.cpp b/src/mongo/db/op_observer_impl.cpp
index b44a8194378..171427154d1 100644
--- a/src/mongo/db/op_observer_impl.cpp
+++ b/src/mongo/db/op_observer_impl.cpp
@@ -50,6 +50,7 @@
 #include "mongo/db/repl/replication_coordinator.h"
 #include "mongo/db/s/collection_sharding_state.h"
 #include "mongo/db/server_options.h"
+#include "mongo/db/server_parameters.h"
 #include "mongo/db/session_catalog.h"
 #include "mongo/db/views/durable_view_catalog.h"
 #include "mongo/s/client/shard_registry.h"
@@ -63,6 +64,7 @@ using repl::OplogEntry;
 namespace {
 
 MONGO_FAIL_POINT_DEFINE(failCollectionUpdates);
+MONGO_EXPORT_SERVER_PARAMETER(storeFindAndModifyImagesInSideCollection, bool, false);
 
 const auto documentKeyDecoration = OperationContext::declareDecoration<BSONObj>();
 
@@ -77,7 +79,8 @@ repl::OpTime logOperation(OperationContext* opCtx,
                           const OperationSessionInfo& sessionInfo,
                           StmtId stmtId,
                           const repl::OplogLink& oplogLink,
-                          const OplogSlot& oplogSlot) {
+                          const OplogSlot& oplogSlot,
+                          boost::optional<repl::RetryImageEnum> needsRetryImage) {
     auto& times = OpObserver::Times::get(opCtx).reservedOpTimes;
     auto opTime = repl::logOp(opCtx,
                               opstr,
@@ -90,7 +93,8 @@ repl::OpTime logOperation(OperationContext* opCtx,
                               sessionInfo,
                               stmtId,
                               oplogLink,
-                              oplogSlot);
+                              oplogSlot,
+                              needsRetryImage);
 
     times.push_back(opTime);
     return opTime;
@@ -169,11 +173,21 @@ OpTimeBundle replLogUpdate(OperationContext* opCtx,
                            Session* session,
                            const OplogUpdateEntryArgs& args) {
     BSONObj storeObj;
+    boost::optional<repl::RetryImageEnum> needsRetryImage;
+    const auto storeImagesInSideCollection = storeFindAndModifyImagesInSideCollection.load() &&
+        serverGlobalParams.featureCompatibility.getVersion() >=
+            ServerGlobalParams::FeatureCompatibility::Version::kUpgradingTo40;
     if (args.storeDocOption == OplogUpdateEntryArgs::StoreDocOption::PreImage) {
         invariant(args.preImageDoc);
         storeObj = *args.preImageDoc;
+        if (storeImagesInSideCollection && opCtx->getTxnNumber()) {
+            needsRetryImage = repl::RetryImageEnum::kPreImage;
+        }
     } else if (args.storeDocOption == OplogUpdateEntryArgs::StoreDocOption::PostImage) {
         storeObj = args.updatedDoc;
+        if (storeImagesInSideCollection && opCtx->getTxnNumber()) {
+            needsRetryImage = repl::RetryImageEnum::kPostImage;
+        }
     }
 
     OperationSessionInfo sessionInfo;
@@ -188,7 +202,7 @@ OpTimeBundle replLogUpdate(OperationContext* opCtx,
     OpTimeBundle opTimes;
     opTimes.wallClockTime = getWallClockTimeForOpLog(opCtx);
 
-    if (!storeObj.isEmpty() && opCtx->getTxnNumber()) {
+    if (!storeImagesInSideCollection && !storeObj.isEmpty() && opCtx->getTxnNumber()) {
         auto noteUpdateOpTime = logOperation(opCtx,
                                              "n",
                                              args.nss,
@@ -200,10 +214,10 @@ OpTimeBundle replLogUpdate(OperationContext* opCtx,
                                              sessionInfo,
                                              args.stmtId,
                                              {},
-                                             OplogSlot());
+                                             OplogSlot(),
+                                             {});
 
         opTimes.prePostImageOpTime = noteUpdateOpTime;
-
         if (args.storeDocOption == OplogUpdateEntryArgs::StoreDocOption::PreImage) {
             oplogLink.preImageOpTime = noteUpdateOpTime;
         } else if (args.storeDocOption == OplogUpdateEntryArgs::StoreDocOption::PostImage) {
@@ -222,7 +236,8 @@ OpTimeBundle replLogUpdate(OperationContext* opCtx,
                                        sessionInfo,
                                        args.stmtId,
                                        oplogLink,
-                                       OplogSlot());
+                                       OplogSlot(),
+                                       needsRetryImage);
 
     return opTimes;
 }
@@ -249,21 +264,29 @@ OpTimeBundle replLogDelete(OperationContext* opCtx,
     OpTimeBundle opTimes;
     opTimes.wallClockTime = getWallClockTimeForOpLog(opCtx);
 
+    boost::optional<repl::RetryImageEnum> needsRetryImage;
     if (deletedDoc && opCtx->getTxnNumber()) {
-        auto noteOplog = logOperation(opCtx,
-                                      "n",
-                                      nss,
-                                      uuid,
-                                      deletedDoc.get(),
-                                      nullptr,
-                                      false,
-                                      opTimes.wallClockTime,
-                                      sessionInfo,
-                                      stmtId,
-                                      {},
-                                      OplogSlot());
-        opTimes.prePostImageOpTime = noteOplog;
-        oplogLink.preImageOpTime = noteOplog;
+        if (storeFindAndModifyImagesInSideCollection.load() &&
+            serverGlobalParams.featureCompatibility.getVersion() >=
+                ServerGlobalParams::FeatureCompatibility::Version::kUpgradingTo40) {
+            needsRetryImage = repl::RetryImageEnum::kPreImage;
+        } else {
+            auto noteOplog = logOperation(opCtx,
+                                          "n",
+                                          nss,
+                                          uuid,
+                                          deletedDoc.get(),
+                                          nullptr,
+                                          false,
+                                          opTimes.wallClockTime,
+                                          sessionInfo,
+                                          stmtId,
+                                          {},
+                                          OplogSlot(),
+                                          {});
+            opTimes.prePostImageOpTime = noteOplog;
+            oplogLink.preImageOpTime = noteOplog;
+        }
     }
 
     auto& documentKey = documentKeyDecoration(opCtx);
@@ -278,7 +301,8 @@ OpTimeBundle replLogDelete(OperationContext* opCtx,
                                        sessionInfo,
                                        stmtId,
                                        oplogLink,
-                                       OplogSlot());
+                                       OplogSlot(),
+                                       needsRetryImage);
     return opTimes;
 }
 
@@ -304,7 +328,8 @@ OpTimeBundle replLogApplyOps(OperationContext* opCtx,
                                      sessionInfo,
                                      stmtId,
                                      oplogLink,
-                                     OplogSlot());
+                                     OplogSlot(),
+                                     {});
     return times;
 }
 
@@ -344,7 +369,8 @@ void OpObserverImpl::onCreateIndex(OperationContext* opCtx,
                      {},
                      kUninitializedStmtId,
                      {},
-                     OplogSlot());
+                     OplogSlot(),
+                     {});
     } else {
         logOperation(opCtx,
                      "i",
@@ -357,7 +383,8 @@ void OpObserverImpl::onCreateIndex(OperationContext* opCtx,
                      {},
                      kUninitializedStmtId,
                      {},
-                     OplogSlot());
+                     OplogSlot(),
+                     {});
     }
 
     AuthorizationManager::get(opCtx->getServiceContext())
@@ -574,7 +601,8 @@ void OpObserverImpl::onInternalOpMessage(OperationContext* opCtx,
                  {},
                  kUninitializedStmtId,
                  {},
-                 OplogSlot());
+                 OplogSlot(),
+                 {});
 }
 
 void OpObserverImpl::onCreateCollection(OperationContext* opCtx,
@@ -619,7 +647,8 @@ void OpObserverImpl::onCreateCollection(OperationContext* opCtx,
                      {},
                      kUninitializedStmtId,
                      {},
-                     createOpTime);
+                     createOpTime,
+                     {});
     }
 
     AuthorizationManager::get(opCtx->getServiceContext())
@@ -666,7 +695,8 @@ void OpObserverImpl::onCollMod(OperationContext* opCtx,
                      {},
                      kUninitializedStmtId,
                      {},
-                     OplogSlot());
+                     OplogSlot(),
+                     {});
     }
 
     AuthorizationManager::get(opCtx->getServiceContext())
@@ -704,7 +734,8 @@ void OpObserverImpl::onDropDatabase(OperationContext* opCtx, const std::string&
                  {},
                  kUninitializedStmtId,
                  {},
-                 OplogSlot());
+                 OplogSlot(),
+                 {});
 
     uassert(
         50714, "dropping the admin database is not allowed.", dbName != NamespaceString::kAdminDb);
@@ -738,7 +769,8 @@ repl::OpTime OpObserverImpl::onDropCollection(OperationContext* opCtx,
                      {},
                      kUninitializedStmtId,
                      {},
-                     OplogSlot());
+                     OplogSlot(),
+                     {});
     }
 
     uassert(50715,
@@ -779,7 +811,8 @@ void OpObserverImpl::onDropIndex(OperationContext* opCtx,
                  {},
                  kUninitializedStmtId,
                  {},
-                 OplogSlot());
+                 OplogSlot(),
+                 {});
 
     AuthorizationManager::get(opCtx->getServiceContext())
         ->logOp(opCtx, "c", cmdNss, cmdObj, &indexInfo);
@@ -815,7 +848,8 @@ repl::OpTime OpObserverImpl::preRenameCollection(OperationContext* const opCtx,
                  {},
                  kUninitializedStmtId,
                  {},
-                 OplogSlot());
+                 OplogSlot(),
+                 {});
 
     return {};
 }
@@ -893,7 +927,8 @@ void OpObserverImpl::onEmptyCapped(OperationContext* opCtx,
                      {},
                      kUninitializedStmtId,
                      {},
-                     OplogSlot());
+                     OplogSlot(),
+                     {});
     }
 
     AuthorizationManager::get(opCtx->getServiceContext())
diff --git a/src/mongo/db/repl/oplog.cpp b/src/mongo/db/repl/oplog.cpp
index 8b1bf69fed5..ce2cf2ed50b 100644
--- a/src/mongo/db/repl/oplog.cpp
+++ b/src/mongo/db/repl/oplog.cpp
@@ -360,7 +360,8 @@ OplogDocWriter _logOpWriter(OperationContext* opCtx,
                             Date_t wallTime,
                             const OperationSessionInfo& sessionInfo,
                             StmtId statementId,
-                            const OplogLink& oplogLink) {
+                            const OplogLink& oplogLink,
+                            boost::optional<repl::RetryImageEnum> needsRetryImage) {
     BSONObjBuilder b(256);
 
     b.append("ts", optime.getTimestamp());
@@ -379,6 +380,10 @@ OplogDocWriter _logOpWriter(OperationContext* opCtx,
     if (o2)
         b.append("o2", *o2);
 
+    if (needsRetryImage) {
+        b.append("needsRetryImage", repl::RetryImage_serializer(*needsRetryImage));
+    }
+
     invariant(wallTime != Date_t{});
     b.appendDate("wall", wallTime);
 
@@ -458,7 +463,8 @@ OpTime logOp(OperationContext* opCtx,
              const OperationSessionInfo& sessionInfo,
              StmtId statementId,
              const OplogLink& oplogLink,
-             const OplogSlot& oplogSlot) {
+             const OplogSlot& oplogSlot,
+             boost::optional<repl::RetryImageEnum> needsRetryImage) {
     auto replCoord = ReplicationCoordinator::get(opCtx);
     // For commands, the test below is on the command ns and therefore does not check for
     // specific namespaces such as system.profile. This is the caller's responsibility.
@@ -499,7 +505,8 @@ OpTime logOp(OperationContext* opCtx,
                                wallClockTime,
                                sessionInfo,
                                statementId,
-                               oplogLink);
+                               oplogLink,
+                               needsRetryImage);
     const DocWriter* basePtr = &writer;
     auto timestamp = slot.opTime.getTimestamp();
     _logOpsInner(opCtx, nss, &basePtr, &timestamp, 1, oplog, slot.opTime);
@@ -571,7 +578,8 @@ std::vector<OpTime> logInsertOps(OperationContext* opCtx,
                                           wallClockTime,
                                           sessionInfo,
                                           begin[i].stmtId,
-                                          oplogLink));
+                                          oplogLink,
+                                          {}));
         oplogLink.prevOpTime = insertStatementOplogSlot.opTime;
         timestamps[i] = oplogLink.prevOpTime.getTimestamp();
         opTimes.push_back(insertStatementOplogSlot.opTime);
diff --git a/src/mongo/db/repl/oplog.h b/src/mongo/db/repl/oplog.h
index bda617909c4..40cd7fea97e 100644
--- a/src/mongo/db/repl/oplog.h
+++ b/src/mongo/db/repl/oplog.h
@@ -38,6 +38,7 @@
 #include "mongo/bson/timestamp.h"
 #include "mongo/db/catalog/collection_options.h"
 #include "mongo/db/logical_session_id.h"
+#include "mongo/db/repl/oplog_entry_gen.h"
 #include "mongo/db/repl/optime.h"
 #include "mongo/db/repl/replication_coordinator.h"
 #include "mongo/stdx/functional.h"
@@ -144,7 +145,8 @@ OpTime logOp(OperationContext* opCtx,
              const OperationSessionInfo& sessionInfo,
              StmtId stmtId,
              const OplogLink& oplogLink,
-             const OplogSlot& oplogSlot);
+             const OplogSlot& oplogSlot,
+             boost::optional<repl::RetryImageEnum> needsRetryImage);
 
 // Flush out the cached pointer to the oplog.
 // Used by the closeDatabase command to ensure we don't cache closed things.
diff --git a/src/mongo/db/repl/oplog_entry.idl b/src/mongo/db/repl/oplog_entry.idl
index f7a7fa6f949..0a30146c53f 100644
--- a/src/mongo/db/repl/oplog_entry.idl
+++ b/src/mongo/db/repl/oplog_entry.idl
@@ -46,6 +46,13 @@ enums:
             kUpdate: "u"
             kDelete: "d"
             kNoop: "n"
+    RetryImage:
+        description: "Dictates whether a pre-image or post-image is to be stored on behalf of this
+                      retryable write."
+        type: string
+        values:
+            kPreImage: "preImage"
+            kPostImage: "postImage"
 
 structs:
     ReplOperation:
@@ -141,3 +148,9 @@ structs:
                 optional: true
                 description: "The optime of another oplog entry that contains the document
                               after an update was applied."
+
+            needsRetryImage:
+                type: RetryImage
+                optional: true
+                description: "Identifies whether a secondary should store a pre-image or post-image
+                              associated with this oplog entry."
diff --git a/src/mongo/db/repl/oplog_test.cpp b/src/mongo/db/repl/oplog_test.cpp
index d98220eebb9..dbc8b434885 100644
--- a/src/mongo/db/repl/oplog_test.cpp
+++ b/src/mongo/db/repl/oplog_test.cpp
@@ -113,7 +113,8 @@ TEST_F(OplogTest, LogOpReturnsOpTimeOnSuccessfulInsertIntoOplogCollection) {
                        {},
                        kUninitializedStmtId,
                        {},
-                       OplogSlot());
+                       OplogSlot(),
+                       {});
         ASSERT_FALSE(opTime.isNull());
         wunit.commit();
     }
@@ -236,7 +237,8 @@ OpTime _logOpNoopWithMsg(OperationContext* opCtx,
                         {},
                         kUninitializedStmtId,
                         {},
-                        OplogSlot());
+                        OplogSlot(),
+                        {});
     ASSERT_FALSE(opTime.isNull());
 
     ASSERT(opTimeNssMap->find(opTime) == opTimeNssMap->end())
diff --git a/src/mongo/db/s/session_catalog_migration_destination.cpp b/src/mongo/db/s/session_catalog_migration_destination.cpp
index bcea0d02b51..c35928016d8 100644
--- a/src/mongo/db/s/session_catalog_migration_destination.cpp
+++ b/src/mongo/db/s/session_catalog_migration_destination.cpp
@@ -291,7 +291,8 @@ ProcessOplogResult processSessionOplog(const BSONObj& oplogBSON,
                                            sessionInfo,
                                            stmtId,
                                            oplogLink,
-                                           OplogSlot());
+                                           OplogSlot(),
+                                           {});
 
             const auto& oplogOpTime = result.oplogTime;
             uassert(40633,
diff --git a/src/mongo/db/session_test.cpp b/src/mongo/db/session_test.cpp
index 60eb1134448..59e2a5ee057 100644
--- a/src/mongo/db/session_test.cpp
+++ b/src/mongo/db/session_test.cpp
@@ -131,7 +131,8 @@ protected:
                            osi,
                            stmtId,
                            link,
-                           OplogSlot());
+                           OplogSlot(),
+                           {});
     }
 
     void bumpTxnNumberFromDifferentOpCtx(Session* session, TxnNumber newTxnNum) {
@@ -529,7 +530,8 @@ TEST_F(SessionTest, ErrorOnlyWhenStmtIdBeingCheckedIsNotInCache) {
                                   osi,
                                   1,
                                   {},
-                                  OplogSlot());
+                                  OplogSlot(),
+                                  {});
         session.onWriteOpCompletedOnPrimary(opCtx(), txnNum, {1}, opTime, wallClockTime);
         wuow.commit();
 
@@ -556,7 +558,8 @@ TEST_F(SessionTest, ErrorOnlyWhenStmtIdBeingCheckedIsNotInCache) {
                                   osi,
                                   kIncompleteHistoryStmtId,
                                   link,
-                                  OplogSlot());
+                                  OplogSlot(),
+                                  {});
 
         session.onWriteOpCompletedOnPrimary(
             opCtx(), txnNum, {kIncompleteHistoryStmtId}, opTime, wallClockTime);