SERVER-56375: [RRFaM] BACKPORT-8905 Have delete path write to config.transactions

5 files changed, 83 insertions, 19 deletions

diff --git a/src/mongo/db/SConscript b/src/mongo/db/SConscript
index 0736608db0f..30058745a85 100644
--- a/src/mongo/db/SConscript
+++ b/src/mongo/db/SConscript
@@ -761,6 +761,7 @@ env.Library(
         'session_catalog',
     ],
     LIBDEPS_PRIVATE=[
+        '$BUILD_DIR/mongo/db/repl/repl_server_parameters',
         '$BUILD_DIR/mongo/idl/idl_parser',
         '$BUILD_DIR/mongo/idl/server_parameter',
         '$BUILD_DIR/mongo/util/concurrency/thread_pool',
diff --git a/src/mongo/db/op_observer_impl.cpp b/src/mongo/db/op_observer_impl.cpp
index e0df138b418..8d01ddd1583 100644
--- a/src/mongo/db/op_observer_impl.cpp
+++ b/src/mongo/db/op_observer_impl.cpp
@@ -171,22 +171,6 @@ struct OpTimeBundle {
     Date_t wallClockTime;
 };
 
-void writeToImagesCollection(OperationContext* opCtx,
-                             BSONObj image,
-                             repl::RetryImageEnum imageKind,
-                             Timestamp ts) {
-    repl::ImageEntry imageEntry;
-    invariant(opCtx->getLogicalSessionId());
-    imageEntry.set_id(*opCtx->getLogicalSessionId());
-    imageEntry.setTs(ts);
-    imageEntry.setImage(std::move(image));
-    imageEntry.setImageKind(imageKind);
-    repl::UnreplicatedWritesBlock unreplicated(opCtx);
-    AutoGetCollection imageCollectionRaii(
-        opCtx, NamespaceString::kConfigImagesNamespace, LockMode::MODE_IX);
-    Helpers::upsert(opCtx, NamespaceString::kConfigImagesNamespace.toString(), imageEntry.toBSON());
-}
-
 /**
  * Write oplog entry(ies) for the update operation.
  */
@@ -354,6 +338,25 @@ OpTimeBundle replLogApplyOps(OperationContext* opCtx,
     return times;
 }
 
+void writeToImageCollection(OperationContext* opCtx,
+                            const LogicalSessionId& sessionId,
+                            const Timestamp timestamp,
+                            repl::RetryImageEnum imageKind,
+                            const BSONObj& dataImage) {
+    repl::ImageEntry imageEntry;
+    invariant(opCtx->getLogicalSessionId());
+    imageEntry.set_id(sessionId);
+    imageEntry.setTxnNumber(opCtx->getTxnNumber().get());
+    imageEntry.setTs(timestamp);
+    imageEntry.setImageKind(imageKind);
+    imageEntry.setImage(dataImage);
+
+    repl::UnreplicatedWritesBlock unreplicated(opCtx);
+    AutoGetCollection imageCollectionRaii(
+        opCtx, NamespaceString::kConfigImagesNamespace, LockMode::MODE_IX);
+    Helpers::upsert(opCtx, NamespaceString::kConfigImagesNamespace.toString(), imageEntry.toBSON());
+}
+
 }  // namespace
 
 BSONObj OpObserverImpl::getDocumentKey(OperationContext* opCtx,
@@ -621,7 +624,11 @@ void OpObserverImpl::onUpdate(OperationContext* opCtx, const OplogUpdateEntryArg
                 imageDoc = args.updateArgs.updatedDoc;
                 imageKind = repl::RetryImageEnum::kPostImage;
             }
-            writeToImagesCollection(opCtx, imageDoc, imageKind, opTime.writeOpTime.getTimestamp());
+            writeToImageCollection(opCtx,
+                                   *opCtx->getLogicalSessionId(),
+                                   opTime.writeOpTime.getTimestamp(),
+                                   imageKind,
+                                   imageDoc);
         }
         SessionTxnRecord sessionTxnRecord;
         sessionTxnRecord.setLastWriteOpTime(opTime.writeOpTime);
@@ -687,6 +694,15 @@ void OpObserverImpl::onDelete(OperationContext* opCtx,
             args.deletedDoc ? boost::optional<BSONObj>(*(args.deletedDoc)) : boost::none;
         opTime = replLogDelete(
             opCtx, nss, uuid, stmtId, args.fromMigrate, deletedDoc, storeImagesInSideCollection);
+
+        if (storeImagesInSideCollection && deletedDoc && opCtx->getTxnNumber()) {
+            writeToImageCollection(opCtx,
+                                   *opCtx->getLogicalSessionId(),
+                                   opTime.writeOpTime.getTimestamp(),
+                                   repl::RetryImageEnum::kPreImage,
+                                   *deletedDoc);
+        }
+
         SessionTxnRecord sessionTxnRecord;
         sessionTxnRecord.setLastWriteOpTime(opTime.writeOpTime);
         sessionTxnRecord.setLastWriteDate(opTime.wallClockTime);
diff --git a/src/mongo/db/ops/delete.cpp b/src/mongo/db/ops/delete.cpp
index 3a95b360a47..78ee9f8d924 100644
--- a/src/mongo/db/ops/delete.cpp
+++ b/src/mongo/db/ops/delete.cpp
@@ -64,4 +64,24 @@ long long deleteObjects(OperationContext* opCtx,
     return DeleteStage::getNumDeleted(*exec);
 }
 
+DeleteResult deleteObject(OperationContext* opCtx,
+                          Collection* collection,
+                          const DeleteRequest& request) {
+    ParsedDelete parsedDelete(opCtx, &request);
+    uassertStatusOK(parsedDelete.parseRequest());
+
+    // This method doesn't support multi-deletes.
+    invariant(!request.isMulti());
+
+    auto exec = uassertStatusOK(
+        getExecutorDelete(opCtx, &CurOp::get(opCtx)->debug(), collection, &parsedDelete));
+
+    BSONObj docImage;
+    if (exec->getNext(&docImage, nullptr) == PlanExecutor::IS_EOF) {
+        return {};
+    }
+
+    return {1, request.shouldReturnDeleted() ? docImage.getOwned() : BSONObj()};
+}
+
 }  // namespace mongo
diff --git a/src/mongo/db/ops/delete.h b/src/mongo/db/ops/delete.h
index b26f583d460..c73c353ff86 100644
--- a/src/mongo/db/ops/delete.h
+++ b/src/mongo/db/ops/delete.h
@@ -30,6 +30,7 @@
 #pragma once
 
 #include "mongo/db/jsobj.h"
+#include "mongo/db/ops/delete_request.h"
 #include "mongo/db/query/plan_executor.h"
 
 
@@ -50,4 +51,14 @@ long long deleteObjects(OperationContext* opCtx,
                         bool justOne,
                         bool god = false,
                         bool fromMigrate = false);
+
+struct DeleteResult {
+    long long nDeleted;
+    boost::optional<BSONObj> requestedPreImage;
+};
+
+DeleteResult deleteObject(OperationContext* opCtx,
+                          Collection* collection,
+                          const DeleteRequest& request);
+
 }  // namespace mongo
diff --git a/src/mongo/db/repl/oplog.cpp b/src/mongo/db/repl/oplog.cpp
index 8e842732f86..cebb59e3990 100644
--- a/src/mongo/db/repl/oplog.cpp
+++ b/src/mongo/db/repl/oplog.cpp
@@ -68,6 +68,7 @@
 #include "mongo/db/namespace_string.h"
 #include "mongo/db/op_observer.h"
 #include "mongo/db/ops/delete.h"
+#include "mongo/db/ops/delete_request.h"
 #include "mongo/db/ops/update.h"
 #include "mongo/db/repl/apply_ops.h"
 #include "mongo/db/repl/bgsync.h"
@@ -1791,6 +1792,7 @@ Status applyOperation_inlock(OperationContext* opCtx,
         }
 
         const StringData ns = fieldNs.valueStringDataSafe();
+        bool upsertConfigImage = true;
         writeConflictRetry(opCtx, "applyOps_delete", ns, [&] {
             WriteUnitOfWork wuow(opCtx);
             if (timestamp != Timestamp::min()) {
@@ -1798,8 +1800,22 @@ Status applyOperation_inlock(OperationContext* opCtx,
             }
 
             if (opType[1] == 0) {
-                const auto justOne = true;
-                deleteObjects(opCtx, collection, requestNss, deleteCriteria, justOne);
+                const bool kNeedsRetryImage =
+                    op.hasField(OplogEntryBase::kNeedsRetryImageFieldName);
+                DeleteRequest request(requestNss);
+                request.setQuery(deleteCriteria);
+                if (kNeedsRetryImage) {
+                    request.setReturnDeleted(true);
+                }
+
+                DeleteResult result = deleteObject(opCtx, collection, request);
+                if (result.nDeleted == 1 && kNeedsRetryImage) {
+                    writeToImageCollection(opCtx,
+                                           op,
+                                           result.requestedPreImage.get(),
+                                           repl::RetryImageEnum::kPreImage,
+                                           &upsertConfigImage);
+                }
             } else
                 verify(opType[1] == 'b');  // "db" advertisement
             wuow.commit();