SERVER-60524 Make retryable internal transactions retryable across data placement changes

5 files changed, 23 insertions, 0 deletions

diff --git a/src/mongo/db/commands/find_and_modify.cpp b/src/mongo/db/commands/find_and_modify.cpp
index 6518620c51c..dab13fcc40c 100644
--- a/src/mongo/db/commands/find_and_modify.cpp
+++ b/src/mongo/db/commands/find_and_modify.cpp
@@ -645,6 +645,7 @@ write_ops::FindAndModifyCommandReply CmdFindAndModify::Invocation::typedRun(
             RetryableWritesStats::get(opCtx)->incrementRetriedCommandsCount();
             RetryableWritesStats::get(opCtx)->incrementRetriedStatementsCount();
             auto findAndModifyReply = parseOplogEntryForFindAndModify(opCtx, req, *entry);
+            findAndModifyReply.setRetriedStmtId(stmtId);
 
             // Make sure to wait for writeConcern on the opTime that will include this
             // write. Needs to set to the system last opTime to get the latest term in an
diff --git a/src/mongo/db/commands/write_commands.cpp b/src/mongo/db/commands/write_commands.cpp
index e90d4ac25ce..865a69dbc7b 100644
--- a/src/mongo/db/commands/write_commands.cpp
+++ b/src/mongo/db/commands/write_commands.cpp
@@ -458,6 +458,9 @@ void populateReply(OperationContext* opCtx,
 
     auto& replyBase = cmdReply->getWriteCommandReplyBase();
     replyBase.setN(nVal);
+    if (!result.retriedStmtIds.empty()) {
+        replyBase.setRetriedStmtIds(result.retriedStmtIds);
+    }
 
     if (!errors.empty()) {
         replyBase.setWriteErrors(errors);
diff --git a/src/mongo/db/ops/write_ops.idl b/src/mongo/db/ops/write_ops.idl
index 91c82db928b..cf58747c7c5 100644
--- a/src/mongo/db/ops/write_ops.idl
+++ b/src/mongo/db/ops/write_ops.idl
@@ -91,6 +91,12 @@ structs:
                 type: array<object_owned>
                 optional: true
                 unstable: false
+            retriedStmtIds:
+                description: "The statement numbers for the write statements that had already been
+                              executed, thus were not executed by this command."
+                type: array<int>
+                optional: true
+                unstable: true
 
     InsertCommandReply:
         description: "Contains information related to insert command reply."
@@ -298,6 +304,12 @@ structs:
                 description: "The document after the write, if the 'new' field of the request is
                               true. Otherwise, the document before the write."
                 unstable: false
+            retriedStmtId:
+                description: "The statement number for this findAndModify statement if it had
+                              already been executed, thus was not executed by this command."
+                type: int
+                optional: true
+                unstable: true
 
 commands:
 
diff --git a/src/mongo/db/ops/write_ops_exec.cpp b/src/mongo/db/ops/write_ops_exec.cpp
index 32213015ea7..748c70da401 100644
--- a/src/mongo/db/ops/write_ops_exec.cpp
+++ b/src/mongo/db/ops/write_ops_exec.cpp
@@ -724,6 +724,7 @@ WriteResult performInserts(OperationContext* opCtx,
         } else if (wasAlreadyExecuted) {
             containsRetry = true;
             RetryableWritesStats::get(opCtx)->incrementRetriedStatementsCount();
+            out.retriedStmtIds.push_back(stmtId);
             out.results.emplace_back(makeWriteResultForInsertOrDeleteRetry());
         }
     }
@@ -1022,6 +1023,7 @@ WriteResult performUpdates(OperationContext* opCtx,
                 containsRetry = true;
                 RetryableWritesStats::get(opCtx)->incrementRetriedStatementsCount();
                 out.results.emplace_back(parseOplogEntryForUpdate(*entry));
+                out.retriedStmtIds.push_back(stmtId);
                 continue;
             }
         }
@@ -1247,6 +1249,7 @@ WriteResult performDeletes(OperationContext* opCtx,
             containsRetry = true;
             RetryableWritesStats::get(opCtx)->incrementRetriedStatementsCount();
             out.results.emplace_back(makeWriteResultForInsertOrDeleteRetry());
+            out.retriedStmtIds.push_back(stmtId);
             continue;
         }
 
diff --git a/src/mongo/db/ops/write_ops_exec.h b/src/mongo/db/ops/write_ops_exec.h
index 377a440524f..548a3034713 100644
--- a/src/mongo/db/ops/write_ops_exec.h
+++ b/src/mongo/db/ops/write_ops_exec.h
@@ -56,6 +56,10 @@ struct WriteResult {
      */
     std::vector<StatusWith<SingleWriteResult>> results;
 
+    // Stores the statement ids for the ops that had already been executed, thus were not executed
+    // by this write.
+    std::vector<StmtId> retriedStmtIds;
+
     // In case of an error, whether the operation can continue.
     bool canContinue = true;
 };