SERVER-55359 pre/post image ops after startFetchingDonorOpTime should be filtered out of oplog buffer

(cherry picked from commit 9457ca34a5e157698602e620042c6c8037023375)

6 files changed, 101 insertions, 41 deletions

diff --git a/buildscripts/resmokeconfig/suites/tenant_migration_kill_primary_jscore_passthrough.yml b/buildscripts/resmokeconfig/suites/tenant_migration_kill_primary_jscore_passthrough.yml
index 39e9f112e13..37072afb156 100644
--- a/buildscripts/resmokeconfig/suites/tenant_migration_kill_primary_jscore_passthrough.yml
+++ b/buildscripts/resmokeconfig/suites/tenant_migration_kill_primary_jscore_passthrough.yml
@@ -120,9 +120,6 @@ selector:
   # Blacklists specific to this suite
   #
 
-  # TODO (SERVER-55359): Docs in tenant migration oplog buffer collection might not be ordered by
-  # timestamp.
-  - jstests/core/*find_and_modify*.js
   - jstests/core/api_version_parameters.js
   - jstests/core/benchrun_pipeline_updates.js
   - jstests/core/bypass_doc_validation.js
diff --git a/buildscripts/resmokeconfig/suites/tenant_migration_stepdown_jscore_passthrough.yml b/buildscripts/resmokeconfig/suites/tenant_migration_stepdown_jscore_passthrough.yml
index 2cc25669001..a6c9a598184 100644
--- a/buildscripts/resmokeconfig/suites/tenant_migration_stepdown_jscore_passthrough.yml
+++ b/buildscripts/resmokeconfig/suites/tenant_migration_stepdown_jscore_passthrough.yml
@@ -120,9 +120,6 @@ selector:
   # Blacklists specific to this suite
   #
 
-  # TODO (SERVER-55359): Docs in tenant migration oplog buffer collection might not be ordered by
-  # timestamp.
-  - jstests/core/*find_and_modify*.js
   - jstests/core/api_version_parameters.js
   - jstests/core/benchrun_pipeline_updates.js
   - jstests/core/bypass_doc_validation.js
diff --git a/buildscripts/resmokeconfig/suites/tenant_migration_terminate_primary_jscore_passthrough.yml b/buildscripts/resmokeconfig/suites/tenant_migration_terminate_primary_jscore_passthrough.yml
index 668cb84c2db..08702a8cbbc 100644
--- a/buildscripts/resmokeconfig/suites/tenant_migration_terminate_primary_jscore_passthrough.yml
+++ b/buildscripts/resmokeconfig/suites/tenant_migration_terminate_primary_jscore_passthrough.yml
@@ -120,9 +120,6 @@ selector:
   # Blacklists specific to this suite
   #
 
-  # TODO (SERVER-55359): Docs in tenant migration oplog buffer collection might not be ordered by
-  # timestamp.
-  - jstests/core/*find_and_modify*.js
   - jstests/core/api_version_parameters.js
   - jstests/core/benchrun_pipeline_updates.js
   - jstests/core/bypass_doc_validation.js
diff --git a/jstests/replsets/tenant_migration_recipient_fetches_retryable_writes_oplog_entries.js b/jstests/replsets/tenant_migration_recipient_fetches_retryable_writes_oplog_entries.js
index 9d32c1b1f60..98ac4e63c62 100644
--- a/jstests/replsets/tenant_migration_recipient_fetches_retryable_writes_oplog_entries.js
+++ b/jstests/replsets/tenant_migration_recipient_fetches_retryable_writes_oplog_entries.js
@@ -139,7 +139,21 @@ batchInsertWorker.join();
 // `lastWriteOpTime` is after `startFetchingDonorOpTime`. The corresponding oplog entries should
 // not be added to the oplog buffer.
 assert.commandWorked(
-    tenantCollection3.insert({_id: "retryableWrite4"}, {writeConcern: {w: "majority"}}));
+    tenantCollection3.insert({_id: "retryableWrite4", count: 0}, {writeConcern: {w: "majority"}}));
+
+// Test that when a post image op's `postImageOpTime` is after `startFetchingDonorOpTime`, it gets
+// filtered out.
+tenantCollection3.findAndModify({
+    query: {_id: "retryableWrite4"},
+    update: {$inc: {count: 1}},
+    new: true,
+    writeConcern: {w: "majority"}
+});
+
+// Test that when a pre image op's `preImageOpTime` is after `startFetchingDonorOpTime`, it gets
+// filtered out.
+tenantCollection3.findAndModify(
+    {query: {_id: "retryableWrite4"}, remove: true, writeConcern: {w: "majority"}});
 
 fpAfterRetrievingStartOpTime.off();
 fpAfterRetrievingRetryableWrites.wait();
@@ -160,6 +174,7 @@ assert.eq(1, recipientOplogBuffer.find({"entry.o._id": "bulkRetryableWrite0"}).i
 // Ensure the retryable write oplog entries that should not be in `kOplogBufferNS` are in fact not.
 assert.eq(0, recipientOplogBuffer.find({"entry.o._id": "retryableWrite1"}).itcount());
 assert.eq(0, recipientOplogBuffer.find({"entry.o._id": "retryableWrite4"}).itcount());
+assert.eq(0, recipientOplogBuffer.find({"entry.o2._id": "retryableWrite4"}).itcount());
 assert.eq(0, recipientOplogBuffer.find({"entry.o._id": "bulkRetryableWrite1"}).itcount());
 
 fpAfterRetrievingRetryableWrites.off();
diff --git a/jstests/replsets/tenant_migration_retryable_write_retry.js b/jstests/replsets/tenant_migration_retryable_write_retry.js
index 9559ab57dc6..0bd659a2d6c 100644
--- a/jstests/replsets/tenant_migration_retryable_write_retry.js
+++ b/jstests/replsets/tenant_migration_retryable_write_retry.js
@@ -224,6 +224,20 @@ assert.commandWorked(donorPrimary.getDB(kDbName).runCommand({
     writeConcern: {w: "majority"}
 }));
 
+const lsid8 = {
+    id: UUID()
+};
+const sessionTag8 = "retryable findAndModify update after migration";
+assert.commandWorked(donorPrimary.getDB(kDbName).runCommand({
+    findAndModify: kCollName,
+    query: {x: 7},
+    update: {$set: {tag: sessionTag8}},
+    new: true,
+    txnNumber: NumberLong(0),
+    lsid: lsid8,
+    writeConcern: {w: "majority"}
+}));
+
 // The aggregation pipeline will return an array of retryable writes oplog entries (pre-image/
 // post-image oplog entries included) with "ts" < "startFetchingTimestamp" and sorted in ascending
 // order of "ts".
@@ -261,20 +275,40 @@ const aggRes = donorPrimary.getDB("config").runCommand({
         // single-element 'lastOps' array field with a single 'lastOp' field.
         {$addFields: {lastOp: {$first: "$lastOps"}}},
         {$unset: "lastOps"},
-        // Fetch preImage oplog entry for findAndModify from the oplog view.
+        // Fetch the preImage oplog entry for findAndModify from the oplog view only if it occurred
+        // before `startFetchingTimestamp`.
         {$lookup: {
             from: {db: "local", coll: "system.tenantMigration.oplogView"},
-            localField: "lastOp.preImageOpTime.ts",
-            foreignField: "ts",
+            let: { preimage_ts: "$lastOp.preImageOpTime.ts"},
+            pipeline: [{
+                $match: {
+                    $expr: {
+                        $and: [
+                            {$eq: [ "$ts", "$$preimage_ts"]},
+                            {$lt: ["$ts", startFetchingTimestamp]}
+                        ]
+                    }
+                }
+            }],
             // This array is expected to contain exactly one element if the 'preImageOpTime'
             // field is not null.
             as: "preImageOps"
         }},
-        // Fetch postImage oplog entry for findAndModify from the oplog view.
+        // Fetch the postImage oplog entry for findAndModify from the oplog view only if it occurred
+        // before `startFetchingTimestamp`.
         {$lookup: {
             from: {db: "local", coll: "system.tenantMigration.oplogView"},
-            localField: "lastOp.postImageOpTime.ts",
-            foreignField: "ts",
+            let: { postimage_ts: "$lastOp.postImageOpTime.ts"},
+            pipeline: [{
+                $match: {
+                    $expr: {
+                        $and: [
+                            {$eq: [ "$ts", "$$postimage_ts"]},
+                            {$lt: ["$ts", startFetchingTimestamp]}
+                        ]
+                    }
+                }
+            }],
             // This array is expected to contain exactly one element if the 'postImageOpTime'
             // field is not null.
             as: "postImageOps"
diff --git a/src/mongo/db/repl/tenant_migration_util.cpp b/src/mongo/db/repl/tenant_migration_util.cpp
index 9989cf4cb9d..860fc442eea 100644
--- a/src/mongo/db/repl/tenant_migration_util.cpp
+++ b/src/mongo/db/repl/tenant_migration_util.cpp
@@ -249,31 +249,51 @@ createRetryableWritesOplogFetchingPipelineForTenantMigrations(
     // 5. Remove `lastOps` in favor of `lastOp`.
     stages.emplace_back(DocumentSourceProject::createUnset(FieldPath("lastOps"), expCtx));
 
-    // 6. Fetch preImage oplog entry for `findAndModify` from the oplog view. `preImageOps` is not
-    //    expected to contain exactly one element if the `preImageOpTime` field is not null.
-    stages.emplace_back(DocumentSourceLookUp::createFromBson(
-        Doc{{"$lookup",
-             Doc{{"from", Doc{{"db", "local"_sd}, {"coll", "system.tenantMigration.oplogView"_sd}}},
-                 {"localField", "lastOp.preImageOpTime.ts"_sd},
-                 {"foreignField", "ts"_sd},
-                 {"as", "preImageOps"_sd}}}}
-            .toBson()
-            .firstElement(),
-        expCtx));
-
-
-    // 7. Fetch postImage oplog entry for `findAndModify` from the oplog view. `postImageOps` is not
-    //    expected to contain exactly one element if the `postImageOpTime` field is not null.
-    stages.emplace_back(DocumentSourceLookUp::createFromBson(
-        Doc{{"$lookup",
-             Doc{{"from", Doc{{"db", "local"_sd}, {"coll", "system.tenantMigration.oplogView"_sd}}},
-                 {"localField", "lastOp.postImageOpTime.ts"_sd},
-                 {"foreignField", "ts"_sd},
-                 {"as", "postImageOps"_sd}}}}
-            .toBson()
-            .firstElement(),
-        expCtx));
+    // 6. Fetch preImage oplog entry for `findAndModify` from the oplog view. `preImageOps` is
+    //    expected to contain exactly one element if the `preImageOpTime` field is not null and is
+    //    earlier than `startFetchingTimestamp`.
+    stages.emplace_back(DocumentSourceLookUp::createFromBson(fromjson("{\
+                    $lookup: {\
+                        from: {db: 'local', coll: 'system.tenantMigration.oplogView'},\
+                        let: { preimage_ts: '$lastOp.preImageOpTime.ts'},\
+                        pipeline: [{\
+                            $match: {\
+                                $expr: {\
+                                    $and: [\
+                                        {$eq: ['$ts', '$$preimage_ts']},\
+                                        {$lt: ['$ts', " + startFetchingTimestamp.toString() +
+                                                                      "]}\
+                                    ]\
+                                }\
+                            }\
+                        }],\
+                        as: 'preImageOps'\
+                    }}")
+                                                                 .firstElement(),
+                                                             expCtx));
 
+    // 7. Fetch postImage oplog entry for `findAndModify` from the oplog view. `postImageOps` is
+    //    expected to contain exactly one element if the `postImageOpTime` field is not null and is
+    //    earlier than `startFetchingTimestamp`.
+    stages.emplace_back(DocumentSourceLookUp::createFromBson(fromjson("{\
+                    $lookup: {\
+                        from: {db: 'local', coll: 'system.tenantMigration.oplogView'},\
+                        let: { postimage_ts: '$lastOp.postImageOpTime.ts'},\
+                        pipeline: [{\
+                            $match: {\
+                                $expr: {\
+                                    $and: [\
+                                        {$eq: ['$ts', '$$postimage_ts']},\
+                                        {$lt: ['$ts', " + startFetchingTimestamp.toString() +
+                                                                      "]}\
+                                    ]\
+                                }\
+                            }\
+                        }],\
+                        as: 'postImageOps'\
+                    }}")
+                                                                 .firstElement(),
+                                                             expCtx));
 
     // 8. Fetch oplog entries in each chain from the oplog view.
     stages.emplace_back(DocumentSourceGraphLookUp::createFromBson(
@@ -288,7 +308,7 @@ createRetryableWritesOplogFetchingPipelineForTenantMigrations(
             .firstElement(),
         expCtx));
 
-    // 9. Filter out all oplog entries from the `history` arrary that occur after
+    // 9. Filter out all oplog entries from the `history` array that occur after
     //    `startFetchingTimestamp`. Since the oplog fetching and application stages will already
     //    capture entries after `startFetchingTimestamp`, we only need the earlier part of the oplog
     //    chain.