SERVER-61709 Introduce implicitly replicated namespaces

Implicitly replicated namespaces are internal namespaces that do not replicate
writes, with the exception of deletions, user-initiated direct writes and some
maintenance operations.

This patch lists config.system.preimages, config.images_collection,
config.transactions and config.changes.* as implicitly replicated namespaces,
and unifies and validates their semantics. It also special-cases some of the
config.transactions replication behaviour that is too specific to be generalised.

11 files changed, 60 insertions, 6 deletions

diff --git a/src/mongo/db/commands/set_feature_compatibility_version_command.cpp b/src/mongo/db/commands/set_feature_compatibility_version_command.cpp
index d7d6d9f6d8e..a1aa4f78199 100644
--- a/src/mongo/db/commands/set_feature_compatibility_version_command.cpp
+++ b/src/mongo/db/commands/set_feature_compatibility_version_command.cpp
@@ -711,7 +711,8 @@ private:
             // Due to the possibility that the shell or drivers have implicit sessions enabled, we
             // cannot write to the config.transactions collection while we're in a session. So we
             // construct a temporary client to as a work around.
-            auto newClient = opCtx->getServiceContext()->makeClient("InternalSessionsCleanup");
+            auto newClient = opCtx->getServiceContext()->makeClient(
+                SessionCatalog::kInternalSessionsCleanupClient.toString());
 
             {
                 stdx::lock_guard<Client> lk(*newClient.get());
diff --git a/src/mongo/db/exec/update_stage.cpp b/src/mongo/db/exec/update_stage.cpp
index bda0a81078b..d6e53a345ef 100644
--- a/src/mongo/db/exec/update_stage.cpp
+++ b/src/mongo/db/exec/update_stage.cpp
@@ -52,6 +52,7 @@
 #include "mongo/db/s/sharding_state.h"
 #include "mongo/db/s/sharding_write_router.h"
 #include "mongo/db/service_context.h"
+#include "mongo/db/session_catalog.h"
 #include "mongo/db/storage/duplicate_key_error_info.h"
 #include "mongo/db/update/path_support.h"
 #include "mongo/db/update/storage_validation.h"
@@ -372,6 +373,17 @@ PlanStage::StageState UpdateStage::doWork(WorkingSetID* out) {
         return PlanStage::IS_EOF;
     }
 
+    boost::optional<repl::UnreplicatedWritesBlock> unReplBlock;
+    const auto isSessionCleanupClient =
+        opCtx()->getClient()->desc() == SessionCatalog::kInternalSessionsCleanupClient;
+    if (collection()->ns().isImplicitlyReplicated() && !_isUserInitiatedWrite &&
+        !isSessionCleanupClient) {
+        // Implictly replicated collections do not replicate updates.
+        // However, user-initiated writes and some background maintenance tasks are allowed
+        // to replicate as they cannot be derived from the oplog.
+        unReplBlock.emplace(opCtx());
+    }
+
     // It is possible that after an update was applied, a WriteConflictException
     // occurred and prevented us from returning ADVANCED with the requested version
     // of the document.
diff --git a/src/mongo/db/exec/upsert_stage.cpp b/src/mongo/db/exec/upsert_stage.cpp
index a6f6bf73d04..f1b684a544d 100644
--- a/src/mongo/db/exec/upsert_stage.cpp
+++ b/src/mongo/db/exec/upsert_stage.cpp
@@ -35,6 +35,7 @@
 #include "mongo/db/curop_failpoint_helpers.h"
 #include "mongo/db/query/query_feature_flags_gen.h"
 #include "mongo/db/s/operation_sharding_state.h"
+#include "mongo/db/session_catalog.h"
 #include "mongo/db/update/storage_validation.h"
 #include "mongo/s/would_change_owning_shard_exception.h"
 
@@ -72,6 +73,14 @@ PlanStage::StageState UpsertStage::doWork(WorkingSetID* out) {
         return StageState::IS_EOF;
     }
 
+    boost::optional<repl::UnreplicatedWritesBlock> unReplBlock;
+    const auto isSessionCleanupClient =
+        opCtx()->getClient()->desc() == SessionCatalog::kInternalSessionsCleanupClient;
+    if (collection()->ns().isImplicitlyReplicated() && !isSessionCleanupClient) {
+        // Implictly replicated collections do not replicate updates.
+        unReplBlock.emplace(opCtx());
+    }
+
     // First, attempt to perform the update on a matching document.
     auto updateState = UpdateStage::doWork(out);
 
diff --git a/src/mongo/db/namespace_string.cpp b/src/mongo/db/namespace_string.cpp
index 631177c4d63..f1e258ba2ce 100644
--- a/src/mongo/db/namespace_string.cpp
+++ b/src/mongo/db/namespace_string.cpp
@@ -347,6 +347,10 @@ bool NamespaceString::isConfigImagesCollection() const {
     return ns() == kConfigImagesNamespace.ns();
 }
 
+bool NamespaceString::isConfigTransactionsCollection() const {
+    return ns() == kSessionTransactionsTableNamespace.ns();
+}
+
 NamespaceString NamespaceString::makeTimeseriesBucketsNamespace() const {
     return {db(), kTimeseriesBucketsCollectionPrefix.toString() + coll()};
 }
@@ -356,6 +360,18 @@ NamespaceString NamespaceString::getTimeseriesViewNamespace() const {
     return {db(), coll().substr(kTimeseriesBucketsCollectionPrefix.size())};
 }
 
+bool NamespaceString::isImplicitlyReplicated() const {
+    if (isChangeStreamPreImagesCollection() || isConfigImagesCollection() ||
+        isConfigTransactionsCollection() || isChangeCollection()) {
+        // Implicitly replicated namespaces are replicated, although they only replicate a subset of
+        // writes.
+        invariant(isReplicated());
+        return true;
+    }
+
+    return false;
+}
+
 bool NamespaceString::isReplicated() const {
     if (isLocal()) {
         return false;
diff --git a/src/mongo/db/namespace_string.h b/src/mongo/db/namespace_string.h
index 17ea8855bc7..339c62ce23f 100644
--- a/src/mongo/db/namespace_string.h
+++ b/src/mongo/db/namespace_string.h
@@ -307,6 +307,9 @@ public:
     bool isSystemDotProfile() const {
         return coll() == "system.profile";
     }
+    bool isChangeCollection() const {
+        return (db() == kConfigDb) && coll().startsWith("changes.");
+    }
     bool isSystemDotViews() const {
         return coll() == kSystemDotViewsCollectionName;
     }
@@ -384,6 +387,11 @@ public:
     bool isConfigImagesCollection() const;
 
     /**
+     * Returns whether the specified namespace is config.transactions.
+     */
+    bool isConfigTransactionsCollection() const;
+
+    /**
      * Returns the time-series buckets namespace for this view.
      */
     NamespaceString makeTimeseriesBucketsNamespace() const;
@@ -394,6 +402,15 @@ public:
     NamespaceString getTimeseriesViewNamespace() const;
 
     /**
+     * Returns whether the namespace is implicitly replicated, based only on its string value.
+     *
+     * An implicitly replicated namespace is an internal namespace which does not replicate writes
+     * via the oplog, with the exception of deletions. Writes are not replicated as an optimization
+     * because their content can be reliably derived from entries in the oplog.
+     */
+    bool isImplicitlyReplicated() const;
+
+    /**
      * Returns whether a namespace is replicated, based only on its string value. One notable
      * omission is that map reduce `tmp.mr` collections may or may not be replicated. Callers must
      * decide how to handle that case separately.
diff --git a/src/mongo/db/op_observer_impl.cpp b/src/mongo/db/op_observer_impl.cpp
index 398c577df06..51bf478673f 100644
--- a/src/mongo/db/op_observer_impl.cpp
+++ b/src/mongo/db/op_observer_impl.cpp
@@ -297,7 +297,6 @@ void writeToImageCollection(OperationContext* opCtx,
     imageEntry.setImageKind(imageKind);
     imageEntry.setImage(dataImage);
 
-    repl::UnreplicatedWritesBlock unreplicated(opCtx);
     DisableDocumentValidation documentValidationDisabler(
         opCtx, DocumentValidationSettings::kDisableInternalValidation);
 
diff --git a/src/mongo/db/op_observer_impl_test.cpp b/src/mongo/db/op_observer_impl_test.cpp
index 373d979feab..d0273e162ea 100644
--- a/src/mongo/db/op_observer_impl_test.cpp
+++ b/src/mongo/db/op_observer_impl_test.cpp
@@ -187,6 +187,7 @@ public:
         reset(opCtx, NamespaceString::kRsOplogNamespace);
         reset(opCtx, NamespaceString::kSessionTransactionsTableNamespace);
         reset(opCtx, NamespaceString::kConfigImagesNamespace);
+        reset(opCtx, NamespaceString::kChangeStreamPreImagesNamespace);
     }
 
 protected:
diff --git a/src/mongo/db/pipeline/change_stream_pre_image_helpers.cpp b/src/mongo/db/pipeline/change_stream_pre_image_helpers.cpp
index 01335139a8d..68e2e6949f8 100644
--- a/src/mongo/db/pipeline/change_stream_pre_image_helpers.cpp
+++ b/src/mongo/db/pipeline/change_stream_pre_image_helpers.cpp
@@ -48,7 +48,6 @@ void writeToChangeStreamPreImagesCollection(OperationContext* opCtx,
     // This lock acquisition can block on a stronger lock held by another operation modifying the
     // pre-images collection. There are no known cases where an operation holding an exclusive lock
     // on the pre-images collection also waits for oplog visibility.
-    repl::UnreplicatedWritesBlock unreplicated(opCtx);
     AllowLockAcquisitionOnTimestampedUnitOfWork allowLockAcquisition(opCtx->lockState());
     AutoGetCollection preimagesCollectionRaii(opCtx, collectionNamespace, LockMode::MODE_IX);
     UpdateResult res = Helpers::upsert(opCtx, collectionNamespace.toString(), preImage.toBSON());
diff --git a/src/mongo/db/repl/dbcheck.cpp b/src/mongo/db/repl/dbcheck.cpp
index 79a98205972..cb909bbd08e 100644
--- a/src/mongo/db/repl/dbcheck.cpp
+++ b/src/mongo/db/repl/dbcheck.cpp
@@ -206,8 +206,7 @@ std::unique_ptr<HealthLogEntry> dbCheckBatchEntry(
         }
         // Implcitily replicated collections and capped collections not replicating truncation are
         // not designed to be consistent, so inconsistency is not necessarily pathological.
-        if (nss.isChangeStreamPreImagesCollection() || nss.isConfigImagesCollection() ||
-            (options && options->capped)) {
+        if (nss.isImplicitlyReplicated() || (options && options->capped)) {
             return SeverityEnum::Warning;
         }
 
diff --git a/src/mongo/db/repl/oplog.cpp b/src/mongo/db/repl/oplog.cpp
index ea9b83dc41b..ec5933b830d 100644
--- a/src/mongo/db/repl/oplog.cpp
+++ b/src/mongo/db/repl/oplog.cpp
@@ -297,7 +297,6 @@ void writeToImageCollection(OperationContext* opCtx,
     request.setFromOplogApplication(true);
     try {
         // This code path can also be hit by things such as `applyOps` and tenant migrations.
-        repl::UnreplicatedWritesBlock dontReplicate(opCtx);
         ::mongo::update(opCtx, autoColl.getDb(), request);
     } catch (const ExceptionFor<ErrorCodes::DuplicateKey>&) {
         // We can get a duplicate key when two upserts race on inserting a document.
diff --git a/src/mongo/db/session_catalog.h b/src/mongo/db/session_catalog.h
index 7bed6ae91a3..bcf696042b7 100644
--- a/src/mongo/db/session_catalog.h
+++ b/src/mongo/db/session_catalog.h
@@ -59,6 +59,8 @@ class SessionCatalog {
     friend class OperationContextSession;
 
 public:
+    static constexpr StringData kInternalSessionsCleanupClient = "InternalSessionsCleanup"_sd;
+
     class ScopedCheckedOutSession;
     class SessionToKill;