SERVER-47155 Limit the number of simultaneous index builds running from user connections to 3

3 files changed, 101 insertions, 10 deletions

diff --git a/src/mongo/db/index_builds_coordinator_mongod.cpp b/src/mongo/db/index_builds_coordinator_mongod.cpp
index 2eda872cbea..48be752b76f 100644
--- a/src/mongo/db/index_builds_coordinator_mongod.cpp
+++ b/src/mongo/db/index_builds_coordinator_mongod.cpp
@@ -60,6 +60,8 @@ namespace {
 MONGO_FAIL_POINT_DEFINE(hangBeforeInitializingIndexBuild);
 MONGO_FAIL_POINT_DEFINE(hangAfterInitializingIndexBuild);
 
+const StringData kMaxNumActiveUserIndexBuildsServerParameterName = "maxNumActiveUserIndexBuilds"_sd;
+
 /**
  * Constructs the options for the loader thread pool.
  */
@@ -67,10 +69,13 @@ ThreadPool::Options makeDefaultThreadPoolOptions() {
     ThreadPool::Options options;
     options.poolName = "IndexBuildsCoordinatorMongod";
     options.minThreads = 0;
-    // We depend on thread pool sizes being equal between primaries and secondaries. If a secondary
-    // has fewer resources than a primary, index build oplog entries can replicate in an order that
-    // the secondary is unable to fulfill, leading to deadlocks. See SERVER-44250.
-    options.maxThreads = 3;
+    // Both the primary and secondary nodes will have an unlimited thread pool size. This is done to
+    // allow secondary nodes to startup as many index builders as necessary in order to prevent
+    // scheduling deadlocks during initial sync or oplog application. When commands are run from
+    // user connections that need to create indexes, those commands will hang until there are less
+    // than 'maxNumActiveUserIndexBuilds' running index build threads, or until the operation is
+    // interrupted.
+    options.maxThreads = ThreadPool::Options::kUnlimited;
 
     // Ensure all threads have a client.
     options.onCreateThread = [](const std::string& threadName) {
@@ -85,10 +90,18 @@ ThreadPool::Options makeDefaultThreadPoolOptions() {
 IndexBuildsCoordinatorMongod::IndexBuildsCoordinatorMongod()
     : _threadPool(makeDefaultThreadPoolOptions()) {
     _threadPool.startup();
-}
-IndexBuildsCoordinatorMongod::IndexBuildsCoordinatorMongod(ThreadPool::Options options)
-    : _threadPool(std::move(options)) {
-    _threadPool.startup();
+
+    // Change the 'setOnUpdate' function for the server parameter to signal the condition variable
+    // when the value changes.
+    ServerParameter* serverParam =
+        ServerParameterSet::getGlobal()->get(kMaxNumActiveUserIndexBuildsServerParameterName);
+    static_cast<
+        IDLServerParameterWithStorage<ServerParameterType::kStartupAndRuntime, AtomicWord<int>>*>(
+        serverParam)
+        ->setOnUpdate([this](const int) -> Status {
+            _indexBuildFinished.notify_all();
+            return Status::OK();
+        });
 }
 
 void IndexBuildsCoordinatorMongod::shutdown() {
@@ -110,6 +123,55 @@ IndexBuildsCoordinatorMongod::startIndexBuild(OperationContext* opCtx,
                                               const UUID& buildUUID,
                                               IndexBuildProtocol protocol,
                                               IndexBuildOptions indexBuildOptions) {
+    const NamespaceStringOrUUID nssOrUuid{dbName, collectionUUID};
+
+    {
+        // Only operations originating from user connections need to wait while there are more than
+        // 'maxNumActiveUserIndexBuilds' index builds currently running.
+        if (opCtx->getClient()->isFromUserConnection()) {
+            // Need to follow the locking order here by getting the global lock first followed by
+            // the mutex. The global lock acquires the RSTL lock which we use to assert that we're
+            // the primary node when running user operations.
+            ShouldNotConflictWithSecondaryBatchApplicationBlock shouldNotConflictBlock(
+                opCtx->lockState());
+            Lock::GlobalLock globalLk(opCtx, MODE_IX);
+
+            stdx::unique_lock<Latch> lk(_mutex);
+
+            auto replCoord = repl::ReplicationCoordinator::get(opCtx);
+            uassert(ErrorCodes::NotMaster,
+                    "Not primary while waiting to start an index build",
+                    replCoord->canAcceptWritesFor(opCtx, nssOrUuid));
+            opCtx->waitForConditionOrInterrupt(_indexBuildFinished, lk, [&] {
+                const int maxActiveBuilds = maxNumActiveUserIndexBuilds.load();
+                if (_numActiveIndexBuilds < maxActiveBuilds) {
+                    _numActiveIndexBuilds++;
+                    return true;
+                }
+
+                LOGV2(4715500,
+                      "Too many index builds running simultaneously, waiting until the number of "
+                      "active index builds is below the threshold",
+                      "numActiveIndexBuilds"_attr = _numActiveIndexBuilds,
+                      "maxNumActiveUserIndexBuilds"_attr = maxActiveBuilds,
+                      "indexSpecs"_attr = specs,
+                      "buildUUID"_attr = buildUUID,
+                      "collectionUUID"_attr = collectionUUID);
+                return false;
+            });
+        } else {
+            // System index builds have no limit and never wait, but do consume a slot.
+            stdx::unique_lock<Latch> lk(_mutex);
+            _numActiveIndexBuilds++;
+        }
+    }
+
+    auto onScopeExitGuard = makeGuard([&] {
+        stdx::unique_lock<Latch> lk(_mutex);
+        _numActiveIndexBuilds--;
+        _indexBuildFinished.notify_one();
+    });
+
     if (indexBuildOptions.twoPhaseRecovery) {
         // Two phase index build recovery goes though a different set-up procedure because the
         // original index will be dropped first.
@@ -150,7 +212,6 @@ IndexBuildsCoordinatorMongod::startIndexBuild(OperationContext* opCtx,
     const auto deadline = opCtx->getDeadline();
     const auto timeoutError = opCtx->getTimeoutError();
 
-    const NamespaceStringOrUUID nssOrUuid{dbName, collectionUUID};
     const auto nss = CollectionCatalog::get(opCtx).resolveNamespaceStringOrUUID(opCtx, nssOrUuid);
 
     const auto& oss = OperationShardingState::get(opCtx);
@@ -179,6 +240,10 @@ IndexBuildsCoordinatorMongod::startIndexBuild(OperationContext* opCtx,
 
 
     auto replState = invariant(_getIndexBuild(buildUUID));
+
+    // The thread pool task will be responsible for signalling the condition variable when the index
+    // build thread is done running.
+    onScopeExitGuard.dismiss();
     _threadPool.schedule([
         this,
         buildUUID,
@@ -195,6 +260,12 @@ IndexBuildsCoordinatorMongod::startIndexBuild(OperationContext* opCtx,
         shardVersion,
         dbVersion
     ](auto status) mutable noexcept {
+        auto onScopeExitGuard = makeGuard([&] {
+            stdx::unique_lock<Latch> lk(_mutex);
+            _numActiveIndexBuilds--;
+            _indexBuildFinished.notify_one();
+        });
+
         // Clean up if we failed to schedule the task.
         if (!status.isOK()) {
             stdx::unique_lock<Latch> lk(_mutex);
diff --git a/src/mongo/db/index_builds_coordinator_mongod.h b/src/mongo/db/index_builds_coordinator_mongod.h
index 8dd2335dd83..7596a4ddb97 100644
--- a/src/mongo/db/index_builds_coordinator_mongod.h
+++ b/src/mongo/db/index_builds_coordinator_mongod.h
@@ -55,7 +55,6 @@ public:
      * Sets up the thread pool.
      */
     IndexBuildsCoordinatorMongod();
-    IndexBuildsCoordinatorMongod(ThreadPool::Options options);
 
     /**
      * Shuts down the thread pool, signals interrupt to all index builds, then waits for all of the
@@ -160,6 +159,12 @@ private:
 
     // Thread pool on which index builds are run.
     ThreadPool _threadPool;
+
+    // Protected by _mutex.
+    int _numActiveIndexBuilds = 0;
+
+    // Condition signalled to indicate that an index build thread finished executing.
+    stdx::condition_variable _indexBuildFinished;
 };
 
 }  // namespace mongo
diff --git a/src/mongo/db/storage/two_phase_index_build_knobs.idl b/src/mongo/db/storage/two_phase_index_build_knobs.idl
index dbcdc45d1e3..c2a176c91f5 100644
--- a/src/mongo/db/storage/two_phase_index_build_knobs.idl
+++ b/src/mongo/db/storage/two_phase_index_build_knobs.idl
@@ -46,3 +46,18 @@ server_parameters:
     cpp_vartype: bool
     cpp_varname: "enableIndexBuildCommitQuorum"
     default: true
+
+  maxNumActiveUserIndexBuilds:
+    description: >
+      Specifies the maximum number of active user index builds that can be built simultaneously on
+      the primary node. Index builds initiated by the system are not subject to this limitation.
+
+      Additionally, active index builds initiated by the system count towards the limit and can
+      delay scheduling user index builds even when the number of active user index builds is below
+      the limit.
+    set_at: [ startup, runtime ]
+    cpp_vartype: AtomicWord<int>
+    cpp_varname: maxNumActiveUserIndexBuilds
+    default: 3
+    validator:
+      gte: 0