Fix agency shutdown bug. (#3683)

* Fix agency shutdown bug.
* Remove precondition that was not needed in AgencyComm::removeValues.
* Fail fatally if threads do not shut down.

arangod/Agency/AgencyComm.cpp

@@ -913,8 +913,7 @@ AgencyCommResult AgencyComm::getValues(std::string const& key) {
 AgencyCommResult AgencyComm::removeValues(std::string const& key,
                                           bool recursive) {
   AgencyWriteTransaction transaction(
-      AgencyOperation(key, AgencySimpleOperationType::DELETE_OP),
+      AgencyOperation(key, AgencySimpleOperationType::DELETE_OP));
-      AgencyPrecondition(key, AgencyPrecondition::Type::EMPTY, false));
 
   return sendTransactionWithFailover(transaction);
 }

arangod/Agency/AgencyFeature.cpp

@@ -333,11 +333,22 @@ void AgencyFeature::stop() {
       }
     }
 
-    // delete the Agent object here ensures it shuts down all of its threads
+    // Wait until all agency threads have been shut down. Note that the
-    // this is a precondition that it must fulfill before we can go on with the
+    // actual agent object is only destroyed in the destructor to allow
-    // shutdown
+    // server jobs from RestAgencyHandlers to complete without incident:
-    _agent.reset();
+    _agent->waitForThreadsStop();
   }
 
   AGENT = nullptr;
 }
+
+void AgencyFeature::unprepare() {
+  if (!isEnabled()) {
+    return;
+  }
+  // delete the Agent object here ensures it shuts down all of its threads
+  // this is a precondition that it must fulfill before we can go on with the
+  // shutdown
+  _agent.reset();
+}
+

arangod/Agency/AgencyFeature.h

@@ -45,6 +45,7 @@ class AgencyFeature : virtual public application_features::ApplicationFeature {
   void start() override final;
   void beginShutdown() override final;
   void stop() override final;
+  void unprepare() override final;
 
  private:
   bool _activated;

arangod/Agency/Agent.cpp

@@ -86,22 +86,31 @@ bool Agent::mergeConfiguration(VPackSlice const& persisted) {
 
 /// Dtor shuts down thread
 Agent::~Agent() {
+  waitForThreadsStop();
+  // This usually was already done called from AgencyFeature::unprepare,
+  // but since this only waits for the threads to stop, it can be done
+  // multiple times, and we do it just in case the Agent object was
+  // created but never really started. Here, we exit with a fatal error
+  // if the threads do not stop in time.
+}
 
-  // Give up if some subthread breaks shutdown
+/// Wait until threads are terminated:
+void Agent::waitForThreadsStop() {
+  // It is allowed to call this multiple times, we do so from the constructor
+  // and from AgencyFeature::unprepare.
   int counter = 0;
   while (_constituent.isRunning() || _compactor.isRunning() ||
          (_config.supervision() && _supervision.isRunning()) ||
          (_inception != nullptr && _inception->isRunning())) {
     usleep(100000);
 
-    // emit warning after 15 seconds
+    // fail fatally after 5 mins:
-    if (++counter == 10 * 15) {
+    if (++counter >= 10 * 60 * 5) {
       LOG_TOPIC(FATAL, Logger::AGENCY) << "some agency thread did not finish";
       FATAL_ERROR_EXIT();
     }
   }
-    
+  shutdown();  // wait for the main Agent thread to terminate
-  shutdown();
 }
 
 /// State machine

arangod/Agency/Agent.h

@@ -51,6 +51,9 @@ class Agent : public arangodb::Thread,
   /// @brief Clean up
   ~Agent();
 
+  /// @brief bring down threads, can be called multiple times.
+  void waitForThreadsStop();
+
   /// @brief Get current term
   term_t term() const;
 

js/server/tests/resilience/resilience-synchronous-repl-cluster.js

@@ -344,7 +344,6 @@ function SynchronousReplicationSuite () {
 
     tearDown : function () {
       db._drop(cn);
-      global.ArangoAgency.remove('Target/FailedServers');
       global.ArangoAgency.set('Target/FailedServers', {});
     },