diff --git a/CHANGELOG b/CHANGELOG
index 914f53ea13..f091a27cca 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -30,6 +30,7 @@ devel
 
 * fixed internal issue #3919: The web UI is now using precompiled ejs templates.
 
+* fixed agency issue in abort of cleanOutServer job
 
 v3.5.0-rc.2 (2019-05-23)
 ------------------------
diff --git a/arangod/Agency/FailedServer.cpp b/arangod/Agency/FailedServer.cpp
index 83267e9c0d..60119a04b3 100644
--- a/arangod/Agency/FailedServer.cpp
+++ b/arangod/Agency/FailedServer.cpp
@@ -179,7 +179,7 @@ bool FailedServer::start(bool& aborts) {
               for (auto const& it : VPackArrayIterator(shard.second->slice())) {
                 auto dbs = it.copyString();
 
-                if (dbs == _server) {
+                if (dbs == _server || dbs == "_" + _server) {
                   if (pos == 0) {
                     FailedLeader(
                       _snapshot, _agent, _jobId + "-" + std::to_string(sub++),
diff --git a/arangod/Agency/Supervision.cpp b/arangod/Agency/Supervision.cpp
index 71a1ff1ec6..eee11d7c6d 100644
--- a/arangod/Agency/Supervision.cpp
+++ b/arangod/Agency/Supervision.cpp
@@ -684,8 +684,8 @@ bool Supervision::updateSnapshot() {
   }
 
   _agent->executeLockedRead([&]() {
-    if (_agent->readDB().has(_agencyPrefix)) {
-      _snapshot = _agent->readDB().get(_agencyPrefix);
+    if (_agent->spearhead().has(_agencyPrefix)) {
+      _snapshot = _agent->spearhead().get(_agencyPrefix);
     }
     if (_agent->transient().has(_agencyPrefix)) {
       _transient = _agent->transient().get(_agencyPrefix);
@@ -1211,7 +1211,7 @@ void Supervision::workJobs() {
   // per second. Therefore, we have - for now - chosen to limit the number of
   // jobs actually worked on to 1000 in ToDo and 1000 in Pending. However,
   // since some jobs are just waiting, we cannot work on the same 1000
-  // jobs in each round. This is where the randomization comes in. We work 
+  // jobs in each round. This is where the randomization comes in. We work
   // on up to 1000 *random* jobs. This will eventually cover everything with
   // very high probability. Note that the snapshot does not change, so
   // `todos.size()` is constant for the loop, even though we do agency
@@ -1222,14 +1222,14 @@ void Supervision::workJobs() {
   LOG_TOPIC("00567", TRACE, Logger::SUPERVISION) << "Begin ToDos of type Failed*";
   bool doneFailedJob = false;
   while (it != todos.end()) {
-    if (selectRandom && RandomGenerator::interval(static_cast<uint64_t>(todos.size())) > maximalJobsPerRound) {
-      LOG_TOPIC("675fe", TRACE, Logger::SUPERVISION) << "Skipped ToDo Job";
-      ++it;
-      continue;
-    }
- 
     auto const& jobNode = *(it->second);
     if (jobNode.hasAsString("type").first.compare(0, FAILED.length(), FAILED) == 0) {
+      if (selectRandom && RandomGenerator::interval(static_cast<uint64_t>(todos.size())) > maximalJobsPerRound) {
+        LOG_TOPIC("675fe", TRACE, Logger::SUPERVISION) << "Skipped ToDo Job";
+        ++it;
+        continue;
+      }
+
       LOG_TOPIC("87812", TRACE, Logger::SUPERVISION) << "Begin JobContext::run()";
       JobContext(TODO, jobNode.hasAsString("jobId").first, _snapshot, _agent)
         .run(_haveAborts);
@@ -1258,7 +1258,6 @@ void Supervision::workJobs() {
       }
     }
   }
-  
   LOG_TOPIC("a55ce", DEBUG, Logger::SUPERVISION) << "Updating snapshot after ToDo";
   updateSnapshot();