added "peakMemoryUsage" query result figure (#7952)

CHANGELOG

@@ -1,6 +1,10 @@
 devel
 -----
 
+* added "peakMemoryUsage" in query results figures, showing the peak memory
+  usage of the executed query. In a cluster, the value the peak memory usage
+  of all shards, but it is not summed up across shards.
+
 * fixed an issue where a crashed coordinator can lead to some Foxx queue jobs
   erroneously either left hanging or being restarted
 
@@ -12,7 +16,7 @@ devel
 
   millisecond parts of AQL date values were limited to up to 3 digits.
   Now the length of the millisecond part is unrestricted, but the 
-  millisecond precision is still limit to up to 3 digits.
+  millisecond precision is still limited to up to 3 digits.
 
 * fix issue #7900: Bind values of `null` are not replaced by empty string 
   anymore, when toggling between JSON and table view in the web UI

Documentation/Books/AQL/ExecutionAndPerformance/QueryStatistics.md

@@ -47,6 +47,13 @@ The meaning of the statistics attributes is as follows:
   This attribute may only be returned if the `fullCount` option was set when starting the 
   query and will only contain a sensible value if the query contained a `LIMIT` operation on
   the top level.
+* *peakMemoryUsage*: the maximum memory usage of the query while it was running. In a cluster,
+  the memory accounting is done per shard, and the memory usage reported is the peak
+  memory usage value from the individual shards.
+  Note that to keep things light-weight, the per-query memory usage is tracked on a relatively 
+  high level, not including any memory allocator overhead nor any memory used for temporary
+  results calculations (e.g. memory allocated/deallocated inside AQL expressions and function 
+  calls).
 * *nodes*: _(optional)_ when the query was executed with the option `profile` set to at least *2*,
   then this value contains runtime statistics per query execution node. This field contains the
   node id (in `id`), the number of calls to this node `calls` and the number of items returned

arangod/Aql/ExecutionStats.cpp

@@ -45,9 +45,10 @@ void ExecutionStats::toVelocyPack(VPackBuilder& builder, bool reportFullCount) c
     // fullCount is optional
     builder.add("fullCount", VPackValue(fullCount > count ? fullCount : count));
   }
-  // builder.add("count", VPackValue(count));
   builder.add("executionTime", VPackValue(executionTime));
   
+  builder.add("peakMemoryUsage", VPackValue(peakMemoryUsage));
+
   if (!nodes.empty()) {
     builder.add("nodes", VPackValue(VPackValueType::Array));
     for (std::pair<size_t const, ExecutionStats::Node> const& pair : nodes) {
@@ -80,6 +81,7 @@ void ExecutionStats::add(ExecutionStats const& summand) {
     fullCount += summand.fullCount;
   }
   count += summand.count;
+  peakMemoryUsage = std::max(summand.peakMemoryUsage, peakMemoryUsage);
   // intentionally no modification of executionTime
 
   for (auto const& pair : summand.nodes) {
@@ -99,7 +101,8 @@ ExecutionStats::ExecutionStats()
       requests(0),
       fullCount(0),
       count(0),
-      executionTime(0.0) {}
+      executionTime(0.0),
+      peakMemoryUsage(0) {}
 
 ExecutionStats::ExecutionStats(VPackSlice const& slice) : ExecutionStats() {
   if (!slice.isObject()) {

arangod/Aql/ExecutionStats.h

@@ -65,6 +65,9 @@ struct ExecutionStats {
   /// @brief sets query execution time from the outside
   void setExecutionTime(double value) { executionTime = value; }
   
+  /// @brief sets the peak memory usage from the outside
+  void setPeakMemoryUsage(size_t value) { peakMemoryUsage = value; }
+
   /// @brief sumarize two sets of ExecutionStats
   void add(ExecutionStats const& summand);
 
@@ -78,6 +81,7 @@ struct ExecutionStats {
     fullCount = 0;
     count = 0;
     executionTime = 0.0;
+    peakMemoryUsage = 0;
   }
 
   /// @brief number of successfully executed write operations
@@ -108,6 +112,9 @@ struct ExecutionStats {
   /// the outside
   double executionTime;
 
+  /// @brief peak memory usage of the query
+  size_t peakMemoryUsage;
+
   ///  @brief statistics per ExecutionNodes
   std::map<size_t, ExecutionStats::Node> nodes;
 };

arangod/Aql/Query.cpp

@@ -279,6 +279,7 @@ void Query::kill() { _killed = true; }
 
 void Query::setExecutionTime() {
   if (_engine != nullptr) {
+    _engine->_stats.setPeakMemoryUsage(_resourceMonitor.currentResources.peakMemoryUsage);
     _engine->_stats.setExecutionTime(TRI_microtime() - _startTime);
   }
 }
@@ -976,7 +977,7 @@ ExecutionState Query::finalize(QueryResult& result) {
         << "Query::finalize: before cleanupPlanAndEngine"
         << " this: " << (uintptr_t)this;
 
-    _engine->_stats.setExecutionTime(runTime());
+    setExecutionTime();
     enterState(QueryExecutionState::ValueType::FINALIZATION);
 
     result.extra = std::make_shared<VPackBuilder>();
@@ -1009,7 +1010,7 @@ ExecutionState Query::finalize(QueryResult& result) {
   // patch executionTime stats value in place
   // we do this because "executionTime" should include the whole span of the
   // execution and we have to set it at the very end
-  double const rt = runTime(now);
+  double const rt = now - _startTime;
   basics::VelocyPackHelper::patchDouble(result.extra->slice().get("stats").get("executionTime"),
                                         rt);
 
@@ -1244,7 +1245,7 @@ void Query::exitContext() {
 /// @brief returns statistics for current query.
 void Query::getStats(VPackBuilder& builder) {
   if (_engine != nullptr) {
-    _engine->_stats.setExecutionTime(TRI_microtime() - _startTime);
+    setExecutionTime();
     _engine->_stats.toVelocyPack(builder, _queryOptions.fullCount);
   } else {
     ExecutionStats::toVelocyPackStatic(builder);

arangod/Aql/Query.h

@@ -150,12 +150,6 @@ class Query {
   /// @brief return the start timestamp of the query
   double startTime() const { return _startTime; }
 
-  /// @brief return the current runtime of the query
-  double runTime(double now) const { return now - _startTime; }
-
-  /// @brief return the current runtime of the query
-  double runTime() const { return runTime(TRI_microtime()); }
-
   /// @brief the part of the query
   inline QueryPart part() const { return _part; }
 

arangod/Aql/ResourceUsage.h

@@ -27,16 +27,26 @@
 #include "Basics/Common.h"
 #include "Basics/Exceptions.h"
 
+#include <algorithm>
+
 namespace arangodb {
 namespace aql {
 
 struct ResourceUsage {
-  constexpr ResourceUsage() : memoryUsage(0) {}
+  constexpr ResourceUsage() 
-  explicit ResourceUsage(size_t memoryUsage) : memoryUsage(memoryUsage) {}
+      : memoryUsage(0), 
+        peakMemoryUsage(0) {}
+  ResourceUsage(ResourceUsage const& other) noexcept
+      : memoryUsage(other.memoryUsage),
+        peakMemoryUsage(other.peakMemoryUsage) {}
 
-  void clear() { memoryUsage = 0; }
+  void clear() { 
+    memoryUsage = 0; 
+    peakMemoryUsage = 0;
+  }
 
   size_t memoryUsage;
+  size_t peakMemoryUsage;
 };
 
 struct ResourceMonitor {
@@ -47,12 +57,16 @@ struct ResourceMonitor {
   void setMemoryLimit(size_t value) { maxResources.memoryUsage = value; }
 
   inline void increaseMemoryUsage(size_t value) {
+    currentResources.memoryUsage += value;
+
     if (maxResources.memoryUsage > 0 &&
-        currentResources.memoryUsage + value > maxResources.memoryUsage) {
+        ADB_UNLIKELY(currentResources.memoryUsage > maxResources.memoryUsage)) {
+      currentResources.memoryUsage -= value;
       THROW_ARANGO_EXCEPTION_MESSAGE(
           TRI_ERROR_RESOURCE_LIMIT, "query would use more memory than allowed");
     }
-    currentResources.memoryUsage += value;
+
+    currentResources.peakMemoryUsage = std::max(currentResources.memoryUsage, currentResources.peakMemoryUsage);
   }
 
   inline void decreaseMemoryUsage(size_t value) noexcept {

js/server/modules/@arangodb/aql-helper.js

@@ -430,6 +430,7 @@ function getQueryMultiplePlansAndExecutions (query, bindVars, testObject, debug)
     delete results[i].stats.filtered;
     delete results[i].stats.executionTime;
     delete results[i].stats.httpRequests;
+    delete results[i].stats.peakMemoryUsage;
     delete results[i].stats.fullCount;
 
     if (debug) {

js/server/modules/@arangodb/aql-profiler-test-helper.js

@@ -250,6 +250,7 @@ function assertIsProfileStatsObject (stats, {level}) {
     'scannedIndex',
     'filtered',
     'httpRequests',
+    'peakMemoryUsage',
     'executionTime',
   ];
 
@@ -266,6 +267,7 @@ function assertIsProfileStatsObject (stats, {level}) {
   expect(stats.scannedIndex).to.be.a('number');
   expect(stats.filtered).to.be.a('number');
   expect(stats.httpRequests).to.be.a('number');
+  expect(stats.peakMemoryUsage).to.be.a('number');
   expect(stats.executionTime).to.be.a('number');
 }
 

tests/js/server/aql/aql-modify-cluster.js

@@ -44,6 +44,7 @@ var sanitizeStats = function (stats) {
   delete stats.executionTime;
   delete stats.httpRequests;
   delete stats.fullCount;
+  delete stats.peakMemoryUsage;
   return stats;
 };
 

tests/js/server/aql/aql-modify-noncluster.js

@@ -48,6 +48,7 @@ var sanitizeStats = function (stats) {
   delete stats.executionTime;
   delete stats.httpRequests;
   delete stats.fullCount;
+  delete stats.peakMemoryUsage;
   return stats;
 };
 

tests/js/server/aql/aql-modify-subqueries.js

@@ -56,6 +56,7 @@ var sanitizeStats = function (stats) {
   delete stats.executionTime;
   delete stats.httpRequests;
   delete stats.fullCount;
+  delete stats.peakMemoryUsage;
   return stats;
 };