[Enhancement] fix connector mem scan limit adjustment when no chunk s…

…ource (#53112) Signed-off-by: yanz <[email protected]> (cherry picked from commit 14ccfcc) # Conflicts: # fe/fe-core/src/main/java/com/starrocks/planner/OlapScanNode.java # fe/fe-core/src/main/java/com/starrocks/planner/ScanNode.java # fe/fe-core/src/main/java/com/starrocks/planner/SchemaScanNode.java
StarRocks · Nov 25, 2024 · 906f250 · 906f250
1 parent 7b672b9
commit 906f250
Show file tree

Hide file tree

Showing 12 changed files with 65 additions and 44 deletions.
diff --git a/be/src/connector/connector.h b/be/src/connector/connector.h
@@ -125,9 +125,10 @@ using DataSourcePtr = std::unique_ptr<DataSource>;
 
 class DataSourceProvider {
 public:
-    static constexpr int64_t MIN_DATA_SOURCE_MEM_BYTES = 16 * 1024 * 1024;  // 16MB
-    static constexpr int64_t MAX_DATA_SOURCE_MEM_BYTES = 256 * 1024 * 1024; // 256MB
-    static constexpr int64_t PER_FIELD_MEM_BYTES = 1 * 1024 * 1024;         // 1MB
+    static constexpr int64_t MIN_DATA_SOURCE_MEM_BYTES = 16 * 1024 * 1024;     // 16MB
+    static constexpr int64_t DEFAULT_DATA_SOURCE_MEM_BYTES = 64 * 1024 * 1024; // 64MB
+    static constexpr int64_t MAX_DATA_SOURCE_MEM_BYTES = 256 * 1024 * 1024;    // 256MB
+    static constexpr int64_t PER_FIELD_MEM_BYTES = 1 * 1024 * 1024;            // 1MB
 
     virtual ~DataSourceProvider() = default;
 

diff --git a/be/src/exec/pipeline/fragment_executor.cpp b/be/src/exec/pipeline/fragment_executor.cpp
@@ -226,12 +226,12 @@ Status FragmentExecutor::_prepare_runtime_state(ExecEnv* exec_env, const Unified
         }
     }
 
-    int scan_node_number = 1;
-    if (query_globals.__isset.scan_node_number) {
-        scan_node_number = query_globals.scan_node_number;
+    int connector_scan_node_number = 1;
+    if (query_globals.__isset.connector_scan_node_number) {
+        connector_scan_node_number = query_globals.connector_scan_node_number;
     }
     _query_ctx->init_mem_tracker(option_query_mem_limit, parent_mem_tracker, big_query_mem_limit, spill_mem_limit_ratio,
-                                 wg.get(), runtime_state, scan_node_number);
+                                 wg.get(), runtime_state, connector_scan_node_number);
 
     auto query_mem_tracker = _query_ctx->mem_tracker();
     SCOPED_THREAD_LOCAL_MEM_TRACKER_SETTER(query_mem_tracker.get());

diff --git a/be/src/exec/pipeline/query_context.cpp b/be/src/exec/pipeline/query_context.cpp
@@ -113,7 +113,7 @@ void QueryContext::cancel(const Status& status) {
 
 void QueryContext::init_mem_tracker(int64_t query_mem_limit, MemTracker* parent, int64_t big_query_mem_limit,
                                     std::optional<double> spill_mem_reserve_ratio, workgroup::WorkGroup* wg,
-                                    RuntimeState* runtime_state, int scan_node_number) {
+                                    RuntimeState* runtime_state, int connector_scan_node_number) {
     std::call_once(_init_mem_tracker_once, [=]() {
         _profile = std::make_shared<RuntimeProfile>("Query" + print_id(_query_id));
         auto* mem_tracker_counter =
@@ -145,7 +145,7 @@ void QueryContext::init_mem_tracker(int64_t query_mem_limit, MemTracker* parent,
             _static_query_mem_limit = std::min(big_query_mem_limit, _static_query_mem_limit);
         }
         _connector_scan_operator_mem_share_arbitrator = _object_pool.add(
-                new ConnectorScanOperatorMemShareArbitrator(_static_query_mem_limit, scan_node_number));
+                new ConnectorScanOperatorMemShareArbitrator(_static_query_mem_limit, connector_scan_node_number));
 
         {
             MemTracker* connector_scan_parent = GlobalEnv::GetInstance()->connector_scan_pool_mem_tracker();

diff --git a/be/src/exec/pipeline/query_context.h b/be/src/exec/pipeline/query_context.h
@@ -164,7 +164,7 @@ class QueryContext : public std::enable_shared_from_this<QueryContext> {
     /// that there is a big query memory limit of this resource group.
     void init_mem_tracker(int64_t query_mem_limit, MemTracker* parent, int64_t big_query_mem_limit = -1,
                           std::optional<double> spill_mem_limit = std::nullopt, workgroup::WorkGroup* wg = nullptr,
-                          RuntimeState* state = nullptr, int scan_node_number = 1);
+                          RuntimeState* state = nullptr, int connector_scan_node_number = 1);
     std::shared_ptr<MemTracker> mem_tracker() { return _mem_tracker; }
     MemTracker* connector_scan_mem_tracker() { return _connector_scan_mem_tracker.get(); }
 

diff --git a/be/src/exec/pipeline/scan/connector_scan_operator.cpp b/be/src/exec/pipeline/scan/connector_scan_operator.cpp
@@ -17,18 +17,18 @@
 #include "exec/connector_scan_node.h"
 #include "exec/pipeline/pipeline_driver.h"
 #include "exec/pipeline/scan/balanced_chunk_buffer.h"
-#include "exec/workgroup/work_group.h"
 #include "runtime/exec_env.h"
 #include "runtime/runtime_state.h"
 
 namespace starrocks::pipeline {
 
 // ==================== ConnectorScanOperatorFactory ====================
 ConnectorScanOperatorMemShareArbitrator::ConnectorScanOperatorMemShareArbitrator(int64_t query_mem_limit,
-                                                                                 int scan_node_number)
+                                                                                 int connector_scan_node_number)
         : query_mem_limit(query_mem_limit),
           scan_mem_limit(query_mem_limit),
-          total_chunk_source_mem_bytes(scan_node_number * connector::DataSourceProvider::MAX_DATA_SOURCE_MEM_BYTES) {}
+          total_chunk_source_mem_bytes(connector_scan_node_number *
+                                       connector::DataSourceProvider::DEFAULT_DATA_SOURCE_MEM_BYTES) {}
 
 int64_t ConnectorScanOperatorMemShareArbitrator::update_chunk_source_mem_bytes(int64_t old_value, int64_t new_value) {
     int64_t diff = new_value - old_value;
@@ -54,9 +54,7 @@ class ConnectorScanOperatorIOTasksMemLimiter {
     int64_t chunk_source_mem_bytes_update_count = 0;
     int64_t arb_chunk_source_mem_bytes = 0;
     mutable int64_t debug_output_timestamp = 0;
-
     std::atomic<int64_t> open_scan_operator_count = 0;
-    std::atomic<int64_t> active_scan_operator_count = 0;
 
 public:
     ConnectorScanOperatorIOTasksMemLimiter(int64_t dop, bool shared_scan) : dop(dop), shared_scan(shared_scan) {}
@@ -127,9 +125,6 @@ class ConnectorScanOperatorIOTasksMemLimiter {
     int64_t update_open_scan_operator_count(int delta) {
         return open_scan_operator_count.fetch_add(delta, std::memory_order_seq_cst);
     }
-    int64_t update_active_scan_operator_count(int delta) {
-        return active_scan_operator_count.fetch_add(delta, std::memory_order_seq_cst);
-    }
 };
 
 ConnectorScanOperatorFactory::ConnectorScanOperatorFactory(int32_t id, ScanNode* scan_node, RuntimeState* state,
@@ -218,9 +213,6 @@ struct ConnectorScanOperatorAdaptiveProcessor {
     int try_add_io_tasks_fail_count = 0;
     int check_slow_io = 0;
     int32_t slow_io_latency_ms = config::connector_io_tasks_adjust_interval_ms;
-
-    // ------------------------
-    bool started_running = false;
 };
 
 // ==================== ConnectorScanOperator ====================
@@ -264,43 +256,36 @@ Status ConnectorScanOperator::do_prepare(RuntimeState* state) {
     _unique_metrics->add_info_string("AdaptiveIOTasks", _enable_adaptive_io_tasks ? "True" : "False");
     _adaptive_processor = state->obj_pool()->add(new ConnectorScanOperatorAdaptiveProcessor());
     _adaptive_processor->op_start_time = GetCurrentTimeMicros();
-    _adaptive_processor->started_running = false;
     if (options.__isset.connector_io_tasks_slow_io_latency_ms) {
         _adaptive_processor->slow_io_latency_ms = options.connector_io_tasks_slow_io_latency_ms;
     }
 
+    // As the first running scan operator, it will update the scan mem limit
     {
         auto* factory = down_cast<ConnectorScanOperatorFactory*>(_factory);
         ConnectorScanOperatorIOTasksMemLimiter* L = factory->_io_tasks_mem_limiter;
-        L->update_open_scan_operator_count(1);
+        int64_t c = L->update_open_scan_operator_count(1);
+        if (c == 0) {
+            _adjust_scan_mem_limit(connector::DataSourceProvider::DEFAULT_DATA_SOURCE_MEM_BYTES,
+                                   L->get_arb_chunk_source_mem_bytes());
+        }
     }
     return Status::OK();
 }
 
 void ConnectorScanOperator::do_close(RuntimeState* state) {
+    // As the last closing scan operator, it will update the scan mem limit.
     auto* factory = down_cast<ConnectorScanOperatorFactory*>(_factory);
     ConnectorScanOperatorIOTasksMemLimiter* L = factory->_io_tasks_mem_limiter;
     int64_t c = L->update_open_scan_operator_count(-1);
     if (c == 1) {
-        if (L->update_active_scan_operator_count(0) > 0) {
-            _adjust_scan_mem_limit(L->get_arb_chunk_source_mem_bytes(), 0);
-        }
+        _adjust_scan_mem_limit(L->get_arb_chunk_source_mem_bytes(), 0);
     }
 }
 
 ChunkSourcePtr ConnectorScanOperator::create_chunk_source(MorselPtr morsel, int32_t chunk_source_index) {
     auto* scan_node = down_cast<ConnectorScanNode*>(_scan_node);
     auto* factory = down_cast<ConnectorScanOperatorFactory*>(_factory);
-    ConnectorScanOperatorIOTasksMemLimiter* L = factory->_io_tasks_mem_limiter;
-
-    if (_adaptive_processor->started_running == false) {
-        _adaptive_processor->started_running = true;
-        int64_t c = L->update_active_scan_operator_count(1);
-        if (c == 0) {
-            _adjust_scan_mem_limit(connector::DataSourceProvider::MAX_DATA_SOURCE_MEM_BYTES,
-                                   L->get_arb_chunk_source_mem_bytes());
-        }
-    }
 
     // Only use one chunk source profile, so we can see metrics on scan operator level.
     // Since there is adaptive io tasks feature, chunk sources will be used unevenly,

diff --git a/be/src/exec/pipeline/scan/connector_scan_operator.h b/be/src/exec/pipeline/scan/connector_scan_operator.h
@@ -35,7 +35,7 @@ struct ConnectorScanOperatorMemShareArbitrator {
     int64_t scan_mem_limit = 0;
     std::atomic<int64_t> total_chunk_source_mem_bytes = 0;
 
-    ConnectorScanOperatorMemShareArbitrator(int64_t query_mem_limit, int scan_node_number);
+    ConnectorScanOperatorMemShareArbitrator(int64_t query_mem_limit, int connector_scan_node_number);
 
     int64_t set_scan_mem_ratio(double mem_ratio) {
         scan_mem_limit = std::max<int64_t>(1, query_mem_limit * mem_ratio);

diff --git a/fe/fe-core/src/main/java/com/starrocks/planner/MetaScanNode.java b/fe/fe-core/src/main/java/com/starrocks/planner/MetaScanNode.java
@@ -117,7 +117,8 @@ public void computeRangeLocations() {
                 boolean tabletIsNull = true;
                 for (Replica replica : allQueryableReplicas) {
                     ComputeNode node =
-                            GlobalStateMgr.getCurrentState().getNodeMgr().getClusterInfo().getBackendOrComputeNode(replica.getBackendId());
+                            GlobalStateMgr.getCurrentState().getNodeMgr().getClusterInfo()
+                                    .getBackendOrComputeNode(replica.getBackendId());
                     if (node == null) {
                         LOG.debug("replica {} not exists", replica.getBackendId());
                         continue;
@@ -183,4 +184,9 @@ protected String getNodeExplainString(String prefix, TExplainLevel detailLevel)
     public boolean canUseRuntimeAdaptiveDop() {
         return true;
     }
+
+    @Override
+    public boolean isRunningAsConnectorOperator() {
+        return false;
+    }
 }
diff --git a/fe/fe-core/src/main/java/com/starrocks/planner/OlapScanNode.java b/fe/fe-core/src/main/java/com/starrocks/planner/OlapScanNode.java
@@ -504,7 +504,6 @@ public List<TScanRangeLocations> updateScanRangeLocations(List<TScanRangeLocatio
         return newLocations;
     }
 
-
     private void checkSomeAliveComputeNode() throws ErrorReportException {
         // Note that it's theoretically possible that there were some living CN earlier in this query's execution, and then
         // they all died, but in that case, the problem this will be surfaced later anyway.
@@ -764,7 +763,8 @@ public void checkIfScanRangeNumSafe(long scanRangeSize) {
             long freeMemory = runtime.freeMemory();
             if (totalScanRangeBytes > freeMemory / 2) {
                 LOG.warn(
-                        "Try to allocate too many scan ranges for table {}, which may cause FE OOM, Partition Num:{}, tablet Num:{}, Scan Range Total Bytes:{}",
+                        "Try to allocate too many scan ranges for table {}, which may cause FE OOM, Partition Num:{}, tablet " +
+                                "Num:{}, Scan Range Total Bytes:{}",
                         olapTable.getName(), totalPartitionNum, totalTabletsNum, totalScanRangeBytes);
             }
         }
@@ -912,11 +912,14 @@ protected String getNodeExplainString(String prefix, TExplainLevel detailLevel)
         return output.toString();
     }
 
+<<<<<<< HEAD
     @Override
     public int getNumInstances() {
         return result.size();
     }
 
+=======
+>>>>>>> 14ccfcc047 ([Enhancement] fix connector mem scan limit adjustment when no chunk source (#53112))
     private void assignOrderByHints(List<String> keyColumnNames) {
         // assign order by hint
         for (RuntimeFilterDescription probeRuntimeFilter : probeRuntimeFilters) {
@@ -1450,4 +1453,9 @@ public void clearScanNodeForThriftBuild() {
         bucketColumns.clear();
         rowStoreKeyLiterals = Lists.newArrayList();
     }
+
+    @Override
+    public boolean isRunningAsConnectorOperator() {
+        return false;
+    }
 }
diff --git a/fe/fe-core/src/main/java/com/starrocks/planner/ScanNode.java b/fe/fe-core/src/main/java/com/starrocks/planner/ScanNode.java
@@ -158,4 +158,18 @@ protected List<TColumnAccessPath> columnAccessPathToThrift() {
     protected boolean supportTopNRuntimeFilter() {
         return false;
     }
+<<<<<<< HEAD
+=======
+
+    @Override
+    public boolean needCollectExecStats() {
+        return true;
+    }
+
+    // We use this flag to know how many connector scan nodes at BE side, and connector framework
+    // will use this number to fair share memory usage between those scan nodes.
+    public boolean isRunningAsConnectorOperator() {
+        return true;
+    }
+>>>>>>> 14ccfcc047 ([Enhancement] fix connector mem scan limit adjustment when no chunk source (#53112))
 }
diff --git a/fe/fe-core/src/main/java/com/starrocks/planner/SchemaScanNode.java b/fe/fe-core/src/main/java/com/starrocks/planner/SchemaScanNode.java
@@ -58,7 +58,6 @@
 import org.apache.logging.log4j.LogManager;
 import org.apache.logging.log4j.Logger;
 
-import java.util.Collections;
 import java.util.List;
 import java.util.stream.Collectors;
 
@@ -372,11 +371,14 @@ public List<TScanRangeLocations> getScanRangeLocations(long maxScanRangeLength)
         return beScanRanges;
     }
 
+<<<<<<< HEAD
     @Override
     public int getNumInstances() {
         return beScanRanges == null ? 1 : beScanRanges.size();
     }
 
+=======
+>>>>>>> 14ccfcc047 ([Enhancement] fix connector mem scan limit adjustment when no chunk source (#53112))
     @Override
     public boolean canUseRuntimeAdaptiveDop() {
         return true;
@@ -389,4 +391,9 @@ public String getCatalogName() {
     public void setCatalogName(String catalogName) {
         this.catalogName = catalogName;
     }
+
+    @Override
+    public boolean isRunningAsConnectorOperator() {
+        return false;
+    }
 }
diff --git a/fe/fe-core/src/main/java/com/starrocks/qe/scheduler/dag/JobSpec.java b/fe/fe-core/src/main/java/com/starrocks/qe/scheduler/dag/JobSpec.java
@@ -110,7 +110,7 @@ public static JobSpec fromQuerySpec(ConnectContext context,
             if (context.getLastQueryId() != null) {
                 queryGlobals.setLast_query_id(context.getLastQueryId().toString());
             }
-            queryGlobals.setScan_node_number(scanNodes.size());
+            queryGlobals.setConnector_scan_node_number(scanNodes.stream().filter(x -> x.isRunningAsConnectorOperator()).count());
 
             return new Builder()
                     .queryId(context.getExecutionId())
@@ -140,7 +140,7 @@ public static JobSpec fromMVMaintenanceJobSpec(ConnectContext context,
             if (context.getLastQueryId() != null) {
                 queryGlobals.setLast_query_id(context.getLastQueryId().toString());
             }
-            queryGlobals.setScan_node_number(scanNodes.size());
+            queryGlobals.setConnector_scan_node_number(scanNodes.stream().filter(x -> x.isRunningAsConnectorOperator()).count());
 
             return new Builder()
                     .queryId(context.getExecutionId())

diff --git a/gensrc/thrift/InternalService.thrift b/gensrc/thrift/InternalService.thrift
@@ -398,7 +398,7 @@ struct TQueryGlobals {
 
   31: optional i64 timestamp_us
 
-  32: optional i64 scan_node_number
+  32: optional i64 connector_scan_node_number
 }