Merge pull request #1414 from Altinity/frontport/antalya-26.1/rendezvous_hashing

26.1 Antalya port - improvements for cluster requests

Author: zvonand

docs/en/sql-reference/statements/system.md

@@ -220,6 +220,12 @@ Normally shuts down ClickHouse (like `service clickhouse-server stop` / `kill {$
 
 Aborts ClickHouse process (like `kill -9 {$ pid_clickhouse-server}`)
 
+## SYSTEM PRESHUTDOWN {#preshutdown}
+
+<CloudNotSupportedBadge/>
+
+Prepare node for graceful shutdown. Unregister in autodiscovered clusters, stop accepting distributed requests to object storages (s3Cluster, icebergCluster, etc.).
+
 ## SYSTEM INSTRUMENT {#instrument}
 
 Manages instrumentation points using LLVM's XRay feature which is available when ClickHouse is built using `ENABLE_XRAY=1`.

programs/server/Server.cpp

@@ -2570,6 +2570,8 @@ try
 
     }
 
+    global_context->startSwarmMode();
+
     {
         std::lock_guard lock(servers_lock);
         /// We should start interserver communications before (and more important shutdown after) tables.
@@ -3042,6 +3044,8 @@ try
 
             is_cancelled = true;
 
+            global_context->stopSwarmMode();
+
             LOG_DEBUG(log, "Waiting for current connections to close.");
 
             size_t current_connections = 0;

src/Access/Common/AccessType.h

@@ -353,6 +353,7 @@ enum class AccessType : uint8_t
     M(SYSTEM_TTL_MERGES, "SYSTEM STOP TTL MERGES, SYSTEM START TTL MERGES, STOP TTL MERGES, START TTL MERGES", TABLE, SYSTEM) \
     M(SYSTEM_FETCHES, "SYSTEM STOP FETCHES, SYSTEM START FETCHES, STOP FETCHES, START FETCHES", TABLE, SYSTEM) \
     M(SYSTEM_MOVES, "SYSTEM STOP MOVES, SYSTEM START MOVES, STOP MOVES, START MOVES", TABLE, SYSTEM) \
+    M(SYSTEM_SWARM, "SYSTEM STOP SWARM MODE, SYSTEM START SWARM MODE, STOP SWARM MODE, START SWARM MODE", GLOBAL, SYSTEM) \
     M(SYSTEM_PULLING_REPLICATION_LOG, "SYSTEM STOP PULLING REPLICATION LOG, SYSTEM START PULLING REPLICATION LOG", TABLE, SYSTEM) \
     M(SYSTEM_CLEANUP, "SYSTEM STOP CLEANUP, SYSTEM START CLEANUP", TABLE, SYSTEM) \
     M(SYSTEM_VIEWS, "SYSTEM REFRESH VIEW, SYSTEM START VIEWS, SYSTEM STOP VIEWS, SYSTEM START VIEW, SYSTEM STOP VIEW, SYSTEM CANCEL VIEW, REFRESH VIEW, START VIEWS, STOP VIEWS, START VIEW, STOP VIEW, CANCEL VIEW", VIEW, SYSTEM) \

src/Client/MultiplexedConnections.cpp

@@ -232,7 +232,7 @@ void MultiplexedConnections::sendIgnoredPartUUIDs(const std::vector<UUID> & uuid
 void MultiplexedConnections::sendClusterFunctionReadTaskResponse(const ClusterFunctionReadTaskResponse & response)
 {
     std::lock_guard lock(cancel_mutex);
-    if (cancelled)
+    if (cancelled || !current_connection || !current_connection->isConnected())
         return;
     current_connection->sendClusterFunctionReadTaskResponse(response);
 }
@@ -241,7 +241,7 @@ void MultiplexedConnections::sendClusterFunctionReadTaskResponse(const ClusterFu
 void MultiplexedConnections::sendMergeTreeReadTaskResponse(const ParallelReadResponse & response)
 {
     std::lock_guard lock(cancel_mutex);
-    if (cancelled)
+    if (cancelled || !current_connection || !current_connection->isConnected())
         return;
     current_connection->sendMergeTreeReadTaskResponse(response);
 }
@@ -527,9 +527,12 @@ MultiplexedConnections::ReplicaState & MultiplexedConnections::getReplicaForRead
 
 void MultiplexedConnections::invalidateReplica(ReplicaState & state)
 {
+    Connection * old_connection = state.connection;
     state.connection = nullptr;
     state.pool_entry = IConnectionPool::Entry();
     --active_connection_count;
+    if (current_connection == old_connection)
+        current_connection = nullptr;
 }
 
 void MultiplexedConnections::setAsyncCallback(AsyncCallback async_callback)

src/Common/CurrentMetrics.cpp

@@ -461,6 +461,7 @@
     M(StartupScriptsExecutionState, "State of startup scripts execution: 0 = not finished, 1 = success, 2 = failure.") \
     \
     M(IsServerShuttingDown, "Indicates if the server is shutting down: 0 = no, 1 = yes") \
+    M(IsSwarmModeEnabled, "Indicates if the swarm mode enabled or not: 0 = disabled, 1 = enabled") \
     \
     M(StatelessWorkerThreads, "Number of threads in the stateless worker thread pool.") \
     M(StatelessWorkerThreadsActive, "Number of threads in the stateless worker thread pool running a task.") \

src/Common/ProfileEvents.cpp

@@ -1298,8 +1298,13 @@ The server successfully detected this situation and will download merged part fr
     M(RuntimeFilterRowsChecked, "Number of rows checked by JOIN Runtime Filters", ValueType::Number) \
     M(RuntimeFilterRowsPassed, "Number of rows that passed (not filtered out by) JOIN Runtime Filters", ValueType::Number) \
     M(RuntimeFilterRowsSkipped, "Number of rows in blocks that were skipped by JOIN Runtime Filters", ValueType::Number) \
+    \
+    M(ObjectStorageClusterSentToMatchedReplica, "Number of tasks in ObjectStorageCluster request sent to matched replica.", ValueType::Number) \
+    M(ObjectStorageClusterSentToNonMatchedReplica, "Number of tasks in ObjectStorageCluster request sent to non-matched replica.", ValueType::Number) \
+    M(ObjectStorageClusterProcessedTasks, "Number of processed tasks in ObjectStorageCluster request.", ValueType::Number) \
+    M(ObjectStorageClusterWaitingMicroseconds, "Time of waiting for tasks in ObjectStorageCluster request.", ValueType::Microseconds) \
     M(ObjectStorageListObjectsCacheHits, "Number of times object storage list objects operation hit the cache.", ValueType::Number) \
-    M(ObjectStorageListObjectsCacheMisses, "Number of times object storage list objects operation miss the cache.", ValueType::Number)                                                                                                                                                                                                                                  \
+    M(ObjectStorageListObjectsCacheMisses, "Number of times object storage list objects operation miss the cache.", ValueType::Number) \
     M(ObjectStorageListObjectsCacheExactMatchHits, "Number of times object storage list objects operation hit the cache with an exact match.", ValueType::Number) \
     M(ObjectStorageListObjectsCachePrefixMatchHits, "Number of times object storage list objects operation miss the cache using prefix matching.", ValueType::Number) \
     M(ParquetMetaDataCacheHits, "Number of times the read from filesystem cache hit the cache.", ValueType::Number) \

src/Core/Protocol.h

@@ -96,8 +96,10 @@ namespace Protocol
             MergeTreeReadTaskRequest = 16,  /// Request from a MergeTree replica to a coordinator
             TimezoneUpdate = 17,            /// Receive server's (session-wide) default timezone
             SSHChallenge = 18,              /// Return challenge for SSH signature signing
+
             MAX = SSHChallenge,
 
+            ConnectionLost = 255,           /// Exception that occurred on the client side.
         };
 
         /// NOTE: If the type of packet argument would be Enum, the comparison packet >= 0 && packet < 10

src/Core/Range.cpp

@@ -2,13 +2,21 @@
 #include <Core/Range.h>
 #include <IO/Operators.h>
 #include <IO/WriteBufferFromString.h>
+#include <IO/ReadBufferFromString.h>
 #include <Common/FieldVisitorToString.h>
 #include <Common/FieldAccurateComparison.h>
+#include <Common/Base64.h>
 
 
 namespace DB
 {
 
+namespace ErrorCodes
+{
+    extern const int INCORRECT_DATA;
+};
+
+
 FieldRef::FieldRef(ColumnsWithTypeAndName * columns_, size_t row_idx_, size_t column_idx_)
     : Field((*(*columns_)[column_idx_].column)[row_idx_]), columns(columns_), row_idx(row_idx_), column_idx(column_idx_)
 {
@@ -151,6 +159,13 @@ bool Range::isInfinite() const
     return left.isNegativeInfinity() && right.isPositiveInfinity();
 }
 
+/// [x, x]
+bool Range::isPoint() const
+{
+    return fullBounded() && left_included && right_included && equals(left, right)
+        && !left.isNegativeInfinity() && !left.isPositiveInfinity();
+}
+
 bool Range::intersectsRange(const Range & r) const
 {
     /// r to the left of me.
@@ -276,6 +291,32 @@ bool Range::nearByWith(const Range & r) const
     return false;
 }
 
+String Range::serialize(bool base64) const
+{
+    WriteBufferFromOwnString str;
+
+    str << left_included << right_included;
+    writeFieldBinary(left, str);
+    writeFieldBinary(right, str);
+
+    if (base64)
+        return base64Encode(str.str());
+    else
+        return str.str();
+}
+
+void Range::deserialize(const String & range, bool base64)
+{
+    if (range.empty())
+        throw Exception(ErrorCodes::INCORRECT_DATA, "Empty range dump");
+
+    ReadBufferFromOwnString str(base64 ? base64Decode(range) : range);
+
+    str >> left_included >> right_included;
+    left = readFieldBinary(str);
+    right = readFieldBinary(str);
+}
+
 Range intersect(const Range & a, const Range & b)
 {
     Range res = Range::createWholeUniverse();

src/Core/Range.h

@@ -94,6 +94,8 @@ struct Range
 
     bool isBlank() const;
 
+    bool isPoint() const;
+
     bool intersectsRange(const Range & r) const;
 
     bool containsRange(const Range & r) const;
@@ -114,6 +116,9 @@ struct Range
     bool nearByWith(const Range & r) const;
 
     String toString() const;
+
+    String serialize(bool base64 = false) const;
+    void deserialize(const String & range, bool base64 = false);
 };
 
 Range intersect(const Range & a, const Range & b);

src/Core/Settings.cpp

@@ -7544,6 +7544,19 @@ Default number of tasks for parallel reading in distributed query. Tasks are spr
     DECLARE(Bool, distributed_plan_optimize_exchanges, true, R"(
 Removes unnecessary exchanges in distributed query plan. Disable it for debugging.
 )", 0) \
+    DECLARE(UInt64, lock_object_storage_task_distribution_ms, 500, R"(
+In object storage distribution queries do not distribute tasks on non-prefetched nodes until prefetched node is active.
+Determines how long the free executor node (one that finished processing all of it assigned tasks) should wait before "stealing" tasks from queue of currently busy executor nodes.
+
+Possible values:
+
+- 0  - steal tasks immediately after freeing up.
+- >0 - wait for specified period of time before stealing tasks.
+
+Having this `>0` helps with cache reuse and might improve overall query time.
+Because busy node might have warmed-up caches for this specific task, while free node needs to fetch lots of data from S3.
+Which might take longer than just waiting for the busy node and generate extra traffic.
+)", EXPERIMENTAL) \
     DECLARE(String, distributed_plan_force_exchange_kind, "", R"(
 Force specified kind of Exchange operators between distributed query stages.
 
@@ -7591,6 +7604,9 @@ If the number of set bits in a runtime bloom filter exceeds this ratio the filte
 )", EXPERIMENTAL) \
     DECLARE(Bool, rewrite_in_to_join, false, R"(
 Rewrite expressions like 'x IN subquery' to JOIN. This might be useful for optimizing the whole query with join reordering.
+)", EXPERIMENTAL) \
+    DECLARE(Bool, allow_experimental_iceberg_read_optimization, true, R"(
+Allow Iceberg read optimization based on Iceberg metadata.
 )", EXPERIMENTAL) \
     \
     /** Experimental timeSeries* aggregate functions. */ \
@@ -7743,7 +7759,8 @@ Allow experimental database engine DataLakeCatalog with catalog_type = 'paimon_r
     MAKE_OBSOLETE(M, Bool, use_json_alias_for_old_object_type, false) \
     MAKE_OBSOLETE(M, Bool, describe_extend_object_types, false) \
     MAKE_OBSOLETE(M, Bool, allow_experimental_object_type, false) \
-    MAKE_OBSOLETE(M, BoolAuto, insert_select_deduplicate, Field{"auto"})
+    MAKE_OBSOLETE(M, BoolAuto, insert_select_deduplicate, Field{"auto"}) \
+    MAKE_OBSOLETE(M, Bool, allow_retries_in_cluster_requests, false) \
     /** The section above is for obsolete settings. Do not add anything there. */
 #endif /// __CLION_IDE__