Merge pull request #70029 from ClickHouse/backport/24.6/69075

Backport #69075 to 24.6: Remove stale moving parts without zookeeper

src/Common/FailPoint.cpp

@@ -57,6 +57,7 @@ static struct InitFiu
     PAUSEABLE_ONCE(finish_clean_quorum_failed_parts) \
     PAUSEABLE(dummy_pausable_failpoint) \
     ONCE(execute_query_calling_empty_set_result_func_on_exception) \
+    PAUSEABLE(stop_moving_part_before_swap_with_active) \
     ONCE(receive_timeout_on_table_status_response)
 
 

src/Storages/MergeTree/MergeTreePartsMover.cpp

@@ -1,4 +1,6 @@
 #include <Storages/MergeTree/MergeTreePartsMover.h>
+#include <Storages/MergeTree/MergeTreeSettings.h>
+#include <Common/FailPoint.h>
 #include <Storages/MergeTree/MergeTreeData.h>
 #include <Common/logger_useful.h>
 
@@ -14,6 +16,11 @@ namespace ErrorCodes
     extern const int DIRECTORY_ALREADY_EXISTS;
 }
 
+namespace FailPoints
+{
+    extern const char stop_moving_part_before_swap_with_active[];
+}
+
 namespace
 {
 
@@ -225,6 +232,7 @@ MergeTreePartsMover::TemporaryClonedPart MergeTreePartsMover::clonePart(const Me
     cloned_part.temporary_directory_lock = data->getTemporaryPartDirectoryHolder(part->name);
 
     MutableDataPartStoragePtr cloned_part_storage;
+    bool preserve_blobs = false;
     if (disk->supportZeroCopyReplication() && settings->allow_remote_fs_zero_copy_replication)
     {
         /// Try zero-copy replication and fallback to default copy if it's not possible
@@ -252,6 +260,7 @@ MergeTreePartsMover::TemporaryClonedPart MergeTreePartsMover::clonePart(const Me
         if (zero_copy_part)
         {
             /// FIXME for some reason we cannot just use this part, we have to re-create it through MergeTreeDataPartBuilder
+            preserve_blobs = true;
             zero_copy_part->is_temp = false;    /// Do not remove it in dtor
             cloned_part_storage = zero_copy_part->getDataPartStoragePtr();
         }
@@ -271,7 +280,17 @@ MergeTreePartsMover::TemporaryClonedPart MergeTreePartsMover::clonePart(const Me
     cloned_part.part = std::move(builder).withPartFormatFromDisk().build();
     LOG_TRACE(log, "Part {} was cloned to {}", part->name, cloned_part.part->getDataPartStorage().getFullPath());
 
-    cloned_part.part->is_temp = data->allowRemoveStaleMovingParts();
+    cloned_part.part->is_temp = false;
+    if (data->allowRemoveStaleMovingParts())
+    {
+        cloned_part.part->is_temp = true;
+        /// Setting it in case connection to zookeeper is lost while moving
+        /// Otherwise part might be stuck in the moving directory due to the KEEPER_EXCEPTION in part's destructor
+        if (preserve_blobs)
+            cloned_part.part->remove_tmp_policy = IMergeTreeDataPart::BlobsRemovalPolicyForTemporaryParts::PRESERVE_BLOBS;
+        else
+            cloned_part.part->remove_tmp_policy = IMergeTreeDataPart::BlobsRemovalPolicyForTemporaryParts::REMOVE_BLOBS;
+    }
     cloned_part.part->loadColumnsChecksumsIndexes(true, true);
     cloned_part.part->loadVersionMetadata();
     cloned_part.part->modification_time = cloned_part.part->getDataPartStorage().getLastModified().epochTime();
@@ -281,6 +300,8 @@ MergeTreePartsMover::TemporaryClonedPart MergeTreePartsMover::clonePart(const Me
 
 void MergeTreePartsMover::swapClonedPart(TemporaryClonedPart & cloned_part) const
 {
+    /// Used to get some stuck parts in the moving directory by stopping moves while pause is active
+    FailPointInjection::pauseFailPoint(FailPoints::stop_moving_part_before_swap_with_active);
     if (moves_blocker.isCancelled())
         throw Exception(ErrorCodes::ABORTED, "Cancelled moving parts.");
 

tests/integration/test_remove_stale_moving_parts/config.xml

@@ -0,0 +1,46 @@
+<clickhouse>
+    <remote_servers>
+        <cluster>
+            <shard>
+                <replica>
+                    <host>ch1</host>
+                    <port>9000</port>
+                </replica>
+            </shard>
+        </cluster>
+    </remote_servers>
+    <macros>
+        <shard>01</shard>
+    </macros>
+    <storage_configuration>
+        <disks>
+            <s3>
+                <type>s3</type>
+                <endpoint>http://minio1:9001/root/data/</endpoint>
+                <access_key_id>minio</access_key_id>
+                <secret_access_key>minio123</secret_access_key>
+            </s3>
+        </disks>
+        <policies>
+            <s3>
+                <volumes>
+                    <default>
+                        <disk>default</disk>
+                        <perform_ttl_move_on_insert>False</perform_ttl_move_on_insert>
+                    </default>
+                    <s3>
+                        <disk>s3</disk>
+                        <perform_ttl_move_on_insert>False</perform_ttl_move_on_insert>
+                    </s3>
+                </volumes>
+                <move_factor>0.0</move_factor>
+            </s3>
+        </policies>
+    </storage_configuration>
+
+    <merge_tree>
+        <allow_remote_fs_zero_copy_replication>true</allow_remote_fs_zero_copy_replication>
+        <storage_policy>s3</storage_policy>
+    </merge_tree>
+    <allow_remove_stale_moving_parts>true</allow_remove_stale_moving_parts>
+</clickhouse>

tests/integration/test_remove_stale_moving_parts/test.py

@@ -0,0 +1,117 @@
+from pathlib import Path
+import time
+import pytest
+from helpers.cluster import ClickHouseCluster
+
+cluster = ClickHouseCluster(__file__)
+ch1 = cluster.add_instance(
+    "ch1",
+    main_configs=[
+        "config.xml",
+    ],
+    macros={"replica": "node1"},
+    with_zookeeper=True,
+    with_minio=True,
+)
+
+DATABASE_NAME = "stale_moving_parts"
+
+
+@pytest.fixture(scope="module")
+def started_cluster():
+    try:
+        cluster.start()
+        yield cluster
+
+    finally:
+        cluster.shutdown()
+
+
+def q(node, query):
+    return node.query(database=DATABASE_NAME, sql=query)
+
+
+# .../disks/s3/store/
+def get_table_path(node, table):
+    return (
+        node.query(
+            sql=f"SELECT data_paths FROM system.tables WHERE table = '{table}' and database = '{DATABASE_NAME}' LIMIT 1"
+        )
+        .strip('"\n[]')
+        .split(",")[1]
+        .strip("'")
+    )
+
+
+def exec(node, cmd, path):
+    return node.exec_in_container(
+        [
+            "bash",
+            "-c",
+            f"{cmd} {path}",
+        ]
+    )
+
+
+def wait_part_is_stuck(node, table_moving_path, moving_part):
+    num_tries = 5
+    while q(node, "SELECT part_name FROM system.moves").strip() != moving_part:
+        if num_tries == 0:
+            raise Exception("Part has not started to move")
+        num_tries -= 1
+        time.sleep(1)
+    num_tries = 5
+    while exec(node, "ls", table_moving_path).strip() != moving_part:
+        if num_tries == 0:
+            raise Exception("Part is not stuck in the moving directory")
+        num_tries -= 1
+        time.sleep(1)
+
+
+def wait_zookeeper_node_to_start(zk_nodes, timeout=60):
+    start = time.time()
+    while time.time() - start < timeout:
+        try:
+            for instance in zk_nodes:
+                conn = cluster.get_kazoo_client(instance)
+                conn.get_children("/")
+            print("All instances of ZooKeeper started")
+            return
+        except Exception as ex:
+            print(("Can't connect to ZooKeeper " + str(ex)))
+            time.sleep(0.5)
+
+
+def test_remove_stale_moving_parts_without_zookeeper(started_cluster):
+    ch1.query(f"CREATE DATABASE IF NOT EXISTS {DATABASE_NAME}")
+
+    q(
+        ch1,
+        "CREATE TABLE test_remove ON CLUSTER cluster ( id UInt32 ) ENGINE ReplicatedMergeTree() ORDER BY id;",
+    )
+
+    table_moving_path = Path(get_table_path(ch1, "test_remove")) / "moving"
+
+    q(ch1, "SYSTEM ENABLE FAILPOINT stop_moving_part_before_swap_with_active")
+    q(ch1, "INSERT INTO test_remove SELECT number FROM numbers(100);")
+    moving_part = "all_0_0_0"
+    move_response = ch1.get_query_request(
+        sql=f"ALTER TABLE test_remove MOVE PART '{moving_part}' TO DISK 's3'",
+        database=DATABASE_NAME,
+    )
+
+    wait_part_is_stuck(ch1, table_moving_path, moving_part)
+
+    cluster.stop_zookeeper_nodes(["zoo1", "zoo2", "zoo3"])
+    # Stop moves in case table is not read-only yet
+    q(ch1, "SYSTEM STOP MOVES")
+    q(ch1, "SYSTEM DISABLE FAILPOINT stop_moving_part_before_swap_with_active")
+
+    assert "Cancelled moving parts" in move_response.get_error()
+    assert exec(ch1, "ls", table_moving_path).strip() == ""
+
+    cluster.start_zookeeper_nodes(["zoo1", "zoo2", "zoo3"])
+    wait_zookeeper_node_to_start(["zoo1", "zoo2", "zoo3"])
+    q(ch1, "SYSTEM START MOVES")
+
+    q(ch1, f"DROP TABLE test_remove")