ClickHouse/dbms/src/Storages/StorageDistributed.cpp

558 lines
19 KiB
C++
Raw Normal View History

2018-12-19 12:38:13 +00:00
#include <Storages/StorageDistributed.h>
#include <DataStreams/OneBlockInputStream.h>
#include <DataStreams/materializeBlock.h>
#include <Databases/IDatabase.h>
#include <DataTypes/DataTypeFactory.h>
2018-12-19 12:38:13 +00:00
#include <DataTypes/DataTypesNumber.h>
#include <Storages/Distributed/DirectoryMonitor.h>
2018-12-19 12:38:13 +00:00
#include <Storages/Distributed/DistributedBlockOutputStream.h>
#include <Storages/StorageFactory.h>
2018-12-25 23:14:39 +00:00
#include <Storages/AlterCommands.h>
#include <Common/Macros.h>
#include <Common/escapeForFileName.h>
2017-07-13 20:58:19 +00:00
#include <Common/typeid_cast.h>
2018-12-19 12:38:13 +00:00
#include <Parsers/ASTDropQuery.h>
#include <Parsers/ASTExpressionList.h>
#include <Parsers/ASTIdentifier.h>
#include <Parsers/ASTInsertQuery.h>
2018-12-19 12:38:13 +00:00
#include <Parsers/ASTLiteral.h>
#include <Parsers/ASTSelectQuery.h>
2018-12-19 12:38:13 +00:00
#include <Parsers/ASTTablesInSelectQuery.h>
#include <Parsers/ParserAlterQuery.h>
2018-12-19 12:38:13 +00:00
#include <Parsers/TablePropertiesQueriesASTs.h>
#include <Parsers/parseQuery.h>
2018-12-19 12:38:13 +00:00
#include <Interpreters/ClusterProxy/SelectStreamFactory.h>
#include <Interpreters/ClusterProxy/executeQuery.h>
#include <Interpreters/ExpressionAnalyzer.h>
#include <Interpreters/InterpreterAlterQuery.h>
#include <Interpreters/InterpreterDescribeQuery.h>
2018-12-19 12:38:13 +00:00
#include <Interpreters/InterpreterSelectQuery.h>
#include <Interpreters/SyntaxAnalyzer.h>
2018-12-19 12:38:13 +00:00
#include <Interpreters/createBlockSelector.h>
#include <Interpreters/evaluateConstantExpression.h>
#include <Interpreters/getClusterName.h>
#include <Core/Field.h>
2012-05-21 20:38:34 +00:00
2018-06-05 19:46:49 +00:00
#include <IO/ReadHelpers.h>
2016-12-12 03:33:34 +00:00
#include <Poco/DirectoryIterator.h>
2015-02-10 21:10:58 +00:00
#include <memory>
#include <boost/filesystem.hpp>
2016-12-12 03:33:34 +00:00
2012-05-21 20:38:34 +00:00
namespace DB
{
namespace ErrorCodes
{
extern const int STORAGE_REQUIRES_PARAMETER;
extern const int BAD_ARGUMENTS;
extern const int READONLY;
extern const int NUMBER_OF_ARGUMENTS_DOESNT_MATCH;
extern const int INCORRECT_NUMBER_OF_COLUMNS;
extern const int INFINITE_LOOP;
2018-06-05 19:46:49 +00:00
extern const int TYPE_MISMATCH;
2018-07-06 16:34:53 +00:00
extern const int NO_SUCH_COLUMN_IN_TABLE;
2018-12-19 12:38:13 +00:00
extern const int TOO_MANY_ROWS;
}
2014-08-21 12:07:29 +00:00
namespace
{
2018-07-25 12:31:47 +00:00
/// select query has database, table and table function names as AST pointers
/// Creates a copy of query, changes database, table and table function names.
ASTPtr rewriteSelectQuery(const ASTPtr & query, const std::string & database, const std::string & table, ASTPtr table_function_ptr = nullptr)
{
2018-02-25 00:50:53 +00:00
auto modified_query_ast = query->clone();
2018-07-25 12:31:47 +00:00
if (table_function_ptr)
typeid_cast<ASTSelectQuery &>(*modified_query_ast).addTableFunction(table_function_ptr);
else
typeid_cast<ASTSelectQuery &>(*modified_query_ast).replaceDatabaseAndTable(database, table);
return modified_query_ast;
}
/// The columns list in the original INSERT query is incorrect because inserted blocks are transformed
/// to the form of the sample block of the Distributed table. So we rewrite it and add all columns from
/// the sample block instead.
ASTPtr createInsertToRemoteTableQuery(const std::string & database, const std::string & table, const Block & sample_block)
{
auto query = std::make_shared<ASTInsertQuery>();
query->database = database;
query->table = table;
auto columns = std::make_shared<ASTExpressionList>();
query->columns = columns;
query->children.push_back(columns);
for (const auto & col : sample_block)
columns->children.push_back(std::make_shared<ASTIdentifier>(col.name));
return query;
}
/// Calculate maximum number in file names in directory and all subdirectories.
/// To ensure global order of data blocks yet to be sent across server restarts.
UInt64 getMaximumFileNumber(const std::string & path)
{
UInt64 res = 0;
boost::filesystem::recursive_directory_iterator begin(path);
boost::filesystem::recursive_directory_iterator end;
for (auto it = begin; it != end; ++it)
{
const auto & path = it->path();
if (it->status().type() != boost::filesystem::regular_file || !endsWith(path.filename().string(), ".bin"))
continue;
UInt64 num = 0;
try
{
num = parse<UInt64>(path.filename().stem().string());
}
catch (Exception & e)
{
e.addMessage("Unexpected file name " + path.filename().string() + " found at " + path.parent_path().string() + ", should have numeric base name.");
throw;
}
if (num > res)
res = num;
}
return res;
}
void initializeFileNamesIncrement(const std::string & path, SimpleIncrement & increment)
{
if (!path.empty())
increment.set(getMaximumFileNumber(path));
}
2018-12-19 12:38:13 +00:00
/// the same as DistributedBlockOutputStream::createSelector, should it be static?
IColumn::Selector createSelector(const ClusterPtr cluster, const ColumnWithTypeAndName & result)
{
const auto & slot_to_shard = cluster->getSlotToShard();
#define CREATE_FOR_TYPE(TYPE) \
if (typeid_cast<const DataType##TYPE *>(result.type.get())) \
return createBlockSelector<TYPE>(*result.column, slot_to_shard);
CREATE_FOR_TYPE(UInt8)
CREATE_FOR_TYPE(UInt16)
CREATE_FOR_TYPE(UInt32)
CREATE_FOR_TYPE(UInt64)
CREATE_FOR_TYPE(Int8)
CREATE_FOR_TYPE(Int16)
CREATE_FOR_TYPE(Int32)
CREATE_FOR_TYPE(Int64)
#undef CREATE_FOR_TYPE
throw Exception{"Sharding key expression does not evaluate to an integer type", ErrorCodes::TYPE_MISMATCH};
}
}
/// For destruction of std::unique_ptr of type that is incomplete in class definition.
StorageDistributed::~StorageDistributed() = default;
static ExpressionActionsPtr buildShardingKeyExpression(const ASTPtr & sharding_key, const Context & context, NamesAndTypesList columns, bool project)
{
ASTPtr query = sharding_key;
auto syntax_result = SyntaxAnalyzer(context, {}).analyze(query, columns);
return ExpressionAnalyzer(query, syntax_result, context).getActions(project);
}
StorageDistributed::StorageDistributed(
const String & database_name,
const String & table_name_,
const ColumnsDescription & columns_,
const String & remote_database_,
const String & remote_table_,
const String & cluster_name_,
const Context & context_,
const ASTPtr & sharding_key_,
const String & data_path_,
bool attach)
: IStorage{columns_},
table_name(table_name_),
remote_database(remote_database_), remote_table(remote_table_),
context(context_), cluster_name(context.getMacros()->expand(cluster_name_)), has_sharding_key(sharding_key_),
sharding_key_expr(sharding_key_ ? buildShardingKeyExpression(sharding_key_, context, getColumns().getAllPhysical(), false) : nullptr),
sharding_key_column_name(sharding_key_ ? sharding_key_->getColumnName() : String{}),
path(data_path_.empty() ? "" : (data_path_ + escapeForFileName(table_name) + '/'))
{
/// Sanity check. Skip check if the table is already created to allow the server to start.
if (!attach && !cluster_name.empty())
{
size_t num_local_shards = context.getCluster(cluster_name)->getLocalShardCount();
if (num_local_shards && remote_database == database_name && remote_table == table_name)
throw Exception("Distributed table " + table_name + " looks at itself", ErrorCodes::INFINITE_LOOP);
}
}
StorageDistributed::StorageDistributed(
const String & database_name,
const String & table_name_,
const ColumnsDescription & columns_,
ASTPtr remote_table_function_ptr_,
const String & cluster_name_,
const Context & context_,
const ASTPtr & sharding_key_,
const String & data_path_,
bool attach)
2018-07-25 12:31:47 +00:00
: StorageDistributed(database_name, table_name_, columns_, String{}, String{}, cluster_name_, context_, sharding_key_, data_path_, attach)
{
remote_table_function_ptr = remote_table_function_ptr_;
}
StoragePtr StorageDistributed::createWithOwnCluster(
const std::string & table_name_,
const ColumnsDescription & columns_,
const String & remote_database_, /// database on remote servers.
const String & remote_table_, /// The name of the table on the remote servers.
ClusterPtr owned_cluster_,
const Context & context_)
{
auto res = ext::shared_ptr_helper<StorageDistributed>::create(
String{}, table_name_, columns_, remote_database_, remote_table_, String{}, context_, ASTPtr(), String(), false);
res->owned_cluster = owned_cluster_;
return res;
}
StoragePtr StorageDistributed::createWithOwnCluster(
const std::string & table_name_,
const ColumnsDescription & columns_,
2018-07-25 12:31:47 +00:00
ASTPtr & remote_table_function_ptr_,
ClusterPtr & owned_cluster_,
const Context & context_)
{
auto res = ext::shared_ptr_helper<StorageDistributed>::create(
2018-07-25 12:31:47 +00:00
String{}, table_name_, columns_, remote_table_function_ptr_, String{}, context_, ASTPtr(), String(), false);
res->owned_cluster = owned_cluster_;
return res;
}
QueryProcessingStage::Enum StorageDistributed::getQueryProcessingStage(const Context & context) const
2012-05-21 20:38:34 +00:00
{
auto cluster = getCluster();
2018-06-29 11:42:44 +00:00
return getQueryProcessingStage(context, cluster);
}
2018-06-29 11:42:44 +00:00
QueryProcessingStage::Enum StorageDistributed::getQueryProcessingStage(const Context & context, const ClusterPtr & cluster) const
{
const Settings & settings = context.getSettingsRef();
size_t num_local_shards = cluster->getLocalShardCount();
size_t num_remote_shards = cluster->getRemoteShardCount();
size_t result_size = (num_remote_shards * settings.max_parallel_replicas) + num_local_shards;
if (settings.distributed_group_by_no_merge)
return QueryProcessingStage::Complete;
else /// Normal mode.
return result_size == 1 ? QueryProcessingStage::Complete
: QueryProcessingStage::WithMergeableState;
}
BlockInputStreams StorageDistributed::read(
const Names & /*column_names*/,
const SelectQueryInfo & query_info,
const Context & context,
QueryProcessingStage::Enum processed_stage,
const size_t /*max_block_size*/,
const unsigned /*num_streams*/)
{
auto cluster = getCluster();
2018-06-29 11:42:44 +00:00
const Settings & settings = context.getSettingsRef();
const auto & modified_query_ast = rewriteSelectQuery(
query_info.query, remote_database, remote_table, remote_table_function_ptr);
2018-07-25 13:16:36 +00:00
Block header = materializeBlock(InterpreterSelectQuery(query_info.query, context, Names{}, processed_stage).getSampleBlock());
2018-09-10 03:59:48 +00:00
ClusterProxy::SelectStreamFactory select_stream_factory = remote_table_function_ptr
? ClusterProxy::SelectStreamFactory(
2018-07-25 12:31:47 +00:00
header, processed_stage, remote_table_function_ptr, context.getExternalTables())
: ClusterProxy::SelectStreamFactory(
header, processed_stage, QualifiedTableName{remote_database, remote_table}, context.getExternalTables());
2015-11-06 17:44:01 +00:00
2018-12-20 17:48:11 +00:00
if (settings.optimize_skip_unused_shards)
2018-11-21 04:02:56 +00:00
{
2018-12-05 15:48:06 +00:00
auto smaller_cluster = skipUnusedShards(cluster, query_info);
if (smaller_cluster)
cluster = smaller_cluster;
}
return ClusterProxy::executeQuery(
select_stream_factory, cluster, modified_query_ast, context, settings);
2012-05-21 20:38:34 +00:00
}
BlockOutputStreamPtr StorageDistributed::write(const ASTPtr &, const Settings & settings)
{
auto cluster = getCluster();
/// Ban an attempt to make async insert into the table belonging to DatabaseMemory
if (path.empty() && !owned_cluster && !settings.insert_distributed_sync.value)
{
throw Exception("Storage " + getName() + " must has own data directory to enable asynchronous inserts",
ErrorCodes::BAD_ARGUMENTS);
}
/// If sharding key is not specified, then you can only write to a shard containing only one shard
if (!has_sharding_key && ((cluster->getLocalShardCount() + cluster->getRemoteShardCount()) >= 2))
{
throw Exception("Method write is not supported by storage " + getName() + " with more than one shard and no sharding key provided",
ErrorCodes::STORAGE_REQUIRES_PARAMETER);
}
2014-08-21 12:07:29 +00:00
/// Force sync insertion if it is remote() table function
bool insert_sync = settings.insert_distributed_sync || owned_cluster;
auto timeout = settings.insert_distributed_timeout;
/// DistributedBlockOutputStream will not own cluster, but will own ConnectionPools of the cluster
return std::make_shared<DistributedBlockOutputStream>(
*this, createInsertToRemoteTableQuery(remote_database, remote_table, getSampleBlock()), cluster, settings, insert_sync, timeout);
}
2016-01-28 01:00:27 +00:00
void StorageDistributed::alter(const AlterCommands & params, const String & database_name, const String & table_name, const Context & context)
2013-09-23 12:01:19 +00:00
{
auto lock = lockStructureForAlter();
2016-05-13 21:08:19 +00:00
ColumnsDescription new_columns = getColumns();
params.apply(new_columns);
context.getDatabase(database_name)->alterTable(context, table_name, new_columns, {});
setColumns(std::move(new_columns));
2013-09-23 12:01:19 +00:00
}
void StorageDistributed::startup()
{
createDirectoryMonitors();
initializeFileNamesIncrement(path, file_names_increment);
}
void StorageDistributed::shutdown()
{
cluster_nodes_data.clear();
}
void StorageDistributed::truncate(const ASTPtr &, const Context &)
2018-04-21 00:35:20 +00:00
{
2018-06-09 15:48:22 +00:00
std::lock_guard lock(cluster_nodes_mutex);
2018-04-21 00:35:20 +00:00
2018-06-09 15:48:22 +00:00
for (auto it = cluster_nodes_data.begin(); it != cluster_nodes_data.end();)
2018-04-21 00:35:20 +00:00
{
2018-06-09 15:48:22 +00:00
it->second.shutdownAndDropAllData();
it = cluster_nodes_data.erase(it);
2018-04-21 00:35:20 +00:00
}
}
namespace
{
/// NOTE This is weird. Get rid of this.
std::map<String, String> virtual_columns =
{
{"_table", "String"},
{"_part", "String"},
{"_part_index", "UInt64"},
{"_partition_id", "String"},
{"_sample_factor", "Float64"},
};
}
NameAndTypePair StorageDistributed::getColumn(const String & column_name) const
2014-07-31 09:11:49 +00:00
{
2018-07-06 16:34:53 +00:00
if (getColumns().hasPhysical(column_name))
return getColumns().getPhysical(column_name);
auto it = virtual_columns.find(column_name);
if (it != virtual_columns.end())
return { it->first, DataTypeFactory::instance().get(it->second) };
2014-07-31 09:11:49 +00:00
2018-07-06 16:34:53 +00:00
throw Exception("There is no column " + column_name + " in table.", ErrorCodes::NO_SUCH_COLUMN_IN_TABLE);
2014-07-31 09:11:49 +00:00
}
2014-07-31 09:11:49 +00:00
bool StorageDistributed::hasColumn(const String & column_name) const
{
return virtual_columns.count(column_name) || getColumns().hasPhysical(column_name);
2014-07-31 09:11:49 +00:00
}
void StorageDistributed::createDirectoryMonitors()
{
if (path.empty())
return;
Poco::File{path}.createDirectory();
boost::filesystem::directory_iterator begin(path);
boost::filesystem::directory_iterator end;
for (auto it = begin; it != end; ++it)
if (it->status().type() == boost::filesystem::directory_file)
requireDirectoryMonitor(it->path().filename().string());
}
void StorageDistributed::requireDirectoryMonitor(const std::string & name)
{
std::lock_guard lock(cluster_nodes_mutex);
cluster_nodes_data[name].requireDirectoryMonitor(name, *this);
}
ConnectionPoolPtr StorageDistributed::requireConnectionPool(const std::string & name)
{
std::lock_guard lock(cluster_nodes_mutex);
auto & node_data = cluster_nodes_data[name];
node_data.requireConnectionPool(name, *this);
return node_data.conneciton_pool;
}
2015-09-18 13:36:10 +00:00
size_t StorageDistributed::getShardCount() const
{
return getCluster()->getShardCount();
}
ClusterPtr StorageDistributed::getCluster() const
{
return owned_cluster ? owned_cluster : context.getCluster(cluster_name);
2015-09-18 13:36:10 +00:00
}
void StorageDistributed::ClusterNodeData::requireConnectionPool(const std::string & name, const StorageDistributed & storage)
{
if (!conneciton_pool)
conneciton_pool = StorageDistributedDirectoryMonitor::createPool(name, storage);
}
void StorageDistributed::ClusterNodeData::requireDirectoryMonitor(const std::string & name, StorageDistributed & storage)
{
requireConnectionPool(name, storage);
if (!directory_monitor)
directory_monitor = std::make_unique<StorageDistributedDirectoryMonitor>(storage, name, conneciton_pool);
}
2018-04-21 00:35:20 +00:00
void StorageDistributed::ClusterNodeData::shutdownAndDropAllData()
{
directory_monitor->shutdownAndDropAllData();
}
2018-12-19 12:38:13 +00:00
/// Returns a new cluster with fewer shards if constant folding for `sharding_key_expr` is possible
/// using constraints from "WHERE" condition, otherwise returns `nullptr`
ClusterPtr StorageDistributed::skipUnusedShards(ClusterPtr cluster, const SelectQueryInfo & query_info)
{
const auto & select = typeid_cast<ASTSelectQuery &>(*query_info.query);
if (!select.where_expression)
{
return nullptr;
}
const auto & blocks = evaluateExpressionOverConstantCondition(select.where_expression, sharding_key_expr);
// Can't get definite answer if we can skip any shards
if (!blocks)
{
return nullptr;
}
std::set<int> shards;
for (const auto & block : *blocks)
{
if (!block.has(sharding_key_column_name))
throw Exception("sharding_key_expr should evaluate as a single row", ErrorCodes::TOO_MANY_ROWS);
const auto result = block.getByName(sharding_key_column_name);
const auto selector = createSelector(cluster, result);
shards.insert(selector.begin(), selector.end());
}
return cluster->getClusterWithMultipleShards({shards.begin(), shards.end()});
}
void registerStorageDistributed(StorageFactory & factory)
{
factory.registerStorage("Distributed", [](const StorageFactory::Arguments & args)
{
/** Arguments of engine is following:
* - name of cluster in configuration;
* - name of remote database;
* - name of remote table;
*
* Remote database may be specified in following form:
* - identifier;
* - constant expression with string result, like currentDatabase();
* -- string literal as specific case;
* - empty string means 'use default database from cluster'.
*/
ASTs & engine_args = args.engine_args;
if (!(engine_args.size() == 3 || engine_args.size() == 4))
throw Exception("Storage Distributed requires 3 or 4 parameters"
" - name of configuration section with list of remote servers, name of remote database, name of remote table,"
" sharding key expression (optional).", ErrorCodes::NUMBER_OF_ARGUMENTS_DOESNT_MATCH);
String cluster_name = getClusterName(*engine_args[0]);
engine_args[1] = evaluateConstantExpressionOrIdentifierAsLiteral(engine_args[1], args.local_context);
engine_args[2] = evaluateConstantExpressionOrIdentifierAsLiteral(engine_args[2], args.local_context);
String remote_database = static_cast<const ASTLiteral &>(*engine_args[1]).value.safeGet<String>();
String remote_table = static_cast<const ASTLiteral &>(*engine_args[2]).value.safeGet<String>();
const auto & sharding_key = engine_args.size() == 4 ? engine_args[3] : nullptr;
/// Check that sharding_key exists in the table and has numeric type.
if (sharding_key)
{
auto sharding_expr = buildShardingKeyExpression(sharding_key, args.context, args.columns.getAllPhysical(), true);
const Block & block = sharding_expr->getSampleBlock();
if (block.columns() != 1)
throw Exception("Sharding expression must return exactly one column", ErrorCodes::INCORRECT_NUMBER_OF_COLUMNS);
auto type = block.getByPosition(0).type;
if (!type->isValueRepresentedByInteger())
throw Exception("Sharding expression has type " + type->getName() +
", but should be one of integer type", ErrorCodes::TYPE_MISMATCH);
}
return StorageDistributed::create(
args.database_name, args.table_name, args.columns,
remote_database, remote_table, cluster_name,
args.context, sharding_key, args.data_path,
args.attach);
});
}
2012-05-21 20:38:34 +00:00
}