共计 12479 个字符，预计需要花费 32 分钟才能阅读完成。
query 启动入口
Databend-query server 的启动入口在 databend/src/binaries/query/main.rs 下，在初始化配置之后，它会创立一个 GlobalServices 和 server 敞开时负责解决 shutdown 逻辑的 shutdown_handle
GlobalServices::init(conf.clone()).await?;
let mut shutdown_handle = ShutdownHandle::create()?;GlobalServices
GlobalServices 负责启动 databend-query 的所有全局服务，这些服务都遵循繁多责任准则。
pub struct GlobalServices {
global_runtime: UnsafeCell<Option<Arc<Runtime>>>,
// 负责解决 query log
query_logger: UnsafeCell<Option<Arc<QueryLogger>>>,
// 负责 databend query 集群发现
cluster_discovery: UnsafeCell<Option<Arc<ClusterDiscovery>>>,
// 负责与 storage 层交互来读写数据
storage_operator: UnsafeCell<Option<Operator>>,
async_insert_manager: UnsafeCell<Option<Arc<AsyncInsertManager>>>,
cache_manager: UnsafeCell<Option<Arc<CacheManager>>>,
catalog_manager: UnsafeCell<Option<Arc<CatalogManager>>>,
http_query_manager: UnsafeCell<Option<Arc<HttpQueryManager>>>,
data_exchange_manager: UnsafeCell<Option<Arc<DataExchangeManager>>>,
session_manager: UnsafeCell<Option<Arc<SessionManager>>>,
users_manager: UnsafeCell<Option<Arc<UserApiProvider>>>,
users_role_manager: UnsafeCell<Option<Arc<RoleCacheManager>>>,
}GlobalServices 中的全局服务都实现了单例 trait，这些全局管理器后续会有对应的源码剖析文章介绍，本文介绍与 Session 解决相干的逻辑。
pub trait SingletonImpl<T>: Send + Sync {fn get(&self) -> T;
fn init(&self, value: T) -> Result<()>;}
pub type Singleton<T> = Arc<dyn SingletonImpl<T>>;ShutdownHandle
接下来会依据网络协议初始化 handlers，并把它们注册到 shutdown_handler 的 services 中，任何实现 Server trait 的类型都能够被增加到 services 中。
#[async_trait::async_trait]
pub trait Server: Send {async fn shutdown(&mut self, graceful: bool);
async fn start(&mut self, listening: SocketAddr) -> Result<SocketAddr>;
}目前 Databend 反对三种协定提交查问申请（mysql, clickhouse http, raw http）。
// MySQL handler.
{let hostname = conf.query.mysql_handler_host.clone();
let listening = format!("{}:{}", hostname, conf.query.mysql_handler_port);
let mut handler = MySQLHandler::create(session_manager.clone());
let listening = handler.start(listening.parse()?).await?;
// 注册服务到 shutdown_handle 来解决 server shutdown 时候的敞开逻辑，下同
shutdown_handle.add_service(handler);
}
// ClickHouse HTTP handler.
{let hostname = conf.query.clickhouse_http_handler_host.clone();
let listening = format!("{}:{}", hostname, conf.query.clickhouse_http_handler_port);
let mut srv = HttpHandler::create(session_manager.clone(), HttpHandlerKind::Clickhouse);
let listening = srv.start(listening.parse()?).await?;
shutdown_handle.add_service(srv);
}
// Databend HTTP handler.
{let hostname = conf.query.http_handler_host.clone();
let listening = format!("{}:{}", hostname, conf.query.http_handler_port);
let mut srv = HttpHandler::create(session_manager.clone(), HttpHandlerKind::Query);
let listening = srv.start(listening.parse()?).await?;
shutdown_handle.add_service(srv);
}之后会创立一些其它服务
- Metric service: 指标服务
- Admin service: 负责解决治理信息
- RPC service: query 节点的 rpc 服务，负责 query 节点之间的通信，应用 arrow flight 协定
// Metric API service.
{let address = conf.query.metric_api_address.clone();
let mut srv = MetricService::create(session_manager.clone());
let listening = srv.start(address.parse()?).await?;
shutdown_handle.add_service(srv);
info!("Listening for Metric API: {}/metrics", listening);
}
// Admin HTTP API service.
{let address = conf.query.admin_api_address.clone();
let mut srv = HttpService::create(session_manager.clone());
let listening = srv.start(address.parse()?).await?;
shutdown_handle.add_service(srv);
info!("Listening for Admin HTTP API: {}", listening);
}
// RPC API service.
{let address = conf.query.flight_api_address.clone();
let mut srv = RpcService::create(session_manager.clone());
let listening = srv.start(address.parse()?).await?;
shutdown_handle.add_service(srv);
info!("Listening for RPC API (interserver): {}", listening);
}最初会将这个 query 节点注册到 meta server 中。
// Cluster register.
{let cluster_discovery = session_manager.get_cluster_discovery();
let register_to_metastore = cluster_discovery.register_to_metastore(&conf);
register_to_metastore.await?;
}Session 相干
session 次要分为 4 个局部
- session_manager: 全局惟一，负责管理 client session
- session: 每当有新的 client 连贯到 server 之后会创立一个新的 session 并且注册到 session_manager
- query_ctx: 每一条查问语句会有一个 query_ctx，用来存储以后查问的一些上下文信息
- query_ctx_shared: 查问语句中的子查问共享的上下文信息
上面逐个来剖析
SessionManager (query/src/sessions/session_mgr.rs)
pub struct SessionManager {pub(in crate::sessions) conf: Config,
pub(in crate::sessions) max_sessions: usize,
pub(in crate::sessions) active_sessions: Arc<RwLock<HashMap<String, Arc<Session>>>>,
pub status: Arc<RwLock<SessionManagerStatus>>,
// When session type is MySQL, insert into this map, key is id, val is MySQL connection id.
pub(crate) mysql_conn_map: Arc<RwLock<HashMap<Option<u32>, String>>>,
pub(in crate::sessions) mysql_basic_conn_id: AtomicU32,
}SessionManager 次要用来创立和销毁 session，对应办法如下
// 依据 client 协定类型来创立 session
pub async fn create_session(self: &Arc<Self>, typ: SessionType) -> Result<SessionRef>
// 依据 session id 来销毁 session
pub fn destroy_session(self: &Arc<Self>, session_id: &String)Session (query/src/sessions/session.rs)
session 次要存储 client-server 的上下文信息，代码命名曾经很清晰了，这里就不再过多赘述。
pub struct Session {pub(in crate::sessions) id: String,
pub(in crate::sessions) typ: RwLock<SessionType>,
pub(in crate::sessions) session_ctx: Arc<SessionContext>,
status: Arc<RwLock<SessionStatus>>,
pub(in crate::sessions) mysql_connection_id: Option<u32>,
}
pub struct SessionContext {
conf: Config,
abort: AtomicBool,
current_catalog: RwLock<String>,
current_database: RwLock<String>,
current_tenant: RwLock<String>,
current_user: RwLock<Option<UserInfo>>,
auth_role: RwLock<Option<String>>,
client_host: RwLock<Option<SocketAddr>>,
io_shutdown_tx: RwLock<Option<Sender<Sender<()>>>>,
query_context_shared: RwLock<Option<Arc<QueryContextShared>>>,
}
pub struct SessionStatus {
pub session_started_at: Instant,
pub last_query_finished_at: Option<Instant>,
}Session 的另一个大的性能是负责创立和获取 QueryContext，每次接管到新的 query 申请都会创立一个 QueryContext 并绑定在对应的 query 语句上。
QueryContext (query/src/sessions/query_ctx.rs)
QueryContext 次要是保护查问的上下文信息，它通过 QueryContext::create_from_shared(query_ctx_shared) 创立。
#[derive(Clone)]
pub struct QueryContext {
version: String,
statistics: Arc<RwLock<Statistics>>,
partition_queue: Arc<RwLock<VecDeque<PartInfoPtr>>>,
shared: Arc<QueryContextShared>,
precommit_blocks: Arc<RwLock<Vec<DataBlock>>>,
fragment_id: Arc<AtomicUsize>,
} 其中 partition_queue 次要存储查问对应的 PartInfo，包含 part 的地址、版本信息、波及数据的行数，part 应用的压缩算法、以及波及到 column 的 meta 信息。在 pipeline build 时候会去设置 partition。pipeline 后续会有专门的文章介绍。precommit_blocks 负责暂存插入操作的时曾经写入到存储，然而尚未提交的元数据，DataBlock 次要蕴含 Column 的元信息援用和 arrow schema 的信息。
QueryContextShared (query/src/sessions/query_ctx_shared.rs)
对于蕴含子查问的查问，须要共享很多上下文信息，这就是 QueryContextShared 存在的理由。
/// 数据须要在查问上下文中被共享，这个很重要，比方:
/// USE database_1;
/// SELECT
/// (SELECT scalar FROM table_name_1) AS scalar_1,
/// (SELECT scalar FROM table_name_2) AS scalar_2,
/// (SELECT scalar FROM table_name_3) AS scalar_3
/// FROM table_name_4;
/// 对于上体面查问, 会共享 runtime, session, progress, init_query_id
pub struct QueryContextShared {/// scan_progress for scan metrics of datablocks (uncompressed)
pub(in crate::sessions) scan_progress: Arc<Progress>,
/// write_progress for write/commit metrics of datablocks (uncompressed)
pub(in crate::sessions) write_progress: Arc<Progress>,
/// result_progress for metrics of result datablocks (uncompressed)
pub(in crate::sessions) result_progress: Arc<Progress>,
pub(in crate::sessions) error: Arc<Mutex<Option<ErrorCode>>>,
pub(in crate::sessions) session: Arc<Session>,
pub(in crate::sessions) runtime: Arc<RwLock<Option<Arc<Runtime>>>>,
pub(in crate::sessions) init_query_id: Arc<RwLock<String>>,
...
}它提供了 query 上下文所须要的所有根本信息。
Handler
之前提到了 Databend 反对多种 handler，上面就以 mysql 为例，看一下 handler 的解决流程以及如何与 session 产生交互。
首先 MySQLHandler 会蕴含一个 SessionManager 的援用
pub struct MySQLHandler {
abort_handle: AbortHandle,
abort_registration: Option<AbortRegistration>,
join_handle: Option<JoinHandle<()>>,}MySQLHandler 在启动后，会 spawn 一个 tokio task 来继续监听 tcp stream，并且创立一个 session 再启动一个 task 去执行之后的查问申请。
fn accept_socket(session_mgr: Arc<SessionManager>, executor: Arc<Runtime>, socket: TcpStream) {
executor.spawn(async move {
// 创立 session
match session_mgr.create_session(SessionType::MySQL).await {Err(error) => Self::reject_session(socket, error).await,
Ok(session) => {info!("MySQL connection coming: {:?}", socket.peer_addr());
// 执行查问
if let Err(error) = MySQLConnection::run_on_stream(session, socket) {error!("Unexpected error occurred during query: {:?}", error);
};
}
}
});
} 在 MySQLConnection::run_on_stream 中，session 会先 attach 到对应的 client host 并且注册一个 shutdown 闭包来解决敞开连贯敞开时须要执行的清理，要害代码如下：
// mysql_session.rs
pub fn run_on_stream(session: SessionRef, stream: TcpStream) -> Result<()> {let blocking_stream = Self::convert_stream(stream)?;
MySQLConnection::attach_session(&session, &blocking_stream)?;
...
}
fn attach_session(session: &SessionRef, blocking_stream: &std::net::TcpStream) -> Result<()> {let host = blocking_stream.peer_addr().ok();
let blocking_stream_ref = blocking_stream.try_clone()?;
session.attach(host, move || {
// 注册 shutdown 逻辑
if let Err(error) = blocking_stream_ref.shutdown(Shutdown::Both) {error!("Cannot shutdown MySQL session io {}", error);
}
});
Ok(())
}
// session.rs
pub fn attach<F>(self: &Arc<Self>, host: Option<SocketAddr>, io_shutdown: F)
where F: FnOnce() + Send + 'static {let (tx, rx) = oneshot::channel();
self.session_ctx.set_client_host(host);
self.session_ctx.set_io_shutdown_tx(Some(tx));
common_base::base::tokio::spawn(async move {
// 在 session quit 时候触发清理
if let Ok(tx) = rx.await {(io_shutdown)();
tx.send(()).ok();}
});
}之后会启动一个 MySQL InteractiveWorker 来解决后续的查问。
let join_handle = query_executor.spawn(async move {let client_addr = non_blocking_stream.peer_addr().unwrap().to_string();
let interactive_worker = InteractiveWorker::create(session, client_addr);
let opts = IntermediaryOptions {process_use_statement_on_query: true,};
let (r, w) = non_blocking_stream.into_split();
let w = BufWriter::with_capacity(DEFAULT_RESULT_SET_WRITE_BUFFER_SIZE, w);
AsyncMysqlIntermediary::run_with_options(interactive_worker, r, w, &opts).await
});
let _ = futures::executor::block_on(join_handle); 该 InteractiveWorker 会实现 AsyncMysqlShim trait 的办法，比方：on_execute、on_query 等。查问到来时会回调这些办法来执行查问。这里以 on_query 为例，要害代码如下：
async fn on_query<'a>(
&'a mut self,
query: &'a str,
writer: QueryResultWriter<'a, W>,
) -> Result<()> {
...
// response writer
let mut writer = DFQueryResultWriter::create(writer);
let instant = Instant::now();
// 执行查问
let blocks = self.base.do_query(query).await;
// 回写后果
let format = self.base.session.get_format_settings()?;
let mut write_result = writer.write(blocks, &format);
...
// metrics 信息
histogram!(
super::mysql_metrics::METRIC_MYSQL_PROCESSOR_REQUEST_DURATION,
instant.elapsed());
write_result
} 在 do_query 中会创立 QueryContext 并开始解析 sql 流程来实现后续的整个 sql 查问。要害代码如下：
// 创立 QueryContext
let context = self.session.create_query_context().await?;
// 关联到查问语句
context.attach_query_str(query);
let settings = context.get_settings();
// parse sql
let stmts_hints = DfParser::parse_sql(query, context.get_current_session().get_type());
...
// 创立并生成查问打算
let mut planner = Planner::new(context.clone());
let interpreter = planner.plan_sql(query).await.and_then(|v| {has_result_set = has_result_set_by_plan(&v.0);
InterpreterFactoryV2::get(context.clone(), &v.0)
})
// 执行查问，返回后果
Self::exec_query(interpreter.clone(), &context).await?;
let schema = interpreter.schema();
Ok(QueryResult::create(
blocks,
extra_info,
has_result_set,
schema,
))序幕
以上就是从 Databend 启动服务到承受 sql 申请并开始解决的流程。最近咱们因为一些起因（Clickhouse tcp 协定偏差 clickhouse 的底层，协定没有公开的文档阐明，同时外面历史包袱比拟重，排查问题节约大量精力）去掉了 ClickHouse native tcp client，具体请参见: https://github.com/datafusela…
如果你浏览完代码有好的提议，欢送来这里探讨，另外如果发现相干的问题，能够提交到 issue 来帮忙咱们进步 Databend 的稳定性。Databend 社区欢送所有善意的意见和倡议 :)
对于 Databend
Databend 是一款开源、弹性、低成本，基于对象存储也能够做实时剖析的旧式数仓。期待您的关注，一起摸索云原生数仓解决方案，打造新一代开源 Data Cloud。
- Databend 文档：https://databend.rs/
- Twitter：https://twitter.com/Datafuse_…
- Slack：https://datafusecloud.slack.com/
- Wechat：Databend
- GitHub：https://github.com/datafusela…
文章首发于公众号：Databend
正文完