PostgreSQL sharding : citus 系列5 - worker节点网络优化
背景
citus 节点间的网络需求:
1、cn节点访问所有worker节点。oltp业务的访问较频繁。
2、重分布数据时,worker节点间相互访问。访问频度不大,OLAP业务常见,一旦有可能数据交换吞吐较大。
citus的cn节点连worker节点为有两种模式,
一种为事务级保持连接模式(每条SQL发起时建立连接,SQL结束后释放连接(除非在事务中,否则SQL结束立即释放连接)。),
另一种为会话级保持连接模式(会话发起时建立连接,会话结束后释放连接。)。
1、跑OLAP类的SQL时,使用的是第一种即时连接模式(OLAP场景并发不高,建立连接带来的额外开销不明显)
可以在worker节点打开参数进行跟踪
postgres=# show log_connections ;
log_connections
-----------------
on
(1 row)
postgres=# show log_disconnections ;
log_disconnections
--------------------
on
(1 row)
例子,
以下两条SQL均为即时短连接模式(Custom Scan (Citus Task-Tracker) Custom Scan (Citus Real-Time)
)。
postgres=# set citus.task_executor_type =task;
ERROR: invalid value for parameter "citus.task_executor_type": "task"
HINT: Available values: real-time, task-tracker.
postgres=# set citus.task_executor_type ='task-tracker';
SET
postgres=# explain select count(*) from pgbench_accounts ;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------
Aggregate (cost=0.00..0.00 rows=0 width=0)
-> Custom Scan (Citus Task-Tracker) (cost=0.00..0.00 rows=0 width=0)
Task Count: 128
Tasks Shown: One of 128
-> Task
Node: host=172.24.211.224 port=1921 dbname=postgres
-> Aggregate (cost=231.85..231.86 rows=1 width=8)
-> Seq Scan on pgbench_accounts_106812 pgbench_accounts (cost=0.00..212.48 rows=7748 width=0)
(8 rows)
postgres=# set citus.task_executor_type ='real-time';
postgres=# explain select count(*) from pgbench_accounts ;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------
Aggregate (cost=0.00..0.00 rows=0 width=0)
-> Custom Scan (Citus Real-Time) (cost=0.00..0.00 rows=0 width=0)
Task Count: 128
Tasks Shown: One of 128
-> Task
Node: host=172.24.211.224 port=1921 dbname=postgres
-> Aggregate (cost=231.85..231.86 rows=1 width=8)
-> Seq Scan on pgbench_accounts_106812 pgbench_accounts (cost=0.00..212.48 rows=7748 width=0)
(8 rows)
2、跑OLTP查询时(通常并发很高,前端有连接池(保持会话)),为会话级保持连接模式(Custom Scan (Citus Router)
)。
以下SQL为长连接模式(不会立即释放,而是等会再释放,以降低高并发时连接带来的开销)
postgres=# explain select * from pgbench_accounts where aid=5;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------
Custom Scan (Citus Router) (cost=0.00..0.00 rows=0 width=0)
Task Count: 1
Tasks Shown: All
-> Task
Node: host=172.24.211.224 port=1921 dbname=postgres
-> Index Scan using pgbench_accounts_pkey_106836 on pgbench_accounts_106836 pgbench_accounts (cost=0.28..2.50 rows=1 width=97)
Index Cond: (aid = 5)
(7 rows)
看以上两种场景,CITUS应该说设计得已经很不错了。既能满足TP也能满足AP。
但是前面也说了,连接保持是在事务或会话层面的,如果查询量大,或者用户使用了短连接,建立连接的开销就会很凸显。
newdb=> \c postgres postgres
You are now connected to database "postgres" as user "postgres".
postgres=# select * from pgbench_accounts where aid=1;
aid | bid | abalance | filler
-----+-----+----------+--------------------------------------------------------------------------------------
1 | 1 | 7214 |
(1 row)
Time: 11.264 ms -- 包括新建连接的开销
postgres=# select * from pgbench_accounts where aid=1;
aid | bid | abalance | filler
-----+-----+----------+--------------------------------------------------------------------------------------
1 | 1 | 7214 |
(1 row)
Time: 0.905 ms -- 已建立连接
使用pgbouncer,解决建立连接的开销
在worker节点上,部署pgbouncer,所有与worker节点建立的连接都通过pgbouncer连接池,以此来保持住连接,降低worker节点频繁新建连接的开销。
部署方法
1、所有worker节点
pgbouncer
yum install -y pgbouncer
配置
vi /etc/pgbouncer/pgb.ini
[databases]
newdb = host=/tmp dbname=newdb port=1921 user=digoal pool_size=128 reserve_pool=10
[pgbouncer]
logfile = /var/log/pgbouncer/pgbouncer.log
pidfile = /var/run/pgbouncer/pgbouncer.pid
listen_addr = 0.0.0.0
listen_port = 8001
auth_type = any
auth_file = /etc/pgbouncer/userlist.txt
pool_mode = session
server_reset_query = DISCARD ALL
max_client_conn = 5000
default_pool_size = 128
; 最大不要超过4倍CPU数
启动
pgbouncer -d -u pgbouncer /etc/pgbouncer/pgb.ini
在一个citus集群中,可以同时存在直连worker或者通过pgbouncer连接worker。
不同的database可以有不同的配置。
以下例子,新建一个database,使用pgbouncer连接worker.
2、所有节点(包括cn, worker)
新建数据库,插件
su - postgres
psql -c "create role digoal login;"
psql -c "create database newdb;"
psql -c "grant all on database newdb to digoal;"
psql -U postgres newdb -c "create extension citus;"
cn节点
将worker添加到集群配置,使用pgbouncer的连接端口
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.224', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.230', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.231', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.225', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.227', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.232', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.226', 8001);\""
su - postgres -c "psql -U postgres newdb -c \"SELECT * from master_add_node('xxx.xxx.xxx.229', 8001);\""
MX配置,同样,将worker节点添加到元数据同步中。
psql newdb postgres
select * from master_add_node('xxx.xxx.xxx.224',8001);
select * from master_add_node('xxx.xxx.xxx.230',8001);
开启同步到元数据。
select start_metadata_sync_to_node('xxx.xxx.xxx.224',8001);
select start_metadata_sync_to_node('xxx.xxx.xxx.230',8001);
测试
1、tpc-b 长连接测试
pgbench -i -s -U digoal newdb
psql -U digoal newdb
select create_distributed_table('pgbench_accounts','aid');
select create_distributed_table('pgbench_branches','bid');
select create_distributed_table('pgbench_tellers','tid');
select create_distributed_table('pgbench_history','aid');
pgbench -M prepared -v -r -P 1 -c 64 -j 64 -T 120 -U digoal newdb -S
性能与不使用pgbouncer差不多,因为使用了长连接测试简单SQL(本身citus就使用了会话级连接保持,没有短连接问题)。
citus第一次QUERY的隐藏耗时
在一个新建的会话中,第一次查询总是需要需要耗费更多的时间(如果没有建立连接,那么包括建立连接的时间。即使已建立连接,也需要耗费额外的一些时间)(具体原因可以跟踪分析一下code)。
以下使用的是pgbouncer连接worker,因此第一次QUERY不包含建立连接的时间。
newdb=> \q
postgres@digoal-citus-gpdb-test001-> psql newdb digoal
psql (10.5)
Type "help" for help.
\timing
newdb=> select * from pgbench_accounts where aid=5;
aid | bid | abalance | filler
-----+-----+----------+--------------------------------------------------------------------------------------
5 | 1 | 0 |
(1 row)
Time: 6.016 ms -- 不包括新建连接(已使用pgbouncer建立),但是也多了几毫秒
-- 但是相比未使用pgbouncer,已经降低了5毫秒左右的延迟。
newdb=> select * from pgbench_accounts where aid=5;
aid | bid | abalance | filler
-----+-----+----------+--------------------------------------------------------------------------------------
5 | 1 | 0 |
(1 row)
Time: 0.989 ms
多出的几毫秒,我们社区的小伙伴邓彪、王健给出了原因如下,很多地方需要检查安装的citus版本与citus.control控制文件的版本是否兼容(比如加载分布式TABLE的RELCACHE时,第一次访问就是这个问题),不兼容报错:
详见代码
https://github.com/citusdata/citus/blob/3fa04d8f2c8f27b1377fe4c1468ee47358117e3c/src/backend/distributed/utils/metadata_cache.c
/*
* CheckAvailableVersion compares CITUS_EXTENSIONVERSION and the currently
* available version from the citus.control file. If they are not compatible,
* this function logs an error with the specified elevel and returns false,
* otherwise it returns true.
*/
bool
CheckAvailableVersion(int elevel)
{
char *availableVersion = NULL;
if (!EnableVersionChecks)
{
return true;
}
availableVersion = AvailableExtensionVersion();
if (!MajorVersionsCompatible(availableVersion, CITUS_EXTENSIONVERSION))
{
ereport(elevel, (errmsg("loaded Citus library version differs from latest "
"available extension version"),
errdetail("Loaded library requires %s, but the latest control "
"file specifies %s.", CITUS_MAJORVERSION,
availableVersion),
errhint("Restart the database to load the latest Citus "
"library.")));
return false;
}
return true;
}
/*
* AvailableExtensionVersion returns the Citus version from citus.control file. It also
* saves the result, thus consecutive calls to CitusExtensionAvailableVersion will
* not read the citus.control file again.
*/
static char *
AvailableExtensionVersion(void)
{
ReturnSetInfo *extensionsResultSet = NULL;
TupleTableSlot *tupleTableSlot = NULL;
FunctionCallInfoData *fcinfo = NULL;
FmgrInfo *flinfo = NULL;
int argumentCount = 0;
EState *estate = NULL;
bool hasTuple = false;
bool goForward = true;
bool doCopy = false;
char *availableExtensionVersion;
InitializeCaches();
estate = CreateExecutorState();
extensionsResultSet = makeNode(ReturnSetInfo);
extensionsResultSet->econtext = GetPerTupleExprContext(estate);
extensionsResultSet->allowedModes = SFRM_Materialize;
fcinfo = palloc0(sizeof(FunctionCallInfoData));
flinfo = palloc0(sizeof(FmgrInfo));
fmgr_info(F_PG_AVAILABLE_EXTENSIONS, flinfo);
InitFunctionCallInfoData(*fcinfo, flinfo, argumentCount, InvalidOid, NULL,
(Node *) extensionsResultSet);
/* pg_available_extensions returns result set containing all available extensions */
(*pg_available_extensions)(fcinfo);
tupleTableSlot = MakeSingleTupleTableSlot(extensionsResultSet->setDesc);
hasTuple = tuplestore_gettupleslot(extensionsResultSet->setResult, goForward, doCopy,
tupleTableSlot);
while (hasTuple)
{
Datum extensionNameDatum = 0;
char *extensionName = NULL;
bool isNull = false;
extensionNameDatum = slot_getattr(tupleTableSlot, 1, &isNull);
extensionName = NameStr(*DatumGetName(extensionNameDatum));
if (strcmp(extensionName, "citus") == 0)
{
MemoryContext oldMemoryContext = NULL;
Datum availableVersion = slot_getattr(tupleTableSlot, 2, &isNull);
/* we will cache the result of citus version to prevent catalog access */
oldMemoryContext = MemoryContextSwitchTo(CacheMemoryContext);
availableExtensionVersion = text_to_cstring(DatumGetTextPP(availableVersion));
MemoryContextSwitchTo(oldMemoryContext);
ExecClearTuple(tupleTableSlot);
ExecDropSingleTupleTableSlot(tupleTableSlot);
return availableExtensionVersion;
}
ExecClearTuple(tupleTableSlot);
hasTuple = tuplestore_gettupleslot(extensionsResultSet->setResult, goForward,
doCopy, tupleTableSlot);
}
ExecDropSingleTupleTableSlot(tupleTableSlot);
ereport(ERROR, (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
errmsg("citus extension is not found")));
return NULL;
}
pgbouncer连接池连接worker的好处
1、对于业务层短连接会有比较好的效果。可以降低至少5毫秒左右的延迟。
2、对于大量复杂查询(需要motion的查询),可以减少节点间的连接数。
同时也需要注意pgbouncer瓶颈:带宽,QPS(单核,单个pgbouncer进程峰值QPS约5万)
参考
man 1 pgbouncer
man 5 pgbouncer