GreatSQL删除分区慢的跟踪
背景
某业务系统，每天凌晨会删除分区表的一个分区(按天分区)，耗时较久，从最开始的30秒，慢慢变为1分钟 ，影响到交易业务的正常进行。 在测试环境进行了模拟，复现了删除分区慢的情况，本次基于GreatSQL8.0.25-17进行测试，官方mysql版本也存在相同问题。
测试环境
代码语言：javascript复制$ mysql -h127.0.0.1 -P8025 -uroot -p
mysql: [Warning] Using a password on the command line interface can be insecure.
Welcome to the MySQL monitor. Commands end with ; or g.
Your MySQL connection id is 10
Server version: 8.0.25-17 GreatSQL, Release 17, Revision 4733775f703
Copyright (c) 2000, 2021, Oracle and/or its affiliates.
Oracle is a registered trademark of Oracle Corporation and/or its
affiliates. Other names may be trademarks of their respective
owners.
Type 'help;' or 'h' for help. Type 'c' to clear the current input statement.
greatsql> select version();
-----------
| version() |
-----------
| 8.0.25-17 |
-----------
1 row in set (0.00 sec)
greatsql> show variables like 'autocommit' ;
--------------- -------
| Variable_name | Value |
--------------- -------
| autocommit | ON |
--------------- -------
1 row in set (0.01 sec)
greatsql> show variables like 'innodb_buffer_pool_size';
------------------------- ------------
| Variable_name | Value |
------------------------- ------------
| innodb_buffer_pool_size | 4294967296 |
------------------------- ------------
1 row in set (0.01 sec)
greatsql> show variables like 'innodb_flush_log_at_trx_commit';
-------------------------------- -------
| Variable_name | Value |
-------------------------------- -------
| innodb_flush_log_at_trx_commit | 1 |
-------------------------------- -------
1 row in set (0.00 sec)
greatsql> show variables like 'sync_binlog';
--------------- -------
| Variable_name | Value |
--------------- -------
| sync_binlog | 1 |
--------------- -------
1 row in set (0.00 sec)
建表
代码语言：javascript复制CREATE TABLE `t_partition` (
`id` bigint NOT NULL AUTO_INCREMENT,
`ua` varchar(100) DEFAULT NULL,
`start_time` datetime NOT NULL,
PRIMARY KEY (`id`,`start_time`)
) PARTITION BY RANGE (to_days(`start_time`))
(PARTITION P20230129 VALUES LESS THAN (738915) ENGINE = InnoDB,
PARTITION P20230130 VALUES LESS THAN (738916) ENGINE = InnoDB,
PARTITION P20230131 VALUES LESS THAN (738917) ENGINE = InnoDB,
PARTITION P20230201 VALUES LESS THAN (738918) ENGINE = InnoDB,
PARTITION P20230202 VALUES LESS THAN (738919) ENGINE = InnoDB,
PARTITION P20230203 VALUES LESS THAN (738920) ENGINE = InnoDB,
PARTITION P20230204 VALUES LESS THAN (738921) ENGINE = InnoDB,
PARTITION P20230302 VALUES LESS THAN (738947) ENGINE = InnoDB,
PARTITION P20230303 VALUES LESS THAN (738948) ENGINE = InnoDB,
PARTITION P20230304 VALUES LESS THAN (738949) ENGINE = InnoDB,
PARTITION P20230305 VALUES LESS THAN (738950) ENGINE = InnoDB,
PARTITION P20230306 VALUES LESS THAN (738951) ENGINE = InnoDB,
PARTITION P20230307 VALUES LESS THAN (738952) ENGINE = InnoDB,
PARTITION P20230308 VALUES LESS THAN (738953) ENGINE = InnoDB,
PARTITION P20230309 VALUES LESS THAN (738954) ENGINE = InnoDB,
PARTITION P20230310 VALUES LESS THAN (738955) ENGINE = InnoDB,
PARTITION P20230311 VALUES LESS THAN (738956) ENGINE = InnoDB,
PARTITION P20230312 VALUES LESS THAN (738957) ENGINE = InnoDB,
PARTITION P20230313 VALUES LESS THAN (738958) ENGINE = InnoDB,
PARTITION P20230314 VALUES LESS THAN (738959) ENGINE = InnoDB,
PARTITION P20230315 VALUES LESS THAN (738960) ENGINE = InnoDB,
PARTITION P20230316 VALUES LESS THAN (738961) ENGINE = InnoDB,
PARTITION P20230317 VALUES LESS THAN (738962) ENGINE = InnoDB,
PARTITION P20230318 VALUES LESS THAN (738963) ENGINE = InnoDB,
PARTITION P20230319 VALUES LESS THAN (738964) ENGINE = InnoDB,
PARTITION P20230320 VALUES LESS THAN (738965) ENGINE = InnoDB,
PARTITION P20230321 VALUES LESS THAN (738966) ENGINE = InnoDB,
PARTITION P20230322 VALUES LESS THAN (738967) ENGINE = InnoDB,
PARTITION P20230323 VALUES LESS THAN (738968) ENGINE = InnoDB,
PARTITION P20230324 VALUES LESS THAN (738969) ENGINE = InnoDB,
PARTITION P20230325 VALUES LESS THAN (738970) ENGINE = InnoDB,
PARTITION P20230326 VALUES LESS THAN (738971) ENGINE = InnoDB,
PARTITION P20230327 VALUES LESS THAN (738972) ENGINE = InnoDB,
PARTITION P20230328 VALUES LESS THAN (738973) ENGINE = InnoDB,
PARTITION p20230329 VALUES LESS THAN (738974) ENGINE = InnoDB,
PARTITION p20230330 VALUES LESS THAN (738975) ENGINE = InnoDB,
PARTITION p20230331 VALUES LESS THAN (738976) ENGINE = InnoDB,
PARTITION p20230401 VALUES LESS THAN (738977) ENGINE = InnoDB,
PARTITION p20230402 VALUES LESS THAN (738978) ENGINE = InnoDB,
PARTITION p20230403 VALUES LESS THAN (738979) ENGINE = InnoDB);
插入数据
代码语言：javascript复制greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-10' ;
Query OK, 1 row affected (0.01 sec)
Records: 1 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-10' ;
Query OK, 1 row affected (0.00 sec)
Records: 1 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-10' ;
Query OK, 1 row affected (0.01 sec)
Records: 1 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),start_time from t_partition;
Query OK, 3 rows affected (0.01 sec)
Records: 3 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),start_time from t_partition;
Query OK, 6 rows affected (0.00 sec)
Records: 6 Duplicates: 0 Warnings: 0
...............
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),start_time from t_partition;
Query OK, 3145728 rows affected (35.68 sec)
Records: 3145728 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),start_time from t_partition;
Query OK, 6291456 rows affected (1 min 11.51 sec)
Records: 6291456 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),start_time from t_partition;
Query OK, 12582912 rows affected (2 min 26.74 sec)
Records: 12582912 Duplicates: 0 Warnings: 0
greatsql> select count(*) from t_partition;
----------
| count(*) |
----------
| 25165824 |
----------
1 row in set (0.50 sec)
greatsql> select count(*) from t_partition partition(P20230310);
----------
| count(*) |
----------
| 25165824 |
----------
向分区插入数据
代码语言：javascript复制insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-11' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-12' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-13' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-14' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-15' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-16' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-17' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-18' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-19' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-20' from t_partition partition(P20230310);
insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-21' from t_partition partition(P20230310);
。。。。。。。。。。。
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-19' from t_partition partition(P20230310);
Query OK, 25165824 rows affected (5 min 17.92 sec)
Records: 25165824 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-20' from t_partition partition(P20230310);
Query OK, 25165824 rows affected (5 min 19.56 sec)
Records: 25165824 Duplicates: 0 Warnings: 0
greatsql> insert into t_partition(ua,start_time) select substring(md5(rand()),1,20),'2023-03-21' from t_partition partition(P20230310);
Query OK, 25165824 rows affected (5 min 27.27 sec)
Records: 25165824 Duplicates: 0 Warnings: 0
更新数据
代码语言：javascript复制greatsql> Update t_partition set ua=concat(ua,'abc') where start_time='2023-03-19';
Query OK, 25165824 rows affected (12 min 55.53 sec)
Rows matched: 25165824 Changed: 25165824 Warnings: 0
删除分区
代码语言：javascript复制greatsql> alter table t_partition drop partition P20230311;
Query OK, 0 rows affected (13.68 sec)
Records: 0 Duplicates: 0 Warnings: 0
greatsql> alter table t_partition drop partition P20230312;
Query OK, 0 rows affected (0.07 sec)
Records: 0 Duplicates: 0 Warnings: 0
两个分区数据量是一样，但删除第一个分区耗时较长。
代码语言：javascript复制$ perf record -ag -p 11222 -o /mysqldb/perf_drop_part_mysql2.data
Warning:
PID/TID switch overriding SYSTEM
^C[ perf record: Woken up 41 times to write data ]
[ perf record: Captured and wrote 10.610 MB /mysqldb/perf_drop_part_mysql2.data (54351 samples) ]
Children Self Command Shared Object Symbol
99.17% 0.00% mysqld libpthread-2.17.so [.] start_thread
99.17% 0.00% mysqld mysqld [.] pfs_spawn_thread
99.17% 0.00% mysqld mysqld [.] handle_connection
99.17% 0.00% mysqld mysqld [.] do_command
99.17% 0.00% mysqld mysqld [.] dispatch_command
99.17% 0.00% mysqld mysqld [.] dispatch_sql_command
99.16% 0.00% mysqld mysqld [.] mysql_execute_command
99.16% 0.00% mysqld mysqld [.] Sql_cmd_alter_table::execute
99.16% 0.00% mysqld mysqld [.] mysql_alter_table
99.09% 0.00% mysqld mysqld [.] mysql_inplace_alter_table
98.56% 0.00% mysqld mysqld [.] ha_innopart::commit_inplace_alter_partition
98.54% 0.00% mysqld mysqld [.] alter_parts::prepare_or_commit_for_new
98.54% 0.00% mysqld mysqld [.] alter_part_normal::try_commit
98.53% 0.00% mysqld mysqld [.] btr_drop_ahi_for_table
98.52% 1.30% mysqld mysqld [.] btr_drop_next_batch
97.20% 0.03% mysqld mysqld [.] btr_search_drop_page_hash_when_freed
96.34% 2.03% mysqld mysqld [.] btr_search_drop_page_hash_index
86.52% 52.25% mysqld mysqld [.] ha_remove_all_nodes_to_page
34.21% 34.15% mysqld mysqld [.] ha_delete_hash_node
4.27% 2.68% mysqld mysqld [.] rec_get_offsets_func
2.11% 2.10% mysqld mysqld [.] ut_fold_binary
1.58% 1.58% mysqld mysqld [.] rec_init_offsets
1.30% 1.30% mysqld mysqld [.] rec_fold
0.57% 0.03% mysqld mysqld [.] buf_page_get_gen
0.57% 0.00% mysqld mysqld [.] execute_native_thread_routine
0.56% 0.01% mysqld [kernel.kallsyms] [k] system_call_fastpath
从系统调用上看有大量的自适应hash相关的调用
重启关闭自适应hash
代码语言：javascript复制greatsql> show variables like '%hash%';
---------------------------------- -------
| Variable_name | Value |
---------------------------------- -------
| innodb_adaptive_hash_index | OFF |
| innodb_adaptive_hash_index_parts | 8 |
---------------------------------- -------
修改配置文件，关闭自适应hash，按照上面的流程从新执行
代码语言：javascript复制greatsql> alter table t_partition drop partition P20230311;
Query OK, 0 rows affected (0.08 sec)
Records: 0 Duplicates: 0 Warnings: 0
greatsql> alter table t_partition drop partition P20230312;
Query OK, 0 rows affected (0.08 sec)
Records: 0 Duplicates: 0 Warnings: 0
关闭自适应hash后，相同的操作过程，删除第一个分区的时间明显变短，删除每个分区的时间基本上一致。
备注：innodb_adaptive_hash_index是全局变量，可以动态修改，不重启数据库。
测试结果汇总
自适应hash对比 | 第一次删分区 | 第二次删分区 |
|---|---|---|
innodb_buffer_pool_instances=8& innodb_adaptive_hash_index=on | 13.68 | 0.07 |
innodb_buffer_pool_instances=8& innodb_adaptive_hash_index=off | 0.08 | 0.08 |
源码分析
代码语言：javascript复制// btr_drop_ahi_for_table
void btr_drop_ahi_for_table(dict_table_t *table) {
const ulint len = UT_LIST_GET_LEN(table->indexes);
if (len == 0) {
return;
}
const dict_index_t *indexes[MAX_INDEXES];
const page_size_t page_size(dict_table_page_size(table));
for (;;) {
ulint ref_count = 0;
const dict_index_t **end = indexes;
for (dict_index_t *index = table->first_index(); index != nullptr;
index = index->next()) {
if (ulint n_refs = index->search_info->ref_count) {
ut_ad(!index->disable_ahi);
ut_ad(index->is_committed());
ref_count = n_refs;
ut_ad(indexes len > end);
*end = index;
}
}
ut_ad((indexes == end) == (ref_count == 0));
if (ref_count == 0) {
return;
}
btr_drop_next_batch(page_size, indexes, end); // breakpoint
std::this_thread::yield();
}
}
// btr_drop_next_batch
static void btr_drop_next_batch(const page_size_t &page_size,
const dict_index_t **first,
const dict_index_t **last) {
static constexpr unsigned batch_size = 1024;
std::vector<page_id_t> to_drop;
to_drop.reserve(batch_size);
for (ulint i = 0; i < srv_buf_pool_instances; i) {
to_drop.clear();
buf_pool_t *buf_pool = buf_pool_from_array(i);
mutex_enter(&buf_pool->LRU_list_mutex);
const buf_page_t *prev;
for (const buf_page_t *bpage = UT_LIST_GET_LAST(buf_pool->LRU);
bpage != nullptr; bpage = prev) {
prev = UT_LIST_GET_PREV(LRU, bpage);
ut_a(buf_page_in_file(bpage));
if (buf_page_get_state(bpage) != BUF_BLOCK_FILE_PAGE ||
bpage->buf_fix_count > 0) {
continue;
}
const dict_index_t *block_index =
reinterpret_cast<const buf_block_t *>(bpage)->ahi.index;
/* index == nullptr means the page is no longer in AHI, so no need to
attempt freeing it */
if (block_index == nullptr) {
continue;
}
/* pages IO fixed for read have index == nullptr */
ut_ad(!bpage->was_io_fix_read());
if (std::find(first, last, block_index) != last) {
to_drop.emplace_back(bpage->id);
if (to_drop.size() == batch_size) { // batch_size = 1024
break;
}
}
}
mutex_exit(&buf_pool->LRU_list_mutex);
for (const page_id_t &page_id : to_drop) {
btr_search_drop_page_hash_when_freed(page_id, page_size); // breakpoint
}
}
}
// btr_search_drop_page_hash_when_freed
void btr_search_drop_page_hash_when_freed(const page_id_t &page_id,
const page_size_t &page_size) {
buf_block_t *block;
mtr_t mtr;
ut_d(export_vars.innodb_ahi_drop_lookups );
mtr_start(&mtr);
/* If the caller has a latch on the page, then the caller must
have a x-latch on the page and it must have already dropped
the hash index for the page. Because of the x-latch that we
are possibly holding, we cannot s-latch the page, but must
(recursively) x-latch it, even though we are only reading. */
block = buf_page_get_gen(page_id, page_size, RW_X_LATCH, nullptr,
Page_fetch::PEEK_IF_IN_POOL, UT_LOCATION_HERE, &mtr);
if (block) {
/* If AHI is still valid, page can't be in free state.
AHI is dropped when page is freed. */
ut_ad(!block->page.file_page_was_freed);
buf_block_dbg_add_level(block, SYNC_TREE_NODE_FROM_HASH);
dict_index_t *index = block->ahi.index;
if (index != nullptr) {
/* In all our callers, the table handle should
be open, or we should be in the process of
dropping the table (preventing eviction). */
ut_ad(index->table->n_ref_count > 0 || dict_sys_mutex_own());
btr_search_drop_page_hash_index(block);
}
}
mtr_commit(&mtr);
}
函数逻辑说明：
代码语言：javascript复制drop 分区和add分区都会清空所有分区的AHI信息，最耗时的如下
循环每个分区调用函数
->btr_drop_ahi_for_table
循环表（或者分区）中的每个索引，如果索引都没有用到AHI，
则退出
循环innodb buffer中的每个实例，根据LRU链表循环每个page
如果page建立了AHI信息，且是要删除表（或者分区）的相关索引
则放入drop vector容器中
如果page没有建立AHI信息
则跳过
如果drop verctor容器中填满1024个page
则清理一次，循环每个page，调用函数
->btr_search_drop_page_hash_index
计算page所在AHI结构的slot信息，以便找到对应的hash_table_t结构
循环page中所有的行
循环行中访问到的索引字段（访问模式），计算出fold信息填入到fold[]数组中
本循环中会通过函数rec_get_offsets进行字段偏移量的获取，为耗用CPU的函数
循环fold[]数组，一个fold代表一行数据,调用函数
->ha_remove_all_nodes_to_page,为耗用CPU的函数
->ha_chain_get_fist
根据fold信息找到hash结构的cell
循环本cell中的链表信息
如果行的地址在本要删除的page上，调用函数
->ha_delete_hash_node，为消耗CPU的函数
进行链表和hash结构的维护
每次处理完1024个page后，yeild线程主动放弃CPU，避免长期占用CPU，醒来后继续处理
drop 分区和add分区都会循环每个分区调用函数btr_drop_ahi_for_table 、btr_search_drop_page_hash_index清空所有分区及索引的的AHI信息， 删除第1个分区的时ahi信息被清空， 删第2个分区的时候buffer中已经没有ahi信息了，所有删除第2个分区就很快了。
避免方式
针对以上原因，线上执行 drop必须遵守以下原则：
1、关闭AHI功能，不使用AHI带来的查询加速功能，需要先在测试环境进行业务测试，确保业务性能不受影响。
2、删除表的第一个分区时，内部会清理该表在每个buffer pool实例中对应的数据块页面，耗时较久，接着删其他分区耗时很小，建议将每天一次的删除分区的操作改为每周或者每月批量执行删除分区的操作，并且需要在业务低峰期操作。
