Metrics Schema

METRICS_SCHEMA 是基于 Prometheus 中 TiDB 监控指标的一组视图。每个表的 PromQL(Prometheus 查询语言)的源均可在 INFORMATION_SCHEMA.METRICS_TABLES 表中找到。

use metrics_schema; SELECT * FROM uptime; SELECT * FROM information_schema.metrics_tables WHERE table_name='uptime'\G
+----------------------------+-----------------+------------+--------------------+ | time | instance | job | value | +----------------------------+-----------------+------------+--------------------+ | 2020-07-06 15:26:26.203000 | 127.0.0.1:10080 | tidb | 123.60300016403198 | | 2020-07-06 15:27:26.203000 | 127.0.0.1:10080 | tidb | 183.60300016403198 | | 2020-07-06 15:26:26.203000 | 127.0.0.1:20180 | tikv | 123.60300016403198 | | 2020-07-06 15:27:26.203000 | 127.0.0.1:20180 | tikv | 183.60300016403198 | | 2020-07-06 15:26:26.203000 | 127.0.0.1:2379 | pd | 123.60300016403198 | | 2020-07-06 15:27:26.203000 | 127.0.0.1:2379 | pd | 183.60300016403198 | | 2020-07-06 15:26:26.203000 | 127.0.0.1:9090 | prometheus | 123.72300004959106 | | 2020-07-06 15:27:26.203000 | 127.0.0.1:9090 | prometheus | 183.72300004959106 | +----------------------------+-----------------+------------+--------------------+ 8 rows in set (0.00 sec) *************************** 1. row *************************** TABLE_NAME: uptime PROMQL: (time() - process_start_time_seconds{$LABEL_CONDITIONS}) LABELS: instance,job QUANTILE: 0 COMMENT: TiDB uptime since last restart(second) 1 row in set (0.00 sec)
show tables;
+---------------------------------------------------+ | Tables_in_metrics_schema | +---------------------------------------------------+ | abnormal_stores | | etcd_disk_wal_fsync_rate | | etcd_wal_fsync_duration | | etcd_wal_fsync_total_count | | etcd_wal_fsync_total_time | | go_gc_count | | go_gc_cpu_usage | | go_gc_duration | | go_heap_mem_usage | | go_threads | | goroutines_count | | node_cpu_usage | | node_disk_available_size | | node_disk_io_util | | node_disk_iops | | node_disk_read_latency | | node_disk_size | .. | tikv_storage_async_request_total_time | | tikv_storage_async_requests | | tikv_storage_async_requests_total_count | | tikv_storage_command_ops | | tikv_store_size | | tikv_thread_cpu | | tikv_thread_nonvoluntary_context_switches | | tikv_thread_voluntary_context_switches | | tikv_threads_io | | tikv_threads_state | | tikv_total_keys | | tikv_wal_sync_duration | | tikv_wal_sync_max_duration | | tikv_worker_handled_tasks | | tikv_worker_handled_tasks_total_num | | tikv_worker_pending_tasks | | tikv_worker_pending_tasks_total_num | | tikv_write_stall_avg_duration | | tikv_write_stall_max_duration | | tikv_write_stall_reason | | up | | uptime | +---------------------------------------------------+ 626 rows in set (0.00 sec)

METRICS_SCHEMA 是监控相关的 summary 表的数据源,例如 metrics_summarymetrics_summary_by_labelinspection_summary

更多例子

下面以 metrics_schema 中的 tidb_query_duration 监控表为例,介绍监控表相关的使用和原理,其他的监控表原理均类似。

查询 information_schema.metrics_tables 中关于 tidb_query_duration 表相关的信息如下:

select * from information_schema.metrics_tables where table_name='tidb_query_duration';
+---------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------+-------------------+----------+----------------------------------------------+ | TABLE_NAME | PROMQL | LABELS | QUANTILE | COMMENT | +---------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------+-------------------+----------+----------------------------------------------+ | tidb_query_duration | histogram_quantile($QUANTILE, sum(rate(tidb_server_handle_query_duration_seconds_bucket{$LABEL_CONDITIONS}[$RANGE_DURATION])) by (le,sql_type,instance)) | instance,sql_type | 0.9 | The quantile of TiDB query durations(second) | +---------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------+-------------------+----------+----------------------------------------------+
  • TABLE_NAME:对应于 metrics_schema 中的表名,这里表名是 tidb_query_duration
  • PROMQL:因为监控表的原理是将 SQL 映射成 PromQL 后向 Prometheus 请求数据,并将 Prometheus 返回的结果转换成 SQL 查询结果。该字段是 PromQL 的表达式模板,查询监控表数据时使用查询条件改写模板中的变量,生成最终的查询表达式。
  • LABELS:监控项定义的 label,tidb_query_duration 有两个 label,分别是 instancesql_type
  • QUANTILE:百分位。直方图类型的监控数据会指定一个默认百分位。如果值为 0,表示该监控表对应的监控不是直方图。tidb_query_duration 默认查询 0.9 ,也就是 P90 的监控值。
  • COMMENT:对这个监控表的解释。可以看出 tidb_query_duration 表是用来查询 TiDB query 执行的百分位时间,如 P999/P99/P90 的查询耗时,单位是秒。

再来看 tidb_query_duration 的表结构:

show create table metrics_schema.tidb_query_duration;
+---------------------+--------------------------------------------------------------------------------------------------------------------+ | Table | Create Table | +---------------------+--------------------------------------------------------------------------------------------------------------------+ | tidb_query_duration | CREATE TABLE `tidb_query_duration` ( | | | `time` datetime unsigned DEFAULT CURRENT_TIMESTAMP, | | | `instance` varchar(512) DEFAULT NULL, | | | `sql_type` varchar(512) DEFAULT NULL, | | | `quantile` double unsigned DEFAULT '0.9', | | | `value` double unsigned DEFAULT NULL | | | ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_bin COMMENT='The quantile of TiDB query durations(second)' | +---------------------+--------------------------------------------------------------------------------------------------------------------+
  • time:监控项的时间。
  • instancesql_type:是 tidb_query_duration 这个监控项的 label。instance 表示监控的地址,sql_type 表示执行 SQL 的类似。
  • quantile,百分位,直方图类型的监控都会有该列,表示查询的百分位时间,如 quantile=0.9 就是查询 P90 的时间。
  • value:监控项的值。

下面是查询时间 [2020-03-25 23:40:00, 2020-03-25 23:42:00] 范围内的 P99 的 TiDB Query 耗时:

select * from metrics_schema.tidb_query_duration where value is not null and time>='2020-03-25 23:40:00' and time <= '2020-03-25 23:42:00' and quantile=0.99;
+---------------------+-------------------+----------+----------+----------------+ | time | instance | sql_type | quantile | value | +---------------------+-------------------+----------+----------+----------------+ | 2020-03-25 23:40:00 | 172.16.5.40:10089 | Insert | 0.99 | 0.509929485256 | | 2020-03-25 23:41:00 | 172.16.5.40:10089 | Insert | 0.99 | 0.494690793986 | | 2020-03-25 23:42:00 | 172.16.5.40:10089 | Insert | 0.99 | 0.493460506934 | | 2020-03-25 23:40:00 | 172.16.5.40:10089 | Select | 0.99 | 0.152058493415 | | 2020-03-25 23:41:00 | 172.16.5.40:10089 | Select | 0.99 | 0.152193879678 | | 2020-03-25 23:42:00 | 172.16.5.40:10089 | Select | 0.99 | 0.140498483232 | | 2020-03-25 23:40:00 | 172.16.5.40:10089 | internal | 0.99 | 0.47104 | | 2020-03-25 23:41:00 | 172.16.5.40:10089 | internal | 0.99 | 0.11776 | | 2020-03-25 23:42:00 | 172.16.5.40:10089 | internal | 0.99 | 0.11776 | +---------------------+-------------------+----------+----------+----------------+

以上查询结果的第一行意思是,在 2020-03-25 23:40:00 时,在 TiDB 实例 172.16.5.40:10089 上,Insert 类型的语句的 P99 执行时间是 0.509929485256 秒。其他各行的含义类似,sql_type 列的其他值含义如下:

  • Select:表示执行的 select 类型的语句。
  • internal:表示 TiDB 的内部 SQL 语句,一般是统计信息更新,获取全局变量相关的内部语句。

进一步再查看上面语句的执行计划如下:

desc select * from metrics_schema.tidb_query_duration where value is not null and time>='2020-03-25 23:40:00' and time <= '2020-03-25 23:42:00' and quantile=0.99;
+------------------+----------+------+---------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +------------------+----------+------+---------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Selection_5 | 8000.00 | root | | not(isnull(Column#5)) | | └─MemTableScan_6 | 10000.00 | root | table:tidb_query_duration | PromQL:histogram_quantile(0.99, sum(rate(tidb_server_handle_query_duration_seconds_bucket{}[60s])) by (le,sql_type,instance)), start_time:2020-03-25 23:40:00, end_time:2020-03-25 23:42:00, step:1m0s | +------------------+----------+------+---------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

可以发现执行计划中有一个 PromQL, 以及查询监控的 start_timeend_time,还有 step 值,在实际执行时,TiDB 会调用 Prometheus 的 query_range HTTP API 接口来查询监控数据。

从以上结果可知,在 [2020-03-25 23:40:00, 2020-03-25 23:42:00] 时间范围内,每个 label 只有三个时间的值,间隔时间是 1 分钟,即执行计划中的 step 值。该间隔时间由以下两个 session 变量决定:

  • tidb_metric_query_step:查询的分辨率步长。从 Prometheus 的 query_range 接口查询数据时需要指定 start_timeend_timestep,其中 step 会使用该变量的值。
  • tidb_metric_query_range_duration:查询监控时,会将 PROMQL 中的 $RANGE_DURATION 替换成该变量的值,默认值是 60 秒。

如果想要查看不同时间粒度的监控项的值,用户可以修改上面两个 session 变量后查询监控表,示例如下:

首先修改两个 session 变量的值,将时间粒度设置为 30 秒。

set @@tidb_metric_query_step=30; set @@tidb_metric_query_range_duration=30;

再查询 tidb_query_duration 监控如下,可以发现在三分钟时间范围内,每个 label 有六个时间的值,每个值时间间隔是 30 秒。

select * from metrics_schema.tidb_query_duration where value is not null and time>='2020-03-25 23:40:00' and time <= '2020-03-25 23:42:00' and quantile=0.99;
+---------------------+-------------------+----------+----------+-----------------+ | time | instance | sql_type | quantile | value | +---------------------+-------------------+----------+----------+-----------------+ | 2020-03-25 23:40:00 | 172.16.5.40:10089 | Insert | 0.99 | 0.483285651924 | | 2020-03-25 23:40:30 | 172.16.5.40:10089 | Insert | 0.99 | 0.484151462113 | | 2020-03-25 23:41:00 | 172.16.5.40:10089 | Insert | 0.99 | 0.504576 | | 2020-03-25 23:41:30 | 172.16.5.40:10089 | Insert | 0.99 | 0.493577384561 | | 2020-03-25 23:42:00 | 172.16.5.40:10089 | Insert | 0.99 | 0.49482474311 | | 2020-03-25 23:40:00 | 172.16.5.40:10089 | Select | 0.99 | 0.189253402185 | | 2020-03-25 23:40:30 | 172.16.5.40:10089 | Select | 0.99 | 0.184224951851 | | 2020-03-25 23:41:00 | 172.16.5.40:10089 | Select | 0.99 | 0.151673410553 | | 2020-03-25 23:41:30 | 172.16.5.40:10089 | Select | 0.99 | 0.127953838989 | | 2020-03-25 23:42:00 | 172.16.5.40:10089 | Select | 0.99 | 0.127455434547 | | 2020-03-25 23:40:00 | 172.16.5.40:10089 | internal | 0.99 | 0.0624 | | 2020-03-25 23:40:30 | 172.16.5.40:10089 | internal | 0.99 | 0.12416 | | 2020-03-25 23:41:00 | 172.16.5.40:10089 | internal | 0.99 | 0.0304 | | 2020-03-25 23:41:30 | 172.16.5.40:10089 | internal | 0.99 | 0.06272 | | 2020-03-25 23:42:00 | 172.16.5.40:10089 | internal | 0.99 | 0.0629333333333 | +---------------------+-------------------+----------+----------+-----------------+

最后查看执行计划,也会发现执行计划中的 PromQL 以及 step 的值都已经变成了 30 秒。

desc select * from metrics_schema.tidb_query_duration where value is not null and time>='2020-03-25 23:40:00' and time <= '2020-03-25 23:42:00' and quantile=0.99;
+------------------+----------+------+---------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +------------------+----------+------+---------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Selection_5 | 8000.00 | root | | not(isnull(Column#5)) | | └─MemTableScan_6 | 10000.00 | root | table:tidb_query_duration | PromQL:histogram_quantile(0.99, sum(rate(tidb_server_handle_query_duration_seconds_bucket{}[30s])) by (le,sql_type,instance)), start_time:2020-03-25 23:40:00, end_time:2020-03-25 23:42:00, step:30s | +------------------+----------+------+---------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+