29.9 Performance Schema Tables for Current and Historical Events

For wait, stage, statement, and transaction events, the Performance Schema can monitor and store current events. In addition, when events end, the Performance Schema can store them in history tables. For each event type, the Performance Schema uses three tables for storing current and historical events. The tables have names of the following forms, where xxx indicates the event type (waits, stages, statements, transactions):

events_xxx_current: The “current events” table stores the current monitored event for each thread (one row per thread).
events_xxx_history: The “recent history” table stores the most recent events that have ended per thread (up to a maximum number of rows per thread).
events_xxx_history_long: The “long history” table stores the most recent events that have ended globally (across all threads, up to a maximum number of rows per table).

The _current table for each event type contains one row per thread, so there is no system variable for configuring its maximum size. The Performance Schema autosizes the history tables, or the sizes can be configured explicitly at server startup using table-specific system variables, as indicated in the sections that describe the individual history tables. Typical autosized values are 10 rows per thread for _history tables, and 10,000 rows total for _history_long tables.

For each event type, the _current, _history, and _history_long tables have the same columns. The _current and _history tables have the same indexing. The _history_long table has no indexing.

The _current tables show what is currently happening within the server. When a current event ends, it is removed from its _current table.

The _history and _history_long tables show what has happened in the recent past. When the history tables become full, old events are discarded as new events are added. Rows expire from the _history and _history_long tables in different ways because the tables serve different purposes:

_history is meant to investigate individual threads, independently of the global server load.
_history_long is meant to investigate the server globally, not each thread.

The difference between the two types of history tables relates to the data retention policy. Both tables contains the same data when an event is first seen. However, data within each table expires differently over time, so that data might be preserved for a longer or shorter time in each table:

For _history, when the table contains the maximum number of rows for a given thread, the oldest thread row is discarded when a new row for that thread is added.
For _history_long, when the table becomes full, the oldest row is discarded when a new row is added, regardless of which thread generated either row.

When a thread ends, all its rows are discarded from the _history table but not from the _history_long table.

The following example illustrates the differences in how events are added to and discarded from the two types of history tables. The principles apply equally to all event types. The example is based on these assumptions:

The Performance Schema is configured to retain 10 rows per thread in the _history table and 10,000 rows total in the _history_long table.
Thread A generates 1 event per second.
Thread B generates 100 events per second.
No other threads are running.

After 5 seconds of execution:

A and B have generated 5 and 500 events, respectively.
_history contains 5 rows for A and 10 rows for B. Because storage per thread is limited to 10 rows, no rows have been discarded for A, whereas 490 rows have been discarded for B.
_history_long contains 5 rows for A and 500 rows for B. Because the table has a maximum size of 10,000 rows, no rows have been discarded for either thread.

After 5 minutes (300 seconds) of execution:

A and B have generated 300 and 30,000 events, respectively.
_history contains 10 rows for A and 10 rows for B. Because storage per thread is limited to 10 rows, 290 rows have been discarded for A, whereas 29,990 rows have been discarded for B. Rows for A include data up to 10 seconds old, whereas rows for B include data up to only .1 seconds old.
_history_long contains 10,000 rows. Because A and B together generate 101 events per second, the table contains data up to approximately 10,000/101 = 99 seconds old, with a mix of rows approximately 100 to 1 from B as opposed to A.