Continuation of the article "«.
The article will explore and illustrate with specific queries and examples the useful information that can be obtained using the pg_locks view.
Warning.
Given the novelty of the topic and the ongoing testing period, the article may contain errors. Critique and feedback are always welcomed and expected.
Input data
The pg_locks View History
archive_locking
CREATE TABLE archive_locking
( timepoint timestamp without time zone ,
locktype text ,
relation oid ,
mode text ,
tid xid ,
vtid text ,
pid integer ,
blocking_pids integer[] ,
granted boolean ,
queryid bigint
);Essentially, the table is similar to the one archive_pg_stat_activity, described in more detail here — and here —
To populate the column queryid the function is used
update_history_locking_by_queryid
--update_history_locking_by_queryid.sql
CREATE OR REPLACE FUNCTION update_history_locking_by_queryid() RETURNS boolean AS $$
DECLARE
result boolean ;
current_minute double precision ;
start_minute integer ;
finish_minute integer ;
start_period timestamp without time zone ;
finish_period timestamp without time zone ;
lock_rec record ;
endpoint_rec record ;
current_hour_diff double precision ;
BEGIN
RAISE NOTICE '***update_history_locking_by_queryid';
result = TRUE ;
current_minute = extract ( minute from now() );
SELECT * FROM endpoint WHERE is_need_monitoring
INTO endpoint_rec ;
current_hour_diff = endpoint_rec.hour_diff ;
IF current_minute < 5
THEN
RAISE NOTICE 'Current time is less than 5 minutes.';
start_period = date_trunc('hour',now()) + (current_hour_diff * interval '1 hour');
finish_period = start_period - interval '5 minutes' ;
ELSE
finish_minute = extract ( minute from now() ) / 5 ;
start_minute = finish_minute - 1 ;
start_period = date_trunc('hour',now()) + interval '1 minute'*start_minute*5+(current_hour_diff * interval '1 hour');
finish_period = date_trunc('hour',now()) + interval '1 minute'*finish_minute*5+(current_hour_diff * interval '1 hour') ;
END IF ;
RAISE NOTICE 'start_period = %', start_period;
RAISE NOTICE 'finish_period = %', finish_period;
FOR lock_rec IN
WITH act_queryid AS
(
SELECT
pid ,
timepoint ,
query_start AS started ,
MAX(timepoint) OVER (PARTITION BY pid , query_start ) AS finished ,
queryid
FROM
activity_hist.history_pg_stat_activity
WHERE
timepoint BETWEEN start_period and
finish_period
GROUP BY
pid ,
timepoint ,
query_start ,
queryid
),
lock_pids AS
(
SELECT
hl.pid ,
hl.locktype ,
hl.mode ,
hl.timepoint ,
MIN ( timepoint ) OVER (PARTITION BY pid , locktype ,mode ) as started
FROM
activity_hist.history_locking hl
WHERE
hl.timepoint between start_period and
finish_period
GROUP BY
hl.pid ,
hl.locktype ,
hl.mode ,
hl.timepoint
)
SELECT
lp.pid ,
lp.locktype ,
lp.mode ,
lp.timepoint ,
aq.queryid
FROM lock_pids lp LEFT OUTER JOIN act_queryid aq ON ( lp.pid = aq.pid AND lp.started BETWEEN aq.started AND aq.finished )
WHERE aq.queryid IS NOT NULL
GROUP BY
lp.pid ,
lp.locktype ,
lp.mode ,
lp.timepoint ,
aq.queryid
LOOP
UPDATE activity_hist.history_locking SET queryid = lock_rec.queryid
WHERE pid = lock_rec.pid AND locktype = lock_rec.locktype AND mode = lock_rec.mode AND timepoint = lock_rec.timepoint ;
END LOOP;
RETURN result ;
END
$$ LANGUAGE plpgsql;Explanation: The value of the queryid column is updated in the history_locking table, and then when a new section is created for the archive_locking table, the value will be preserved in historical values.
Output data
General information on processes as a whole.
WAITING FOR LOCKS BY LOCKTYPES
Query
WITH
t AS
(
SELECT
locktype ,
mode ,
count(*) as total
FROM
activity_hist.archive_locking
WHERE
timepoint between pg_stat_history_begin+(current_hour_diff * interval '1 hour') AND pg_stat_history_end+(current_hour_diff * interval '1 hour') AND
NOT granted
GROUP BY
locktype ,
mode
)
SELECT
locktype ,
mode ,
total * interval '1 second' as duration
FROM t
ORDER BY 3 DESC Example
| WAITING FOR LOCKS BY LOCKTYPES +--------------------+------------------------------+-------------------- | locktype| mode| duration +--------------------+------------------------------+-------------------- | transactionid| ShareLock| 19:39:26 | tuple| AccessExclusiveLock| 00:03:35 +--------------------+------------------------------+--------------------
TAKINGS OF LOCKS BY LOCKTYPES
Query
WITH
t AS
(
SELECT
locktype ,
mode ,
count(*) as total
FROM
activity_hist.archive_locking
WHERE
timepoint between pg_stat_history_begin+(current_hour_diff * interval '1 hour') AND pg_stat_history_end+(current_hour_diff * interval '1 hour') AND
granted
GROUP BY
locktype ,
mode
)
SELECT
locktype ,
mode ,
total * interval '1 second' as duration
FROM t
ORDER BY 3 DESC Example
| TAKINGS OF LOCKS BY LOCKTYPES +--------------------+------------------------------+-------------------- | locktype| mode| duration +--------------------+------------------------------+-------------------- | relation| RowExclusiveLock| 51:11:10 | virtualxid| ExclusiveLock| 48:10:43 | transactionid| ExclusiveLock| 44:24:53 | relation| AccessShareLock| 20:06:13 | tuple| AccessExclusiveLock| 17:58:47 | tuple| ExclusiveLock| 01:40:41 | relation| ShareUpdateExclusiveLock| 00:26:41 | object| RowExclusiveLock| 00:00:01 | transactionid| ShareLock| 00:00:01 | extend| ExclusiveLock| 00:00:01 +--------------------+------------------------------+--------------------
Detailed information on specific queryid requests
WAITING FOR LOCKS BY LOCKTYPES BY QUERYID
Query
WITH
lt AS
(
SELECT
pid ,
locktype ,
mode ,
timepoint ,
queryid ,
blocking_pids ,
MIN ( timepoint ) OVER (PARTITION BY pid , locktype ,mode ) as started
FROM
activity_hist.archive_locking
WHERE
timepoint between pg_stat_history_begin+(current_hour_diff * interval '1 hour') AND
pg_stat_history_end+(current_hour_diff * interval '1 hour') AND
NOT granted AND
queryid IS NOT NULL
GROUP BY
pid ,
locktype ,
mode ,
timepoint ,
queryid ,
blocking_pids
)
SELECT
lt.pid ,
lt.locktype ,
lt.mode ,
lt.started ,
lt.queryid ,
lt.blocking_pids ,
COUNT(*) * interval '1 second' as duration
FROM lt
GROUP BY
lt.pid ,
lt.locktype ,
lt.mode ,
lt.started ,
lt.queryid ,
lt.blocking_pids
ORDER BY 4Example
| WAITING FOR LOCKS BY LOCK TYPES BY QUERY ID
+----------+-------------------------+--------------------+------------------------------+--------------------+--------------------+--------------------
| pid| lock type| mode| started| query id| blocking pids| duration
+----------+-------------------------+--------------------+------------------------------+--------------------+--------------------+--------------------
| 11288| transaction id| Share Lock| 2019-09-17 10:00:00.302936| 389015618226997618| {11092}| 00:03:34
| 11626| transaction id| Share Lock| 2019-09-17 10:00:21.380921| 389015618226997618| {12380}| 00:00:29
| 11626| transaction id| Share Lock| 2019-09-17 10:00:21.380921| 389015618226997618| {11092}| 00:03:25
| 11626| transaction id| Share Lock| 2019-09-17 10:00:21.380921| 389015618226997618| {12213}| 00:01:55
| 11626| transaction id| Share Lock| 2019-09-17 10:00:21.380921| 389015618226997618| {12751}| 00:00:01
| 11629| transaction id| Share Lock| 2019-09-17 10:00:24.331935| 389015618226997618| {11092}| 00:03:22
| 11629| transaction id| Share Lock| 2019-09-17 10:00:24.331935| 389015618226997618| {12007}| 00:00:01
| 12007| transaction id| Share Lock| 2019-09-17 10:05:03.327933| 389015618226997618| {11629}| 00:00:13
| 12007| transaction id| Share Lock| 2019-09-17 10:05:03.327933| 389015618226997618| {11092}| 00:01:10
| 12007| transaction id| Share Lock| 2019-09-17 10:05:03.327933| 389015618226997618| {11288}| 00:00:05
| 12213| transaction id| Share Lock| 2019-09-17 10:06:07.328019| 389015618226997618| {12007}| 00:00:10TAKING LOCKS BY LOCK TYPES BY QUERY ID
Query
WITH
lt AS
(
SELECT
pid ,
lock type ,
mode ,
time point ,
query id ,
blocking pids ,
MIN ( time point ) OVER (PARTITION BY pid , lock type ,mode ) as started
FROM
activity_hist.archive_locking
WHERE
time point between pg_stat_history_begin+(current_hour_diff * interval '1 hour') AND
pg_stat_history_end+(current_hour_diff * interval '1 hour') AND
granted AND
query id IS NOT NULL
GROUP BY
pid ,
lock type ,
mode ,
time point ,
query id ,
blocking pids
)
SELECT
lt.pid ,
lt.lock type ,
lt.mode ,
lt.started ,
lt.query id ,
lt.blocking pids ,
COUNT(*) * interval '1 second' as duration
FROM lt
GROUP BY
lt.pid ,
lt.lock type ,
lt.mode ,
lt.started ,
lt.query id ,
lt.blocking pids
ORDER BY 4Example
| TAKING LOCKS BY LOCKTYPES BY QUERYID
+----------+-------------------------+--------------------+------------------------------+--------------------+--------------------+--------------------
| pid| locktype| mode| started| queryid| blocking_pids| duration
+----------+-------------------------+--------------------+------------------------------+--------------------+--------------------+--------------------
| 11288| relation| RowExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {11092}| 00:03:34
| 11092| transactionid| ExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {}| 00:03:34
| 11288| relation| RowExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {}| 00:00:10
| 11092| relation| RowExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {}| 00:03:34
| 11092| virtualxid| ExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {}| 00:03:34
| 11288| virtualxid| ExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {11092}| 00:03:34
| 11288| transactionid| ExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {11092}| 00:03:34
| 11288| tuple| AccessExclusiveLock| 2019-09-17 10:00:00.302936| 389015618226997618| {11092}| 00:03:34Using lock history for performance incident analysis.
- The request with queryid=389015618226997618 executed by the process with pid=11288 has been waiting for a lock since 2019-09-17 10:00:00 for 3 minutes.
- The lock was held by the process with pid=11092.
- The process with pid=11092, executing the request with queryid=389015618226997618 since 2019-09-17 10:00:00, held the lock for 3 minutes.
Summary
Now, hopefully, the most interesting and useful part begins — collecting statistics and analyzing cases based on the history of waits and locks.
In the future, I hope to create a set of notes (similar to Oracle's metalink).
Overall, this is precisely why the methodology we use is quickly made available for public awareness.
I will try to upload the project to GitHub very soon.
Source: habr.com
