Or a bit of applied Tetrisology.
Everything new is well forgotten old.
Epigraphs.
Task Definition
It is necessary to periodically download the current PostgreSQL log file from AWS cloud to a local Linux host. Not in real time, but let's say with a slight delay.
The log file update download interval is 5 minutes.
The log file in AWS is rotated every hour.
Tools used
A bash script is used to download the log file to the host, calling the AWS API “».
Parameters:
- —db-instance-identifier: Instance name in AWS;
- —log-file-name: name of the currently generated log file
- —max-item: Total number of items returned in the command output.Size of the downloaded file portion.
- —starting-token: Starting portion label
In this specific case, the task of downloading logs arose during work on
Moreover, it is an interesting task for training and variety during work hours.
I assume that this task has already been solved due to its common nature. However, a quick Google search did not provide any solutions, and there wasn’t much desire to delve deeper. In any case, it's a good exercise.
Task formalization
The final log file consists of a multitude of variable-length lines. Graphically, the log file can be represented approximately as follows:
Already looks somewhat familiar? What does Tetris have to do with it? Well, here's the connection.
If we graphically represent the possible scenarios that arise when loading the next file (for simplicity, let's assume the lines have the same length), we will get standard Tetris shapes:
1) The file is fully downloaded and is final. The portion size is larger than the final file size:
2) The file has a continuation. The portion size is smaller than the final file size:
3) The file is a continuation of the previous file and has a continuation. The portion size is smaller than the remainder of the final file:
4) The file is a continuation of the previous file and is final. The portion size is larger than the remainder of the final file:
The task is to assemble a rectangle or play Tetris at a new level.
Problems arising while solving the task
1) Stitching a line from 2 portions
In general, no particular problems arose. A standard task from the introductory programming course.
Optimal portion size
Now that's a bit more interesting.
Unfortunately, it is not possible to use an offset after the initial token label:
As you already know, the option —starting-token is used to specify where to start paginating. This option takes String values, which means that if you try to add an offset value in front of the Next Token string, the option will not be considered as an offset.
And so, we have to read in chunks.
If reading in large portions, the number of reads will be minimal, but the volume will be maximal.
If reading in small portions, the number of reads will be maximal, but the volume will be minimal.
Therefore, to reduce traffic and for the overall elegance of the solution, we had to come up with some sort of workaround, which unfortunately resembles a bit of a hack.
To illustrate, let’s consider the process of loading a log file in 2 heavily simplified scenarios. The number of reads in both cases depends on the size of the chunk.
1) Loading in small portions:
2) Loading in large portions:
As usual, the optimal solution is in the middle..
The portion size is minimal, but during the reading process, the size can be increased to reduce the number of reads.
It should be noted, that the complete task of finding the optimal size of the read portion has not yet been solved and requires further in-depth work and analysis. Maybe later.
General description of the implementation
Service tables used
CREATE TABLE endpoint
(
id SERIAL ,
host text
);
TABLE database
(
id SERIAL ,
…
last_aws_log_time text ,
last_aws_nexttoken text ,
aws_max_item_size integer
);
last_aws_log_time — timestamp of the last loaded log file in the format YYYY-MM-DD-HH24.
last_aws_nexttoken — text label of the last loaded portion.
aws_max_item_size - empirically determined initial size of the portion.
Full script text
download_aws_piece.sh
#!/bin/bash
#########################################################
# download_aws_piece.sh
# downloan piece of log from AWS
# version HABR
let min_item_size=1024
let max_item_size=1048576
let growth_factor=3
let growth_counter=1
let growth_counter_max=3
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:''STARTED'
AWS_LOG_TIME=$1
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:AWS_LOG_TIME='$AWS_LOG_TIME
database_id=$2
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:database_id='$database_id
RESULT_FILE=$3
endpoint=`psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE_DATABASE -A -t -c "select e.host from endpoint e join database d on e.id = d.endpoint_id where d.id = $database_id "`
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:endpoint='$endpoint
db_instance=`echo $endpoint | awk -F"." '{print toupper($1)}'`
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:db_instance='$db_instance
LOG_FILE=$RESULT_FILE'.tmp_log'
TMP_FILE=$LOG_FILE'.tmp'
TMP_MIDDLE=$LOG_FILE'.tmp_mid'
TMP_MIDDLE2=$LOG_FILE'.tmp_mid2'
current_aws_log_time=`psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -A -t -c "select last_aws_log_time from database where id = $database_id "`
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:current_aws_log_time='$current_aws_log_time
if [[ $current_aws_log_time != $AWS_LOG_TIME ]];
then
is_new_log='1'
if ! psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -v ON_ERROR_STOP=1 -A -t -q -c "update database set last_aws_log_time = '$AWS_LOG_TIME' where id = $database_id "
then
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: FATAL_ERROR - update database set last_aws_log_time .'
exit 1
fi
else
is_new_log='0'
fi
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh:is_new_log='$is_new_log
let last_aws_max_item_size=`psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -A -t -c "select aws_max_item_size from database where id = $database_id "`
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: last_aws_max_item_size='$last_aws_max_item_size
let count=1
if [[ $is_new_log == '1' ]];
then
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: START DOWNLOADING OF NEW AWS LOG'
if ! aws rds download-db-log-file-portion
--max-items $last_aws_max_item_size
--region REGION
--db-instance-identifier $db_instance
--log-file-name error/postgresql.log.$AWS_LOG_TIME > $LOG_FILE
then
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: FATAL_ERROR - Could not get log from AWS .'
exit 2
fi
else
next_token=`psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -v ON_ERROR_STOP=1 -A -t -c "select last_aws_nexttoken from database where id = $database_id "`
if [[ $next_token == '' ]];
then
next_token='0'
fi
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: CONTINUE DOWNLOADING OF AWS LOG'
if ! aws rds download-db-log-file-portion
--max-items $last_aws_max_item_size
--starting-token $next_token
--region REGION
--db-instance-identifier $db_instance
--log-file-name error/postgresql.log.$AWS_LOG_TIME > $LOG_FILE
then
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: FATAL_ERROR - Could not get log from AWS .'
exit 3
fi
line_count=`cat $LOG_FILE | wc -l`
let lines=$line_count-1
tail -$lines $LOG_FILE > $TMP_MIDDLE
mv -f $TMP_MIDDLE $LOG_FILE
fi
next_token_str=`cat $LOG_FILE | grep NEXTTOKEN`
next_token=`echo $next_token_str | awk -F" " '{ print $2}' `
grep -v NEXTTOKEN $LOG_FILE > $TMP_FILE
if [[ $next_token == '' ]];
then
cp $TMP_FILE $RESULT_FILE
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: NEXTTOKEN NOT FOUND - FINISH '
rm $LOG_FILE
rm $TMP_FILE
rm $TMP_MIDDLE
rm $TMP_MIDDLE2
exit 0
else
psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -v ON_ERROR_STOP=1 -A -t -q -c "update database set last_aws_nexttoken = '$next_token' where id = $database_id "
fi
first_str=`tail -1 $TMP_FILE`
line_count=`cat $TMP_FILE | wc -l`
let lines=$line_count-1
head -$lines $TMP_FILE > $RESULT_FILE
###############################################
# MAIN CIRCLE
let count=2
while [[ $next_token != '' ]];
do
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: count='$count
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: START DOWNLOADING OF AWS LOG'
if ! aws rds download-db-log-file-portion
--max-items $last_aws_max_item_size
--starting-token $next_token
--region REGION
--db-instance-identifier $db_instance
--log-file-name error/postgresql.log.$AWS_LOG_TIME > $LOG_FILE
then
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: FATAL_ERROR - Could not get log from AWS .'
exit 4
fi
next_token_str=`cat $LOG_FILE | grep NEXTTOKEN`
next_token=`echo $next_token_str | awk -F" " '{ print $2}' `
TMP_FILE=$LOG_FILE'.tmp'
grep -v NEXTTOKEN $LOG_FILE > $TMP_FILE
last_str=`head -1 $TMP_FILE`
if [[ $next_token == '' ]];
then
concat_str=$first_str$last_str
echo $concat_str >> $RESULT_FILE
line_count=`cat $TMP_FILE | wc -l`
let lines=$line_count-1
tail -$lines $TMP_FILE >> $RESULT_FILE
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: NEXTTOKEN NOT FOUND - FINISH '
rm $LOG_FILE
rm $TMP_FILE
rm $TMP_MIDDLE
rm $TMP_MIDDLE2
exit 0
fi
if [[ $next_token != '' ]];
then
let growth_counter=$growth_counter+1
if [[ $growth_counter -gt $growth_counter_max ]];
then
let last_aws_max_item_size=$last_aws_max_item_size*$growth_factor
let growth_counter=1
fi
if [[ $last_aws_max_item_size -gt $max_item_size ]];
then
let last_aws_max_item_size=$max_item_size
fi
psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -A -t -q -c "update database set last_aws_nexttoken = '$next_token' where id = $database_id "
concat_str=$first_str$last_str
echo $concat_str >> $RESULT_FILE
line_count=`cat $TMP_FILE | wc -l`
let lines=$line_count-1
#############################
#Get middle of file
head -$lines $TMP_FILE > $TMP_MIDDLE
line_count=`cat $TMP_MIDDLE | wc -l`
let lines=$line_count-1
tail -$lines $TMP_MIDDLE > $TMP_MIDDLE2
cat $TMP_MIDDLE2 >> $RESULT_FILE
first_str=`tail -1 $TMP_FILE`
fi
let count=$count+1
done
#
#################################################################
exit 0
Script snippets with some explanations:
Input parameters of the script:
- Timestamp of the log file name in the format YYYY-MM-DD-HH24: AWS_LOG_TIME=$1
- Database ID: database_id=$2
- Name of the assembled log file: RESULT_FILE=$3
Get the timestamp of the last loaded log file:
current_aws_log_time=`psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -A -t -c "select last_aws_log_time from database where id = $database_id "`If the timestamp of the last loaded log file does not match the input parameter, a new log file is loaded:
if [[ $current_aws_log_time != $AWS_LOG_TIME ]]; then is_new_log='1' if ! psql -h ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -v ON_ERROR_STOP=1 -A -t -c "update database set last_aws_log_time = '$AWS_LOG_TIME' where id = $database_id " then echo '***download_aws_piece.sh -FATAL_ERROR - update database set last_aws_log_time .' exit 1 fi else is_new_log='0' fi
Retrieving the value of the nexttoken label from the uploaded file:
next_token_str=`cat $LOG_FILE | grep NEXTTOKEN` next_token=`echo $next_token_str | awk -F" " '{ print $2}' `
The end of the upload is indicated by an empty nexttoken value.
In a loop, we count portions of the file while concatenating lines and increasing the portion size:
Main Loop
# MAIN CIRCLE
let count=2
while [[ $next_token != '' ]];
do
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: count='$count
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: START DOWNLOADING OF AWS LOG'
if ! aws rds download-db-log-file-portion
--max-items $last_aws_max_item_size
--starting-token $next_token
--region REGION
--db-instance-identifier $db_instance
--log-file-name error/postgresql.log.$AWS_LOG_TIME > $LOG_FILE
then
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: FATAL_ERROR - Could not get log from AWS .'
exit 4
fi
next_token_str=`cat $LOG_FILE | grep NEXTTOKEN`
next_token=`echo $next_token_str | awk -F" " '{ print $2}' `
TMP_FILE=$LOG_FILE'.tmp'
grep -v NEXTTOKEN $LOG_FILE > $TMP_FILE
last_str=`head -1 $TMP_FILE`
if [[ $next_token == '' ]];
then
concat_str=$first_str$last_str
echo $concat_str >> $RESULT_FILE
line_count=`cat $TMP_FILE | wc -l`
let lines=$line_count-1
tail -$lines $TMP_FILE >> $RESULT_FILE
echo $(date +%Y%m%d%H%M)': download_aws_piece.sh: NEXTTOKEN NOT FOUND - FINISH '
rm $LOG_FILE
rm $TMP_FILE
rm $TMP_MIDDLE
rm $TMP_MIDDLE2
exit 0
fi
if [[ $next_token != '' ]];
then
let growth_counter=$growth_counter+1
if [[ $growth_counter -gt $growth_counter_max ]];
then
let last_aws_max_item_size=$last_aws_max_item_size*$growth_factor
let growth_counter=1
fi
if [[ $last_aws_max_item_size -gt $max_item_size ]];
then
let last_aws_max_item_size=$max_item_size
fi
psql -h MONITOR_ENDPOINT.rds.amazonaws.com -U USER -d MONITOR_DATABASE -A -t -q -c "update database set last_aws_nexttoken = '$next_token' where id = $database_id "
concat_str=$first_str$last_str
echo $concat_str >> $RESULT_FILE
line_count=`cat $TMP_FILE | wc -l`
let lines=$line_count-1
#############################
#Get middle of file
head -$lines $TMP_FILE > $TMP_MIDDLE
line_count=`cat $TMP_MIDDLE | wc -l`
let lines=$line_count-1
tail -$lines $TMP_MIDDLE > $TMP_MIDDLE2
cat $TMP_MIDDLE2 >> $RESULT_FILE
first_str=`tail -1 $TMP_FILE`
fi
let count=$count+1
done
What’s next?
So, the first intermediate task — "upload the log file from the cloud" is resolved. What to do with the uploaded log?
First, it is necessary to parse the log file and extract the requests from it.
The task isn't very difficult. A simple bash script can handle it quite well.
upload_log_query.sh
#!/bin/bash
#########################################################
# upload_log_query.sh
# Upload table table from dowloaded aws file
# version HABR
###########################################################
echo 'TIMESTAMP:'$(date +%c)' Upload log_query table '
source_file=$1
echo 'source_file='$source_file
database_id=$2
echo 'database_id='$database_id
beginer=' '
first_line='1'
let "line_count=0"
sql_line=' '
sql_flag=' '
space=' '
cat $source_file | while read line
do
line="$space$line"
if [[ $first_line == "1" ]]; then
beginer=`echo $line | awk -F" " '{ print $1}' `
first_line='0'
fi
current_beginer=`echo $line | awk -F" " '{ print $1}' `
if [[ $current_beginer == $beginer ]]; then
if [[ $sql_flag == '1' ]]; then
sql_flag='0'
log_date=`echo $sql_line | awk -F" " '{ print $1}' `
log_time=`echo $sql_line | awk -F" " '{ print $2}' `
duration=`echo $sql_line | awk -F" " '{ print $5}' `
#replace ' to ''
sql_modline=`echo "$sql_line" | sed 's/'''/''''''/g'`
sql_line=' '
################
#PROCESSING OF THE SQL-SELECT IS HERE
if ! psql -h ENDPOINT.rds.amazonaws.com -U USER -d DATABASE -v ON_ERROR_STOP=1 -A -t -c "select log_query('$ip_port',$database_id , '$log_date' , '$log_time' , '$duration' , '$sql_modline' )"
then
echo 'FATAL_ERROR - log_query '
exit 1
fi
################
fi #if [[ $sql_flag == '1' ]]; then
let "line_count=line_count+1"
check=`echo $line | awk -F" " '{ print $8}' `
check_sql=${check^^}
#echo 'check_sql='$check_sql
if [[ $check_sql == 'SELECT' ]]; then
sql_flag='1'
sql_line="$sql_line$line"
ip_port=`echo $sql_line | awk -F":" '{ print $4}' `
fi
else
if [[ $sql_flag == '1' ]]; then
sql_line="$sql_line$line"
fi
fi #if [[ $current_beginer == $beginer ]]; then
done
Now, with the extracted request from the log file, you can work on it.
And several useful opportunities open up.
The parsed requests need to be stored somewhere. For this, a service table is used. log_query
CREATE TABLE log_query
(
id SERIAL,
queryid bigint,
query_md5hash text not null,
database_id integer not null,
timepoint timestamp without time zone not null,
duration double precision not null,
query text not null,
explained_plan text[],
plan_md5hash text,
explained_plan_wo_costs text[],
plan_hash_value text,
baseline_id integer,
ip text,
port text
);
ALTER TABLE log_query ADD PRIMARY KEY (id);
ALTER TABLE log_query ADD CONSTRAINT queryid_timepoint_unique_key UNIQUE (queryid, timepoint);
ALTER TABLE log_query ADD CONSTRAINT query_md5hash_timepoint_unique_key UNIQUE (query_md5hash, timepoint);
CREATE INDEX log_query_timepoint_idx ON log_query (timepoint);
CREATE INDEX log_query_queryid_idx ON log_query (queryid);
ALTER TABLE log_query ADD CONSTRAINT database_id_fk FOREIGN KEY (database_id) REFERENCES database (id) ON DELETE CASCADE;
The processing of the parsed request is carried out in plpgsql the function "log_query».
log_query.sql
--log_query.sql
--version HABR
CREATE OR REPLACE FUNCTION log_query( ip_port text , log_database_id integer , log_date text , log_time text , duration text , sql_line text ) RETURNS boolean AS $$
DECLARE
result boolean ;
log_timepoint timestamp without time zone ;
log_duration double precision ;
pos integer ;
log_query text ;
activity_string text ;
log_md5hash text ;
log_explain_plan text[] ;
log_planhash text ;
log_plan_wo_costs text[] ;
database_rec record ;
pg_stat_query text ;
test_log_query text ;
log_query_rec record ;
found_flag boolean ;
pg_stat_history_rec record ;
port_start integer ;
port_end integer ;
client_ip text ;
client_port text ;
log_queryid bigint ;
log_query_text text ;
pg_stat_query_text text ;
BEGIN
result = TRUE ;
RAISE NOTICE '***log_query';
port_start = position('(' in ip_port);
port_end = position(')' in ip_port);
client_ip = substring( ip_port from 1 for port_start-1 );
client_port = substring( ip_port from port_start+1 for port_end-port_start-1 );
SELECT e.host , d.name , d.owner_pwd
INTO database_rec
FROM database d JOIN endpoint e ON e.id = d.endpoint_id
WHERE d.id = log_database_id ;
log_timepoint = to_timestamp(log_date||' '||log_time,'YYYY-MM-DD HH24-MI-SS');
log_duration = duration:: double precision;
pos = position ('SELECT' in UPPER(sql_line) );
log_query = substring( sql_line from pos for LENGTH(sql_line));
log_query = regexp_replace(log_query,' +',' ','g');
log_query = regexp_replace(log_query,';+','','g');
log_query = trim(trailing ' ' from log_query);
log_md5hash = md5( log_query::text );
--Explain execution plan--
EXECUTE 'SELECT dblink_connect(''LINK1'',''host='||database_rec.host||' dbname='||database_rec.name||' user=DATABASE password='||database_rec.owner_pwd||' '')';
log_explain_plan = ARRAY ( SELECT * FROM dblink('LINK1', 'EXPLAIN '||log_query ) AS t (plan text) );
log_plan_wo_costs = ARRAY ( SELECT * FROM dblink('LINK1', 'EXPLAIN ( COSTS FALSE ) '||log_query ) AS t (plan text) );
PERFORM dblink_disconnect('LINK1');
--------------------------
BEGIN
INSERT INTO log_query
(
query_md5hash ,
database_id ,
timepoint ,
duration ,
query ,
explained_plan ,
plan_md5hash ,
explained_plan_wo_costs ,
plan_hash_value ,
ip ,
port
)
VALUES
(
log_md5hash ,
log_database_id ,
log_timepoint ,
log_duration ,
log_query ,
log_explain_plan ,
md5(log_explain_plan::text) ,
log_plan_wo_costs ,
md5(log_plan_wo_costs::text),
client_ip ,
client_port
);
activity_string = 'New query has been logged '||
' database_id = '|| log_database_id ||
' query_md5hash='||log_md5hash||
' , timepoint = '||to_char(log_timepoint,'YYYYMMDD HH24:MI:SS');
RAISE NOTICE '%',activity_string;
PERFORM pg_log( log_database_id , 'log_query' , activity_string);
EXCEPTION
WHEN unique_violation THEN
RAISE NOTICE '*** unique_violation *** query already has been logged';
END;
SELECT queryid
INTO log_queryid
FROM log_query
WHERE query_md5hash = log_md5hash AND
timepoint = log_timepoint;
IF log_queryid IS NOT NULL
THEN
RAISE NOTICE 'log_query with query_md5hash = % and timepoint = % has already had a QUERYID = %',log_md5hash,log_timepoint , log_queryid ;
RETURN result;
END IF;
------------------------------------------------
RAISE NOTICE 'Update queryid';
SELECT *
INTO log_query_rec
FROM log_query
WHERE query_md5hash = log_md5hash AND timepoint = log_timepoint ;
log_query_rec.query=regexp_replace(log_query_rec.query,';+','','g');
FOR pg_stat_history_rec IN
SELECT
queryid ,
query
FROM
pg_stat_db_queries
WHERE
database_id = log_database_id AND
queryid is not null
LOOP
pg_stat_query = pg_stat_history_rec.query ;
pg_stat_query=regexp_replace(pg_stat_query,'n+',' ','g');
pg_stat_query=regexp_replace(pg_stat_query,'t+',' ','g');
pg_stat_query=regexp_replace(pg_stat_query,' +',' ','g');
pg_stat_query=regexp_replace(pg_stat_query,'$.','%','g');
log_query_text = trim(trailing ' ' from log_query_rec.query);
pg_stat_query_text = pg_stat_query;
--SELECT log_query_rec.query like pg_stat_query INTO found_flag ;
IF (log_query_text LIKE pg_stat_query_text) THEN
found_flag = TRUE ;
ELSE
found_flag = FALSE ;
END IF;
IF found_flag THEN
UPDATE log_query SET queryid = pg_stat_history_rec.queryid WHERE query_md5hash = log_md5hash AND timepoint = log_timepoint ;
activity_string = ' updated queryid = '||pg_stat_history_rec.queryid||
' for log_query with id = '||log_query_rec.id
;
RAISE NOTICE '%',activity_string;
EXIT ;
END IF ;
END LOOP ;
RETURN result ;
END
$$ LANGUAGE plpgsql;
A service table is used during processing pg_stat_db_queries, containing a snapshot of current queries from the table pg_stat_history (The usage of the table is described here — )
TABLE pg_stat_db_queries
(
database_id integer,
queryid bigint ,
query text ,
max_time double precision
);
TABLE pg_stat_history
(
…
database_id integer ,
…
queryid bigint ,
…
max_time double precision ,
…
);
The function allows for a range of useful capabilities for processing requests from the log file. Specifically:
Feature #1 — Query Execution History
Very useful for beginning to address a performance incident. First, review the history — when did the slowdown begin?
Then, as is classic — look for external causes. Perhaps the database load simply increased sharply, and a specific query is not to blame.
Add a new record to the log_query table
port_start = position('(' in ip_port);
port_end = position(')' in ip_port);
client_ip = substring( ip_port from 1 for port_start-1 );
client_port = substring( ip_port from port_start+1 for port_end-port_start-1 );
SELECT e.host , d.name , d.owner_pwd
INTO database_rec
FROM database d JOIN endpoint e ON e.id = d.endpoint_id
WHERE d.id = log_database_id ;
log_timepoint = to_timestamp(log_date||' '||log_time,'YYYY-MM-DD HH24-MI-SS');
log_duration = to_number(duration,'99999999999999999999D9999999999');
pos = position ('SELECT' in UPPER(sql_line) );
log_query = substring( sql_line from pos for LENGTH(sql_line));
log_query = regexp_replace(log_query,' +',' ','g');
log_query = regexp_replace(log_query,';+','','g');
log_query = trim(trailing ' ' from log_query);
RAISE NOTICE 'log_query=%',log_query ;
log_md5hash = md5( log_query::text );
--Explain execution plan--
EXECUTE 'SELECT dblink_connect(''LINK1'',''host='||database_rec.host||' dbname='||database_rec.name||' user=DATABASE password='||database_rec.owner_pwd||' '')';
log_explain_plan = ARRAY ( SELECT * FROM dblink('LINK1', 'EXPLAIN '||log_query ) AS t (plan text) );
log_plan_wo_costs = ARRAY ( SELECT * FROM dblink('LINK1', 'EXPLAIN ( COSTS FALSE ) '||log_query ) AS t (plan text) );
PERFORM dblink_disconnect('LINK1');
--------------------------
BEGIN
INSERT INTO log_query
(
query_md5hash ,
database_id ,
timepoint ,
duration ,
query ,
explained_plan ,
plan_md5hash ,
explained_plan_wo_costs ,
plan_hash_value ,
ip ,
port
)
VALUES
(
log_md5hash ,
log_database_id ,
log_timepoint ,
log_duration ,
log_query ,
log_explain_plan ,
md5(log_explain_plan::text) ,
log_plan_wo_costs ,
md5(log_plan_wo_costs::text),
client_ip ,
client_port
);
Feature #2 — Save Execution Plans for Queries
At this point, an objection-clarification-comment may arise: "But there is already autoexplain" It exists, but what good is it if the execution plan is stored in the same log file, and to save it for further analysis, one would have to parse the log file?
What I needed was:
First: to store the execution plan in the monitoring database's service table;
Secondly, it is important to compare execution plans with each other to immediately see if the execution plan of the query has changed.
There is a request with specific execution parameters. Obtaining and saving its execution plan using EXPLAIN is a straightforward task.
Moreover, by using the EXPLAIN (COSTS FALSE) statement, one can obtain the framework of the plan, which will be used to generate a hash value for the plan, aiding in the subsequent analysis of plan changes history.
Obtain the execution plan template
--Explain execution plan--
EXECUTE 'SELECT dblink_connect(''LINK1'',''host='||database_rec.host||' dbname='||database_rec.name||' user=DATABASE password='||database_rec.owner_pwd||' '')';
log_explain_plan = ARRAY ( SELECT * FROM dblink('LINK1', 'EXPLAIN '||log_query ) AS t (plan text) );
log_plan_wo_costs = ARRAY ( SELECT * FROM dblink('LINK1', 'EXPLAIN ( COSTS FALSE ) '||log_query ) AS t (plan text) );
PERFORM dblink_disconnect('LINK1');
Option #3 — Using the query log for monitoring
Since performance metrics are set up not on the text of the query, but on its ID, it's necessary to link queries from the log file to those for which performance metrics are configured.
At least to have the exact time of the performance incident occurrence.
Thus, in the event of a performance incident for the query ID, there will be a link to the specific query with specific parameter values and the exact time of execution and duration of the query. This information can only be obtained using the view. pg_stat_statements — cannot.
Find the query ID of the request and update the record in the log_query table
SELECT *
INTO log_query_rec
FROM log_query
WHERE query_md5hash = log_md5hash AND timepoint = log_timepoint ;
log_query_rec.query=regexp_replace(log_query_rec.query,';+','','g');
FOR pg_stat_history_rec IN
SELECT
queryid ,
query
FROM
pg_stat_db_queries
WHERE
database_id = log_database_id AND
queryid is not null
LOOP
pg_stat_query = pg_stat_history_rec.query ;
pg_stat_query=regexp_replace(pg_stat_query,'n+',' ','g');
pg_stat_query=regexp_replace(pg_stat_query,'t+',' ','g');
pg_stat_query=regexp_replace(pg_stat_query,' +',' ','g');
pg_stat_query=regexp_replace(pg_stat_query,'$.','%','g');
log_query_text = trim(trailing ' ' from log_query_rec.query);
pg_stat_query_text = pg_stat_query;
--SELECT log_query_rec.query like pg_stat_query INTO found_flag ;
IF (log_query_text LIKE pg_stat_query_text) THEN
found_flag = TRUE ;
ELSE
found_flag = FALSE ;
END IF;
IF found_flag THEN
UPDATE log_query SET queryid = pg_stat_history_rec.queryid WHERE query_md5hash = log_md5hash AND timepoint = log_timepoint ;
activity_string = ' updated queryid = '||pg_stat_history_rec.queryid||
' for log_query with id = '||log_query_rec.id
;
RAISE NOTICE '%',activity_string;
EXIT ;
END IF ;
END LOOP ;
Afterword
The described methodology has ultimately found its application in , allowing for more information for analysis when resolving performance incidents related to queries.
However, in my personal and professional view, further work is needed on the algorithm for selecting and resizing the loaded batch. The task is still unsolved in general. It might be interesting.
But that's a completely different story…
Source: habr.com
