Improving Zabbix + PostgreSQL Performance with Partitioning and Indexing

import argparse """     .   20   5000    ( 2500  :  echo,  ;  ping,  ) """ TEMP_HEAD = """ <?xml version="1.0" encoding="UTF-8"?> <zabbix_export> <version>2.0</version> <date>2015-08-17T23:15:01Z</date> <groups> <group> <name>Templates</name> </group> </groups> <templates> <template> <template>Template Zabbix Srv Stress {count} passive {char}</template> <name>Template Zabbix Srv Stress {count} passive {char}</name> <description/> <groups> <group> <name>Templates</name> </group> </groups> <applications/> <items> """ TEMP_END = """</items> <discovery_rules/> <macros/> <templates/> <screens/> </template> </templates> </zabbix_export> """ TEMP_ITEM = """<item> <name>{k}</name> <type>0</type> <snmp_community/> <multiplier>0</multiplier> <snmp_oid/> <key>{k}</key> <delay>1m</delay> <history>3</history> <trends>365</trends> <status>0</status> <value_type>{t}</value_type> <allowed_hosts/> <units/> <delta>0</delta> <snmpv3_contextname/> <snmpv3_securityname/> <snmpv3_securitylevel>0</snmpv3_securitylevel> <snmpv3_authprotocol>0</snmpv3_authprotocol> <snmpv3_authpassphrase/> <snmpv3_privprotocol>0</snmpv3_privprotocol> <snmpv3_privpassphrase/> <formula>1</formula> <delay_flex/> <params/> <ipmi_sensor/> <data_type>0</data_type> <authtype>0</authtype> <username/> <password/> <publickey/> <privatekey/> <port/> <description/> <inventory_link>0</inventory_link> <applications/> <valuemap/> <logtimefmt/> </item> """ TMP_FNAME_DEFAULT = "Template_App_Zabbix_Server_Stress_{count}_passive_{char}.xml" chars = "ABCDEFGHIJKLMNOPQRSTUVWXYZ" if __name__ == "__main__": parser = argparse.ArgumentParser( description='     zabbix') parser.add_argument('--items', dest='items', type=int, default=1000, help='-   (default: 1000)') parser.add_argument('--templates', dest='templates', type=int, default=1, help=f'-  [1-{len(chars)}] (default: 1)') args = parser.parse_args() items_count = args.items tmps_count = args.templates if not (tmps_count >= 1 and tmps_count <= len(chars)): sys.exit(f"Templates must be in range 1 - {len(chars)}") for i in range(tmps_count): fname = TMP_FNAME_DEFAULT.format(count=items_count, char=chars[i]) with open(fname, "w") as output: output.write(TEMP_HEAD.format(count=items_count, char=chars[i])) for k,t in [('stress.ping[{}-I-{:06d}]',3), ('stress.echo[{}-S-{:06d}]',4)]: for j in range(int(items_count/2)): output.write(TEMP_ITEM.format(k=k.format(chars[i],j),t=t)) output.write(TEMP_END)
      
      

        
        
        
      

    
        
        
        
      
      

        
        
        
      

    
     

 # DB Version: 11
 # OS Type: linux
 # DB Type: web
 # Total Memory (RAM): 8 GB
 # CPUs num: 1
 # Connections num: 100
 # Data Storage: hdd

 max_connections = 100 # maximum number of simultaneous database connections
 shared_buffers = 2GB # memory size for various buffers (mainly cache of table blocks and indexes) in shared memory
 effective_cache_size = 6GB # maximum size of required memory for query execution using indexes
 maintenance_work_mem = 512MB # affects the speed of operations VACUUM, ANALYZE, CREATE INDEX
 checkpoint_completion_target = 0.7 # target time to complete the checkpoint procedure
 wal_buffers = 16MB # amount of memory used by Shared Memory to maintain transaction logs
 default_statistics_target = 100 # amount of statistics collected by the ANALYZE command - when increasing, the optimizer builds queries more slowly, but better
 random_page_cost = 4 # conditional cost of index access to data pages - affects the decision to use the index
 effective_io_concurrency = 2 # number of asynchronous I / O operations that the DBMS will try to perform in a separate session
 work_mem = 10485kB # amount of memory used for sorting and hash tables before temporary files on disk are used
 min_wal_size = 1GB # limits below the number of WAL files that will be recycled for future use
 max_wal_size = 2GB # limits the number of WAL files that will be recycled for future use 

 # managing concurrent request handlers
 max_worker_processes = 8 # maximum number of background processes - at least one per database
 max_parallel_workers_per_gather = 4 # maximum number of parallel processes within a single request
 max_parallel_workers = 8 # the maximum number of work processes that the system can support for parallel operations

 # logging settings (an easy way to find out about the execution time of requests without using the pg_stat_statements extension)
 log_min_duration_statement = 3000 # write to the logs the duration of the execution of all commands whose operating time> = of the specified value in ms
 log_duration = off # record the duration of each completed command
 log_statement = 'none' # which SQL commands to write to the log, values: none (disabled), ddl, mod and all (all commands)
 debug_print_plan = off # output of the query plan tree for further analysis

 # squeeze the maximum out of the database and be ready to receive it for any failure (for the most repressed, who ignore the existence of ssd and a distributed cluster)
 #fsync = off # physical write to the changes disk, disabling fsync gives a speed gain, but can lead to permanent failures
 #synchronous_commit = off # allows you to respond to the client even before the transaction information is in the WAL - an almost safe alternative to disabling fsync
 #full_page_writes = off # shutdown speeds up normal operations, but can lead to data corruption or data corruption if the system crashes 

  mount / dev / sdc1 / var / lib / pgsql / 10 / data / base -o noatime 

 # where UUID is the identifier of the disk, you can see it using the blkid utility
 UUID = 121efe29-70bf-410b-bc71-90704568ce3b / var / lib / pgsql / 10 / data / base ext4 defaults, noatime 0 0 

 yum install pg_pathman10
 nano /var/pgsqldb/postgresql.conf
 shared_preload_libraries = 'pg_pathman' # important - here write pg_pathman last in the list

 systemctl restart postgresql-10.service
 psql -d zabbix -U postgres
 CREATE EXTENSION pg_pathman;
 # configure one day for the tables of primary monitoring data
 # 1552424400 - countdown as unix timestamp, 86400 - seconds in days
 select create_range_partitions ('history', 'clock', 1552424400, 86400);
 select create_range_partitions ('history_uint', 'clock', 1552424400, 86400);
 select create_range_partitions ('history_text', 'clock', 1552424400, 86400);
 select create_range_partitions ('history_str', 'clock', 1552424400, 86400);
 select create_range_partitions ('history_log', 'clock', 1552424400, 86400);
 # configure for three days for aggregated monitoring data tables
 # 1552424400 - countdown as unix timestamp, 259200 - seconds in three days
 select create_range_partitions ('trends', 'clock', 1545771600, 259200);  
 select create_range_partitions ('trends_uint', 'clock', 1545771600, 259200); 

 # general list of partitioned tables, main configuration storage:
 select * from pathman_config;
 # representation with all existing sections, as well as their parents and range boundaries:
 select * from pathman_partition_list;
 # additional parameters that override the standard pg_pathman behavior:
 select * from pathman_config_params;
 # copy the contents back to the parent table and delete partitions:
 select drop_partitions ('table_name' :: regclass, false);

 /*       */ SELECT nspname AS schemaname, relname, relkind, cast (reltuples as int), pg_size_pretty(pg_relation_size(C.oid)) AS "size" FROM pg_class C LEFT JOIN pg_namespace N ON (N.oid = C.relnamespace) WHERE nspname NOT IN ('pg_catalog', 'information_schema') and (relname like 'history%' or relname like 'trends%') and relkind = 'r' -- and reltuples > 0 -- and pg_relation_size(C.oid) >= 0 ORDER BY schemaname, relname
      
      

        
        
        
      

    
        
        
        
      
      

        
        
        
      

    
     

 CREATE OR REPLACE FUNCTION public.delete_old_partitions (history_days integer, trends_days integer, str_days integer)
  RETURNS text
  LANGUAGE plpgsql
 AS $ function $
 / *
 The function deletes all partitions older than the specified number of days:
 history_days - for partitions history_x, history_uint_x
 trends_days - for partitions trends_x, trends_uint_x
 str_days - for partitions history_str_x, history_text_x, history_log_x
 * /
 declare clock_today_start int;
 declare clock_delete_less_history int = 0;
 declare clock_delete_less_trends int = 0;
 declare clock_delete_less_strings int = 0;
 clock_delete_less int = 0;
 declare iterator int = 0;
 declare result_str text = '';
 declare buf_table_size text;
 declare buf_table_len text;
 declare partition_name text;
 declare clock_max text;
 declare err_detail text;
 declare t_start timestamp = clock_timestamp ();
 declare t_end timestamp;
 begin
     if $ 1 <= 0 then return 'ups, something wrong: history_days argument must be positive integer value';  end if;
     if $ 2 <= 0 then return 'ups, something wrong: trends_days argument must be positive integer value';  end if;
     if $ 3 <= 0 then return 'ups, something wrong: str_days argument must be positive integer value';  end if;
     clock_today_start = extract (epoch from date_trunc ('day', now ())) :: int;
     clock_delete_less_history = extract (epoch from date_trunc ('day', now ()) - ($ 1 :: text || 'days') :: interval) :: int;
     clock_delete_less_trends = extract (epoch from date_trunc ('day', now ()) - ($ 2 :: text || 'days') :: interval) :: int;
     clock_delete_less_strings = extract (epoch from date_trunc ('day', now ()) - ($ 3 :: text || 'days') :: interval) :: int;
     clock_delete_less = least (clock_delete_less_history, clock_delete_less_trends, clock_delete_less_strings);
     --raise notice 'clock_today_start% (%)', to_timestamp (clock_today_start), clock_today_start;
     --raise notice 'clock_delete_less_history% (%)% days', to_timestamp (clock_delete_less_history), clock_delete_less_history, $ 1;
     --raise notice 'clock_delete_less_trends% (%)% days', to_timestamp (clock_delete_less_trends), clock_delete_less_trends, $ 2;
     --raise notice 'clock_delete_less_strings% (%)% days', to_timestamp (clock_delete_less_strings), clock_delete_less_strings, $ 3;
     for partition_name, clock_max in select partition, range_max from pathman_partition_list where 
     range_max :: int <= greatest (clock_delete_less_history, clock_delete_less_trends, clock_delete_less_strings) and 
     (partition :: text like 'history%' or partition :: text like 'trends%') order by partition asc
     loop
         if (partition_name ~ 'history_uint_ \ d' and clock_max :: int <= clock_delete_less_history)
         or (partition_name ~ 'history_ \ d' and clock_max :: int <= clock_delete_less_history)
         or (partition_name ~ 'trends_ \ d' and clock_max :: int <= clock_delete_less_trends)
         or (partition_name ~ 'history_log_ \ d' and clock_max :: int <= clock_delete_less_strings)
         or (partition_name ~ 'history_str_ \ d' and clock_max :: int <= clock_delete_less_strings)
         or (partition_name ~ 'history_text_ \ d' and clock_max :: int <= clock_delete_less_strings)
         then 
             iterator = iterator + 1;
             raise notice '%', format ('!!! delete% s% s', partition_name, clock_max);
             select max (reltuples :: int), pg_size_pretty (sum (pg_relation_size (pg_class.oid))) as "size" from pg_class where relname like partition_name ||  '%' into strict buf_table_len, buf_table_size;
             if result_str! = '' then result_str = result_str ||  ',';  end if;
             result_str = result_str ||  format ('% s (dt <% s, len% s,% s)', partition_name, to_char (to_timestamp (clock_max :: int), 'YYYY-MM-DD'), buf_table_len, buf_table_size);
             execute format ('drop table if exists% s', partition_name);
         end if;
     end loop;
     if iterator = 0 then result_str = format ('there is no partitions to delete older, then% s date', to_char (to_timestamp (clock_delete_less), 'YYYY-MM-DD')); 
     else result_str = format ('deleted% s partitions in% s seconds:', iterator, trunc (extract (seconds from (clock_timestamp () - t_start)) :: numeric, 3)) ||  result_str;
     end if;
     --raise notice '%', result_str;
     return result_str;
 exception when others then
    get stacked diagnostics err_detail = PG_EXCEPTION_CONTEXT;
    return format ('ups, something wrong:% s [err code% s],% s', sqlerrm, sqlstate, err_detail);
 end; 
 $ function $;

 - type: database monitor
 - name: delete_old_history_partitions
 - key: db.odbc.select [delete_old_history_partitions, zabbix]
 - sql expression: select delete_old_partitions (3, 30, 30);
 # here, the parameters for calling the delete_old_partitions function indicate the storage time in days 
 # for numeric values, aggregated numeric values ​​and string values
 - data type: Text
 - update interval: 0
 - user interval: scheduled in h2
 - history storage period: 90 days
 - data element group: Database

  2019-09-16 02:00:00, deleted 3 partitions in 0.024 seconds: trends_78 (dt <2019-08-17, len 1, 48 kB), history_193 (dt <2019-09-13, len 85343, 9448 kB ), history_uint_186 (dt <2019-09-13, len 27969, 3480 kB)

/ *
     gin    biginteger  integer    .
 -   gin     int2, int4, int8,
       bigint     ,     bigint (<= 2147483647)
        intger_ops,  :
create index on tablename using gin(columnname int8_family_ops) with (fastupdate = false);
 * /

--       btree_gin
CREATE EXTENSION btree_gin;

CREATE OPERATOR FAMILY integer_ops using gin;

CREATE OPERATOR CLASS int4_family_ops
FOR TYPE int4 USING gin FAMILY integer_ops
AS
    OPERATOR 1 <,
    OPERATOR 2 <=,
    OPERATOR 3 =,
    OPERATOR 4 >=,
    OPERATOR 5 >,
    FUNCTION 1 btint4cmp(int4,int4),
    FUNCTION 2 gin_extract_value_int4(int4, internal),
    FUNCTION 3 gin_extract_query_int4(int4, internal, int2, internal, internal),
    FUNCTION 4 gin_btree_consistent(internal, int2, anyelement, int4, internal, internal),
    FUNCTION 5 gin_compare_prefix_int4(int4,int4,int2, internal),
STORAGE int4;

CREATE OPERATOR CLASS int8_family_ops
FOR TYPE int8 USING gin FAMILY integer_ops
AS
    OPERATOR 1 <,
    OPERATOR 2 <=,
    OPERATOR 3 =,
    OPERATOR 4 >=,
    OPERATOR 5 >,
    FUNCTION 1 btint8cmp(int8,int8),
    FUNCTION 2 gin_extract_value_int8(int8, internal),
    FUNCTION 3 gin_extract_query_int8(int8, internal, int2, internal, internal),
    FUNCTION 4 gin_btree_consistent(internal, int2, anyelement, int4, internal, internal),
    FUNCTION 5 gin_compare_prefix_int8(int8,int8,int2, internal),
STORAGE int8;

ALTER OPERATOR FAMILY integer_ops USING gin add
  OPERATOR 1 <(int4,int8),
  OPERATOR 2 <=(int4,int8),
  OPERATOR 3 =(int4,int8),
  OPERATOR 4 >=(int4,int8),
  OPERATOR 5 >(int4,int8);

ALTER OPERATOR FAMILY integer_ops USING gin add
  OPERATOR 1 <(int8,int4),
  OPERATOR 2 <=(int8,int4),
  OPERATOR 3 =(int8,int4),
  OPERATOR 4 >=(int8,int4),
  OPERATOR 5 >(int8,int4);

/*
        
 * /

--   
drop index history_1;
drop index history_uint_1;
drop index history_str_1;
drop index history_text_1;
drop index history_log_1;
--          PK 
-- (   ,         )
alter table trends drop constraint trends_pk;
alter table trends_uint drop constraint trends_uint_pk;

--     bree-gin   itemid    
--   btree-gin  bigint       
-- https://habr.com/ru/company/postgrespro/blog/340978/#comment_10545932
--    create extension btree_gin;
create index on history using gin(itemid int8_family_ops) with (fastupdate = false);
create index on history_uint using gin(itemid int8_family_ops) with (fastupdate = false);
create index on history_str using gin(itemid int8_family_ops) with (fastupdate = false);
create index on history_text using gin(itemid int8_family_ops) with (fastupdate = false);
create index on history_log using gin(itemid int8_family_ops) with (fastupdate = false);
create index on trends using gin(itemid int8_family_ops) with (fastupdate = false);
create index on trends_uint using gin(itemid int8_family_ops) with (fastupdate = false);

--     bree-gin   itemid    
--     brin    128 ,    
--           ,
--      https://habr.com/ru/company/postgrespro/blog/346460/
create index on history using brin(clock) with (pages_per_range = 128);
create index on history_uint using brin(clock) with (pages_per_range = 128);
create index on history_str using brin(clock) with (pages_per_range = 128);
create index on history_text using brin(clock) with (pages_per_range = 128);
create index on history_log using brin(clock) with (pages_per_range = 128);
create index on trends using brin(clock) with (pages_per_range = 128);
create index on trends_uint using brin(clock) with (pages_per_range = 128);

 CREATE EXTENSION pg_stat_statements; 

shared_preload_libraries = 'pg_stat_statements'
pg_stat_statements.max = 10000 #   sql ,     (     );
pg_stat_statements.track = all # all -   (    ), top -   /, none -  
pg_stat_statements.save = true #     

 SELECT pg_stat_statements_reset(); 

 select substring(query from '[^(]*') as query_sub, sum(calls) as calls, avg(mean_time) as mean_time from pg_stat_statements where query ~ 'insert into' or query ~ 'update trends' group by substring(query from '[^(]*') order by calls desc