PostgreSQL Workload Analyzer¶
This is the documentation for PoWA 1.x series. If you’re looking for the new (2.x series), please visit the latest documentation If you are using PostgreSQL >= 9.4, you should use the 2.x series.
PoWA is an extension designed to historize informations given by the pg_stat_statements extension. It provides sql SRF (Set Returning Functions) to gather useful information on a specified time interval. If possible (verify with pg_test_timing), also activate track_io_timing in postgresql.conf.
PoWA requires PostgreSQL 9.3 or more.
Connecting on the GUI requires a PostgreSQL user with SUPERUSER and LOGIN privileges.
For a detailed installation procedure, please read the installation guide.
Optionally, you can create a dedicated user for PoWA. For instance, connected on PostgreSQL : CREATE USER powa SUPERUSER ENCRYPTED PASSWORD ‘mypassword’ (don’t forget to change the password).
- make install in the main directory
- Make sure you have installed and configured pg_stat_statements
- create a dedicated database (powa for instance)
- create extension powa in this databse
- add “powa” in the shared_preload_libraries in postgresql.conf (you should already have configured “pg_stat_statements”)
- configure GUC in postgresql.conf (see the §Configuration below)
- configure connections in pg_hba.conf to allow connection from the server that will run the GUI
- restart instance
Upgrade from previous version:
- make install in the main directory
- restart your PostgreSQL engine to use the new powa library
- ALTER EXTENSION powa UPDATE; – This will take some time, a lot of things are rewritten as the schema is upgraded
- If you have deadlock messages, it means that the powa extension is trying to update data, while your update is doing conflicting operations. To solve this, put powa.frequency=-1 to deactivate powa temporarily, then do the extension update, and put powa.frequency back to what it was before. Don’t forget to reload your configuration each time.
Here are the configuration parameters (GUC) available:
- powa.frequency : Defines the frequency of the snapshots. Minimum 5s. You can use the usual postgresql time abbreviations. If not specified, the unit is seconds. Defaults to 5 minutes. Setting it to -1 will disable powa (powa will still start, but it won’t collect anything anymore, and wont connect to the database).
- powa.retention : Automatically purge data older than that. If not specified, the unit is minutes. Defaults to 1 day.
- powa.database : Defines the database of the workload repository. Defaults to powa.
- powa.coalesce : Defines the amount of records to group together in the table.
The more you coalesce, the more PostgreSQL can compress. But the more it has to uncompact when queried. Defaults to 100.
If you can afford it, put a rather high work_mem for the database powa. It will help, as the queries used to display the ui are doing lots of sampling, implying lots of sorts.
We use this: ALTER DATABASE powa SET work_mem TO ‘256MB’;
It’s only used for the duration of the queries anyway, this is not statically allocated memory.
Reset the stats:¶
SELECT powa_stats_reset(); (in the powa database of course)
Impact on performances¶
Using POWA will have a small negative impact on your PostgreSQL server performances. It is hard to evaluate precisely this impact but we can analyze it in 3 parts :
- First of all, you need to activate the pg_stat_statements module. This module itself may slow down your instance, but some benchmarks show that the impact is not that big.
- Second, the POWA collector should have a very low impact, but of course that depends on the frequency at which you collect data. If you do it every 5 seconds, you’ll definitely see something. At 5 minutes, the impact should be minimal.
- And finally the POWA GUI will have an impact too if you run it on the PostgreSQL instance, but it really depends on many user will have access to it.
All in all, we strongly feel that the performance impact of POWA is nothing compared to being in the dark and not knowing what is running on your database. And in most cases the impact is lower than setting
log_min_duration_statement = 0.
See our own benchmark for more details: POWA vs The Badger