Scheduler and external tool invoker for PostgreSQL
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

Postgresql Sidekick

This is a small collection of programs that allow a postgresql database to trigger actions outside the database and to schedule commands to be run in the future, without installing any extensions into the database itself.

This makes it a particularly useful approach for postgresql installations where extensions can't easily be installed, because they're a hosted service without shell access, or they're on a secured network where third-party extensions, or extensions not in the vendors repo, aren't allowed, or where the risk of instability due to potentially unstable extensions is too risky.

How does the magic happen?

The core programs connect to the database as a normal client, and listen for notifications sent from SQL using the NOTIFY command or the pg_notify() function.


pglisten listens to one or more notification channels, specified with one or more --listen=channel arguments, and prints the payload of each notification it receives to stdout.

This is mildly useful on it's own for watching for notifications, for debugging or for ad-hoc logging, but is really useful when combined with some shell scripting.

By default notification payloads are separated with newlines, but the -0 or --print0 flag will cause them to be separated with an ASCII NUL instead (much the same as the -print0 flag for the find commmand). This allows payloads containing newlines to be passed safely.

It can be used with xargs to perform almost any function on the payload.

As a slightly dangerous example, this ...

pglisten --listen=del --print0 | xargs -0 -n1 rm

... will delete /etc/passwd if you run `select pg_notify('del', '/etc/passwd').

An included sample, lets you delete files in a particular subdirectory (such as you might use from a delete trigger on a table that maintains metadata information about filesystem hosted files). It also shows how to read NUL separated arguments from a bash script:

while IFS= read -r -d '' arg; do
  # Do something with "$arg"

It could also be used to run external shell commands on demand, trigger sending of email, invalidating an external cache, such as memcache, or all sorts of other things.


pglater listens to a single notification channel, called pglater by default, and waits for notifications that consist of an integer and a SQL command, separated by a space, e.g.

30 select some_function()

The first parameter is the number of seconds in the future that the remainder of the command line should be executed. In this example it will execute some_function() after approximately thirty seconds.

pglater only keeps track of one future event at once. If it receives a second notification it will discard any pending events. This is just enough functionality to build an event scheduler within the database.

pgcron.sql is a simple scheduler based on pglater. It stores a list of scheduled SQL commands in a table and runs them at the right time.

For example, this

insert into scheduled_events (run_at, period, run_action)
    values (current_timestamp, '20 minutes',
            $$delete from spool where expires < current_timestamp$$);

will run the command to clean up the spool table immediately and then every 20 minutes.


Anyone in the database can issue a notify command. That means that any database user can use pglater to run arbitrary SQL as the database user that pglater runs as. That means that the user pglater connects as should have no more privileges than the least privileged user. (Security definer functions could be used, with some sanity checking, to allow more access for specific operations).

That also means that any database user can cause pglisten to print out any string. If it's being used with a script that performs any destructive or security related action it should include some sanity checking.


Run make to build both binaries.

This was developed on OS X and Linux, and should build with minor effort on anything unix-ish. Porting to Windows would take a little more effort.