On Tue, 11 Jul 2023 at 10:43, Greg Sabino Mullane <htamf...@gmail.com> wrote:
> This has been a long-standing annoyance of mine. Who hasn't done something > like this?: > > psql> SET random_page_cost = 2.5; > (do some stuff, realize that rpc was too high) > > Let's put that inside of postgresql.conf: > > > #------------------------------------------------------------------------------ > # CUSTOMIZED OPTIONS > > #------------------------------------------------------------------------------ > > # Add settings for extensions here > > random_page_cost = 2.5; > > > Boom! Server will not start. Surely, we can be a little more liberal in > what we accept? Attached patch allows a single trailing semicolon to be > silently discarded. As this parsing happens before the logging collector > starts up, the error about the semicolon is often buried somewhere in a > separate logfile or journald - so let's just allow postgres to start up > since there is no ambiguity about what random_page_cost (or any other GUC) > is meant to be set to. > Please, no! There is no end to accepting sloppy syntax. What next, allow "SET random_page_cost = 2.5;" (with or without semicolon) in config files? I'd be more interested in improvements in visibility of errors. For example, maybe if I try to start the server and there is a config file problem, I could somehow get a straightforward error message right in the terminal window complaining about the line of the configuration which is wrong. Or maybe there could be a "check configuration" subcommand which checks the configuration. If it's fine, say so and set a flag saying the server is clear to be started/restarted. If not, give useful error messages and don't set the flag. Then make the start/restart commands only do their thing if the "config OK" flag is set. Make sure that editing the configuration clears the flag (or have 2 copies of the configuration, copied over by the "check" subcommand: one for editing, one for running with). This might properly belong outside of Postgres itself, I don't know. But I think it would be way more useful than a potentially never-ending series of patches to liberalize the config parser.