[oauth] Per connection auth hooks in libpq
Hi,
When implementing OAuth support in the PostgreSQL driver for Ruby[1] the
main pain point I ran into was that the PQauthDataHook is global which
does not work well at all given how Ruby supports multiple sub-instances
of the Ruby runtime (so called Ractors) plus that if you would load e.g.
Python and Ruby into the same process and both would use libpq they
would have to fight over the global hook.
So attached is a proof of concept patch which implements a
per-connection hook. It is just a rough patch to start a discussion. If
people like it I can clean it up and add tests and documentation.
Notes
1. See these two PRs: one which uses a per-runtime map and one which
uses a per-PGConn hash map. I am not a fan of either.
- https://github.com/ged/ruby-pg/pull/693
- https://github.com/ged/ruby-pg/pull/700
--
Andreas Karlsson
Percona
On Thu, Feb 26, 2026 at 3:36 PM Andreas Karlsson <andreas@proxel.se> wrote:
So attached is a proof of concept patch which implements a
per-connection hook. It is just a rough patch to start a discussion. If
people like it I can clean it up and add tests and documentation.
Thank you! So let me start by saying I agree that global is bad, and
per-connection is what we would ideally allow. The original patchset
started out with per-connection settings (for what eventually became
PQAUTHDATA_PROMPT_OAUTH_DEVICE, if I recall correctly).
But I quickly ran into chicken-and-egg problems. We don't expose the
ability to get a connection handle before the connection attempt
starts, nor do we expose the ability to synchronously finish a
connection that's already been started. (So you're locked into
polling-only if you want this.)
You can argue that PQconnectStart() *currently* always returns before
writing its first data, but I feel uncomfortable pinning that
implementation detail for OAuth alone (unless the senior maintainers
are fine with it?). It's already odd, in my mind, that we return
before knowing that polling is even necessary, and I think somebody
might reasonably change that (on purpose, or not) in the future,
especially if PQconnectPoll() ever gets rewritten like I'm hoping.
Not providing a clear "setopt" operation on an empty PGconn is
reminding me of my complaint that we put both application settings and
network settings into the connection string [1]/messages/by-id/CAOYmi+nQu-X7rL3GAB=z2p7qLS4RrNARsz5VSagx=3q+h_03zg@mail.gmail.com, which is reminding me
of Zsolt's complaint that it's difficult to establish security
boundaries for connection settings in the face of environment
variables [2]/messages/by-id/CAOYmi+mVX5rKfvsCVVYM8ZDEBaNDO83TBLPURFuEgA8gbsU8Zw@mail.gmail.com, which recalls earlier discussions with Tom around
sslkeylogfile [3]/messages/by-id/1774813.1736385450@sss.pgh.pa.us. This is like an evil twin to the HBA-setting
problem. I think ideally these _would_ be solved with a single
coherent plan, but is there a way to tackle it incrementally...?
Thanks,
--Jacob
[1]: /messages/by-id/CAOYmi+nQu-X7rL3GAB=z2p7qLS4RrNARsz5VSagx=3q+h_03zg@mail.gmail.com
[2]: /messages/by-id/CAOYmi+mVX5rKfvsCVVYM8ZDEBaNDO83TBLPURFuEgA8gbsU8Zw@mail.gmail.com
[3]: /messages/by-id/1774813.1736385450@sss.pgh.pa.us
On Thu, Feb 26, 2026 at 3:36 PM Andreas Karlsson <andreas@proxel.se> wrote:
- https://github.com/ged/ruby-pg/pull/693
- https://github.com/ged/ruby-pg/pull/700
Related, but a bit tangential: neither proposal there implements
cooperative use of the authdata hook, which is going to make your
problems worse. In particular,
/* TODO: a hook can return 1, 0 or -1 */
is wrong. Zero is reserved for the implementation [1]https://www.postgresql.org/docs/current/libpq-oauth.html#LIBPQ-PQSETAUTHDATAHOOK; you want to
defer to the next link in the chain instead. It's up to the top-level
application (or else, something with *sole* threadsafe control of
libpq) to decide what that chain looks like.
--Jacob
[1]: https://www.postgresql.org/docs/current/libpq-oauth.html#LIBPQ-PQSETAUTHDATAHOOK