I've attached a patch set that adds the restore_library,
archive_cleanup_library, and recovery_end_library parameters to allow
archive recovery via loadable modules. This is a follow-up to the
archive_library parameter added in v15 [0] [1].
The motivation behind this change is similar to that of archive_library
(e.g., robustness, performance). The recovery functions are provided via a
similar interface to archive modules (i.e., an initialization function that
returns the function pointers). Also, I've extended basic_archive to work
as a restore_library, which makes it easy to demonstrate both archiving and
recovery via a loadable module in a TAP test.
A few miscellaneous design notes:
* Unlike archive modules, recovery libraries cannot be changed at runtime.
There isn't a safe way to unload a library, and archive libraries work
around this restriction by restarting the archiver process. Since recovery
libraries are loaded via the startup and checkpointer processes (which
cannot be trivially restarted like the archiver), the same workaround is
not feasible.
* pg_rewind uses restore_command, but there isn't a straightforward path to
support restore_library. I haven't addressed this in the attached patches,
but perhaps this is a reason to allow specifying both restore_command and
restore_library at the same time. pg_rewind would use restore_command, and
the server would use restore_library.
* I've combined the documentation to create one "Archive and Recovery
Modules" chapter. They are similar enough that it felt silly to write a
separate chapter for recovery modules. However, I've still split them up
within the chapter, and they have separate initialization functions. This
retains backward compatibility with v15 archive modules, keeps them
logically separate, and hopefully hints at the functional differences.
Even so, if you want to create one library for both archive and recovery,
there is nothing stopping you.
* Unlike archive modules, I didn't add any sort of "check" or "shutdown"
callbacks. The recovery_end_library parameter makes a "shutdown" callback
largely redundant, and I couldn't think of any use-case for a "check"
callback. However, new callbacks could be added in the future if needed.
* Unlike archive modules, restore_library and recovery_end_library may be
loaded in single-user mode. I believe this works out-of-the-box, but it's
an extra thing to be cognizant of.
* If both the library and command parameter for a recovery action is
specified, the server should fail to startup, but if a misconfiguration is
detected after SIGHUP, we emit a WARNING and continue using the library. I
originally thought about emitting an ERROR like the archiver does in this
case, but failing recovery and stopping the server felt a bit too harsh.
I'm curious what folks think about this.
* Іt could be nice to rewrite pg_archivecleanup for use as an
archive_cleanup_library, but I don't think that needs to be a part of this
patch set.
[0] https://postgr.es/m/668D2428-F73B-475E-87AE-F89D67942270%40amazon.com
[1] https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=5ef1eef
--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com
