Re: missing data/global

Daniel Kalchev <daniel(at)digsys(dot)bg> writes:
> Is there ANY chance to recover data from a database system that suffered disk
> crash, and is not missing the data/global directory?
> Version is 7.2.4. Database files seem to be intact as well as pg_clog and
> pg_xlog directories.

The hard part I think would be reconstructing pg_database, because you'd
need to get the database OIDs right. I can't think of any way to do
that that doesn't involve poking at the file with a hex editor.

Here's a sketch of how I'd proceed:

1. Make a tar backup of what you have! That way you can start over
after you screw up ;-)

2. I assume you know the names and properties of your databases, users,
and groups if any; also the SYSID numbers for the users and groups.
A recent pg_dumpall script would be a good place to get this info.

3. You're also going to need to figure out the OIDs of your databases
(the OIDs are the same as the names of their subdirectories under
$PGDATA/base). Possibly you can do this just from directory/file sizes.
Note that template1 should be OID 1, and template0 will have the next
lowest number (probably 16555, in 7.2).

4. Initdb a scratch database in some other place (or move aside your
existing files, if that seems safer). In this scratch DB, create
databases, users, and groups to match your old setup. You should be
able to duplicate everything except the database OIDs using standard
SQL commands.

5. Shut down scratch postmaster, then hex-edit pg_database to insert the
correct OIDs. Use pg_filedump or a similar tool to verify that you did
this properly.

6. Restart scratch postmaster, and VACUUM FREEZE pg_database, pg_shadow,
and pg_group (from any database). This will make the next step safe.

7. Stop scratch postmaster, and then copy over its $PGDATA/global
directory into the old DB.

8. Cross your fingers and start postmaster ...

This will probably *not* work if you had been doing anything to
pg_database, pg_shadow, or pg_group between your last checkpoint and the
crash, because the reconstructed tables are not going to be physically
identical to what they were before, so any actions replayed from WAL
against those tables will be wrong. Hopefully you won't have that
problem. If you do, it might work to shut down the postmaster and again
copy the scratch $PGDATA/global directory into the old DB, thereby
overwriting what the WAL replay did. This is getting into the realm of
speculation though.

regards, tom lane