On Tue, May 21, 2019 at 4:17 AM Michael Paquier <michael@paquier.xyz> wrote:
> > 2. Suppose the system reaches the end of
> > heapam_relation_set_new_filenode and then the system crashes. Because
> > of the smgrimmedsync(), and only because of the smgrimmedsync(), the
> > init fork would exist at the start of recovery. However, unlike the
> > heap, some of the index AMs don't actually have a call to
> > smgrimmedsync() in their *buildempty() functions. If I'm not
> > mistaken, the ones that write pages through shared_buffers do not do
> > smgrimmedsync, and the ones that use private buffers do perform
> > smgrimmedsync.
>
> Yes, that maps with what I can see in the code for the various empty()
> callbacks.
>
> > Therefore, the ones that write pages through
> > shared_buffers are vulnerable to a crash right after the unlogged
> > table is created. The init fork could fail to survive the crash, and
> > therefore the start-of-recovery code would again fail to know that
> > it's dealign with an unlogged relation.
>
> Taking the example of gist which uses shared buffers for the init
> fork logging, we take an exclusive lock on the buffer involved while
> logging the init fork, meaning that the checkpoint is not able to
> complete until the lock is released and the buffer is flushed. Do you
> have a specific sequence in mind?
Yes. I thought I had described it. You create an unlogged table,
with an index of a type that does not smgrimmedsync(), your
transaction commits, and then the system crashes, losing the _init
fork for the index.
There's no checkpoint involved in this scenario, so any argument that
it can't be a problem because of checkpoints is necessarily incorrect.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
