Dean, Robert,
Afaics the problem described below was introduced in b4e07417, do you
have a different/better proposal than
s/CacheInvalidateSmgr/CacheInvalidateRelcache/? Doing that doesn't feel
quite right either, because it only makes the file extension visible at
end-of-xact - which is mostly harmless, but still.
On 2016-04-23 19:51:17 -0700, Andres Freund wrote:
> It fixes the problem at hand, but essentially it's just luck that the
> patch is sufficient. The first layer of the issue is that queued
> invalidation messages aren't sent; but for vm_extend() there's another
> side to it. Namely vm_extend() does
>
> /*
> * Send a shared-inval message to force other backends to close any smgr
> * references they may have for this rel, which we are about to change.
> * This is a useful optimization because it means that backends don't have
> * to keep checking for creation or extension of the file, which happens
> * infrequently.
> */
> CacheInvalidateSmgr(rel->rd_smgr->smgr_rnode);
>
> but CacheInvalidateSmgr is non-transactional as it's comment explains:
> *
> * Note: because these messages are nontransactional, they won't be captured
> * in commit/abort WAL entries. Instead, calls to CacheInvalidateSmgr()
> * should happen in low-level smgr.c routines, which are executed while
> * replaying WAL as well as when creating it.
> *
>
> as far as I can see vm_extend() is the only current caller forgetting
> that rule. I don't think it's safe to use CacheInvalidateSmgr() outside
> smgr.c.
>
> The reason this all ends up working nonetheless is that the
> heap_inplace_update()s in vacuum triggers a CacheInvalidateHeapTuple()
> which queues a relcache invalidation, which in turn does a
> RelationCloseSmgr() in RelationClearRelation(). Thereby effectively
> doing a transactional CacheInvalidateSmgr(). But that seems more than
> fragile.
>
> ISTM we should additionally replace the CacheInvalidateSmgr() with a
> CacheInvalidateRelcache() and document that that implies an smgr
> invalidation. Alternatively we could log smgr (and relmapper)
> invalidations as well, but that's not quite non-invasive either; but
> might be a good long-term idea to keep things simpler.
- Andres
