Re: CLUSTER not lose indexes

Bruce Momjian dijo:

> FYI, RENAME also deals with relfilenode renaming in setNewRelfilenode().
> The difference is that RENAME doesn't need to access the old index, just
> build a new one, so it can take shortcuts in how it handles things. It
> uses two methods to modify the tuple, one directly modifying pg_class,
> rather than inserting a new heap row. and another of doing it the
> traditional way.  It does this because REINDEX is used to fix corrupt
> indexes, including corrupt system indexes.  You will not be using that
> type of code in CLUSTER because there is a real temp heap associated
> with this operation.  Just heap_update() like normal for both relations.

Well, I think my approach is somewhat more naive.  What I'm actually
doing is something like:

1.  save information on extant indexes
2.  create a new heap, and populate (clustered)
3.  swap the relfilenodes of the new and old heaps
4.  drop new heap (with old relfilenode)
5.  for each index saved in 1:
5.1.  create a new index with the same attributes
5.2.  swap relfilenodes of original and new index
5.3.  drop new index (with old relfilenode)

But now I'm lost.  It has worked sometimes; then I change a minimal
thing, recompile and then it doesn't work (bufmgr fails an assertion).
I can tell that the new (table) filenode is correct if I skip step 4
above, and check the temp table manually (but of course it has no
indexes).  I've never gotten as far as completing all steps right, so I
cannot tell whether the new indexes' filenodes are correct.

I'm posting the new patch.  Please review it; I've turned it upside down
a few times and frankly don't know what's happening.  I sure am
forgetting a lock or something but cannot find what is it.

--
Alvaro Herrera (<alvherre[a]atentus.com>)
"Aprender sin pensar es inutil; pensar sin aprender, peligroso" (Confucio)