Horton

Thread

Commits

  1. Fix data loss on crash after sorted GiST index build.

  1. GiST index build missing smgrimmedsync()?

    Melanie Plageman <melanieplageman@gmail.com> — 2022-02-23T21:30:50Z

    I brought this up in [1] but wanted to start a dedicated thread.

Since 16fa9b2b30a357 GiST indexes are not built in shared buffers.
However, smgrimmedsync() is not done anywhere and skipFsync=true is
always passed to smgrwrite() and smgrextend(). So, if a checkpoint
starts and finishes after WAL-logging some of the index build pages and
there is a crash sometime before the dirty pages make it to permanent
storage, won't that data be lost?

(Also, as Heikki has pointed out before, passing skipFsync=false isn't
sufficient either if checkpoint finishes before
register_dirty_segment(), so seems like smgrimmesync() is needed.)

Seems like the following would address this:

diff --git a/src/backend/access/gist/gistbuild.c
b/src/backend/access/gist/gistbuild.c
index 9854116fca..364a7455ea 100644
--- a/src/backend/access/gist/gistbuild.c
+++ b/src/backend/access/gist/gistbuild.c
@@ -454,9 +454,11 @@ gist_indexsortbuild(GISTBuildState *state)
        smgrwrite(RelationGetSmgr(state->indexrel), MAIN_FORKNUM,
GIST_ROOT_BLKNO,
                          pagestate->page, true);
        if (RelationNeedsWAL(state->indexrel))
+       {
                log_newpage(&state->indexrel->rd_node, MAIN_FORKNUM,
GIST_ROOT_BLKNO,
                                        pagestate->page, true);
-
+               smgrimmedsync(RelationGetSmgr(state->indexrel), MAIN_FORKNUM);
+       }
        pfree(pagestate->page);
        pfree(pagestate);
 }

- Melanie

[1] https://www.postgresql.org/message-id/CAAKRu_ZmkCNz-%3D5U5wZyyi3%3DUsfs1btsTKc_L9r1ceFvbxE3kg%40mail.gmail.com

  2. Re: GiST index build missing smgrimmedsync()?

    Heikki Linnakangas <hlinnaka@iki.fi> — 2022-02-24T14:32:53Z

    On 23/02/2022 23:30, Melanie Plageman wrote:
> I brought this up in [1] but wanted to start a dedicated thread.
> 
> Since 16fa9b2b30a357 GiST indexes are not built in shared buffers.
> However, smgrimmedsync() is not done anywhere and skipFsync=true is
> always passed to smgrwrite() and smgrextend(). So, if a checkpoint
> starts and finishes after WAL-logging some of the index build pages and
> there is a crash sometime before the dirty pages make it to permanent
> storage, won't that data be lost?

Yes, good catch!

> Seems like the following would address this:

Committed essentially that, except that I put the smgrimmedsync in a 
separate if-block, and copied the comment from the similar piece of code 
from nbtsort.c to explain the issue.

Thanks!

- Heikki