Re: Fix checkpoint skip logic on idle systems by tracking LSN progress

On 9/27/16 6:16 AM, Kyotaro HORIGUCHI wrote:

> I apologize in advance that the comments in this message might
> one of the ideas discarded in the past thread.. I might not grasp
> the discussion completely X(
>
> The attached patches are rebased to the master and additional one
> mentioned below.

I tried the attached patch set and noticed an interesting behavior.
With archive_timeout=5 whenever I made a change I would get a WAL
segment within a few seconds as expected then another one would follow a
few minutes later.

Database init:
16M Sep 27 20:05 000000010000000000000001
16M Sep 27 20:09 000000010000000000000002

Create test table:
16M Sep 27 20:13 000000010000000000000003
16M Sep 27 20:15 000000010000000000000004

Insert row into test table:
16M Sep 27 20:46 000000010000000000000005
16M Sep 27 20:49 000000010000000000000006

The cluster was completely idle with no sessions connected in between
those three commands. Is it possible this is caused by:

+ * switch segment only when any substantial progress have made from
+ * the last segment switching by timeout. Segment switching by other
+ * reasons will cause last_xlog_switch_lsn stay behind but it doesn't
+ * matter since it just causes possible one excessive segment switch.
*/

I would like to give Michael a chance to respond to the updated patches
before delving deeper.

--
-David
david(at)pgmasters(dot)net