Hello Ryohei-san,
I have further remarks after Kirk-san extensive review on these patches.
* About TCP interface v18.
For homogeneity with the surrounding cases, ISTM that "TCP_user_timeout"
should be ""TCP-user-timeout".
* About TCP backend v19 patch
I still disagree with "on other systems, it must be zero.": I do not see
why "it must be zero", I think that the parameter should simply be ignored
if it does not apply or is not implemented on a platform?
If there are consistency constraint with other timeout parameters,
probably the documentation should mention it?
* About socket_timeout v12 patch, I'm not sure there is a consensus.
I still think that there should be an attempt at cancelling before
severing.
Robert pointed out that it is not a timeout wrt the query, but this is not
clearly explained in the documentation nor the comments. The doc says that
it is the time for socket read/write operations, but it is somehow the
time between messages, some of which may not be linked to read/write
operations. I feel that the documentation is not very precise about what
it really does.
ISTM that the implementation could make the cancelling as low as 1 second
because of rounding. This could be said somewhere, maybe in the doc,
surely in a comment.
I still think that this parameter should be preservered on psql's
reconnections when explicitely set to non zero.
--
Fabien.