Re: proposal: rounding up time value less than its unit.

On 9/23/14, 1:21 AM, David Johnston wrote:

This patch should fix the round-to-zero issue.  If someone wants to get rid of zero as a special case let them supply a separate patch for that "improvement".

I am going to wander into this fresh after just reading everything once (but with more than a little practice with real-world GUC mucking) and say that, no, it should not even do that.  The original complaint here is real and can be straightforward to fix, but I don't like any of the proposals so far.

My suggestion:  fix the one really bad one in 9.5 with an adjustment to the units of log_rotation_age.  That's a small commit that should thank Tomonari Katsumata for discovering the issue and even suggesting a fix (that we don't use) and a release note; sample draft below.  Stop there.

We could just change the units for log_rotation_age to seconds, then let the person who asks for a 10s rotation time suffer for that decision only with many log files.  The person who instead chooses 10ms may find log rotation disabled altogether because that rounds to zero, but now we are not talking about surprises on something that seems sane--that's a fully unreasonable user request.

Following that style of fix all the way through to the sort of release notes needed would give something like this:

log_rotation_age is now set in units of seconds instead of minutes. Earlier installations of PostgreSQL that set this value directly, to a value in minutes, should change that setting to use a time unit before migrating to this version.

If I were feeling ambitious about breaking configurations with a long-term eye toward improvement, I'd be perfectly happy converting everything on this list to ms.  We live in 64 bit integer land now; who cares if the numbers are bigger? 

Then the rounding issues only impact sub-millisecond values, making all them squarely in nonsense setting land.  Users will be pushed very clearly to always use time units in their postgresql.conf files instead of guessing whether something is set in ms vs. seconds.  Seeing the reaction to a units change on log_rotation_age might be a useful test for how that sort of change plays out in the real world.