Re: [PATCH]: Spread checks

On Mon, 15 Oct 2007, Willy Tarreau wrote:

> Hi Krzysztof,

Hi Willy,

> I've merged your spread-checks work. I've brought minor adaptations :

> - I renamed the parameter "spread-checks" instead of "spread-check".
> I realized that it's how I called it everywhere intuitively, so
> it should be better that way.

You are completely right, "spread-checks" is much better.

> - I added a lower bound for the minimal interval. I had a case where
> two backends shared 5 servers with 1000 ms interval, and one backend
> had one server with 100 ms interval. All 11 servers would be started
> within 100 ms, which meant about 45 ms between two consecutive checks
> for the same server in each backend. Now, servers with inter < 1000
> will not affect the "mininter" variable, and will have their start
> date relative to their own inter instead of mininter. It gave fairly
> better results for the case above, as identical servers would then
> be checked 450 ms apart.

Nice solution.

> - I found that the randomization was only applied in the code path where
> a health check had failed. This may explain why it took 45 minutes to
> stabilize on your server. Here, with that fixed, it already stabilizes
> after 2 or 3 checks.

Ahh, sorry for that. I should have tested it more carefully.

> - I added some minimal doc on the new variable, and suggested 2..5 as
> good values since they were those which showed me good results.
Indeed, with above fix 2..5 should be fine.

> All in all, I'm satisfied with this feature, and I've merged it for 1.3.12.3.
> I'll send another mail here on the ML with a preview of the changes.
OK. Thank you.

Best regards,

                                 Krzysztof Olêdzki Received on 2007/10/15 22:25