It sounds like balance url_param is exactly what i need. In regards to a node going down, hash-type consistent sounds great as i would like to keep my clients using the same node if it is still available.
Thanks for your help! -karl On Sep 15, 2010, at 3:34 AM, Willy Tarreau wrote: > On Mon, Sep 13, 2010 at 08:06:37AM -0400, Karl Baum wrote: >> Hi Willy. The "balance url_param" looks like what I need. >> >> In regards to setting a cookie, in my case each of the http clients is >> actually an email fetching worker calling an imap api which will eventually >> sit behind HAProxy. Because each api node will have a connection pool of >> imap connections, depending on which email address the worker is processing, >> i want the request to be directed to the server which already has a pool of >> connections open for that email address. If i didn't do this, the more api >> nodes behind HAProxy, the more connections i will have open to the imap >> server and imap limits open connections to each email account. Each worker >> will be serving multiple email accounts and workers will process the same >> email account in parallel so i don't think the cookie based routing applies >> to this use case (but i could be wrong). > > OK, so what you describe perfectly matches the typical usage of a URL > parameter hash. > >> Is there significant overhead in the url_param hashing algorithm? > > No, it's still quite cheap. You should not notice it on a mailing system. > >> Will the load be spread evenly across all nodes? > > yes, the hashing algorithm used tends to be quite smooth, but for that > you obviously need a somewhat large population. I'd say that if your > number of possible keys is about 100 times bigger than the number of > nodes, you'll get quite a smooth distribution. > >> Also, in the case of node failure, will the "balance url_param" pick another >> node? > > Yes, that's the same for all algorithms, provided that you enable health > checks of course. The point to note with hashing algorithms is that when > a node goes down, the farm size changes and everyone gets redispatched. > This may or may not be desired. If this is a problem, you could use > "hash-type consistent" to change the hashing algorithm to a consistent > one : only the users of the failed node will be uniformly redispatched > to remaining nodes. The downside is that consistent hashing is not as > smooth at the normal one, but it might still be acceptable in your case. > > Regards, > Willy >
