>From a recent post I made for the trackrecord list: Status: O > We're currently running dbmail with: > - Dual Xeon 2.4G processors / SCSI Disks > - MySQL/InnoDB > - roughly 90k mailboxes > - 100G of mail data > - alias selection offloaded to smtp servers (we don't use aliases table we deliver straight to userid) > - Load average between 0.4 and 2.5 on a dual processor machine, averaging at around the .8 mark (so say 50% load) > > We have 4 machines handling SMTP/Virus/Spam filtering and then handing it to dbmail-smtp for delviery, and 3 machines > doing POP/IMAP -- all of which are light-moderate loaded.
--------------------- I'd be *really keen* to hear from people with larger installations (even if privately and off list), who have ideas on how dbmail can scale. I need to work out what the next evolution for my cluster will be for the question below, and also just for general organic growth, etc. > I've been posed the question by management lately -- how would dbmail scale if we wanted to offer say 3 Million > mailboxes, rather than just the current limit I see of 100-200k mailboxes using my current setup? > > > Ideas that spring to mind are: > > Clustered databases > ------------------- > Oracle or some database with multiple write masters and > cluser-wide locking for writes springs to mind. I understand > there is problems with MySQL one-way replication and > directing all writes to a master server, but reads from > multiple slaves - due to write-locking not being supported > thus the slave getting out of sync with the master, etc. I > scanned the list again recently but couldn't find any obvious > posts with solutions for this. > > Last minute addition >> I notice SAP DB is now MySQL MaxDB > and is available. Anyone looked at this as a viable backend > for dbmail? Can't for the life of me find the details on > replication support/etc on the site... > > > Partitioned databases > --------------------- > The old customea A-E on database server 1, F-L on database > server 2, etc. This would seem to do the trick, although with > dbmail it may be user_idnr 1-50,000 on DB1, 50,001-100,000 on > DB2 I guess, with some sort of central mailbox -> store index > table. I haven't looked at the code for 2.0 recently, but > from memory 1.x maintained a persistent connection with the > mailstore, thus this could be quite bad if there were say 100 > database servers. I'm not sure what the overhead would be for > connection setup/teardown if doing dynamic connections to > different servers. > > > Any other ideas? (even if not likely to be implemented > officially by dbmail -- may code our own variation if needed). ADDITIONAL QUESTIONS: - Has anyone ported dbmail to Oracle or something that has two-way clustering? - Cheers, Mark.
