Cyrus mail spool entirely on SSD

[Jason L Tibbitts III]

>>>>> "MU" == Michael Ulitskiy via Info-cyrus <info-cyrus at lists.andrew.cmu.edu> writes:

MU> My question is: assuming cost is not an issue, would it be advisable
MU> to put an entire cyrus mail spool on SSD?  I'm thinking about
MU> combinining multiple SSDs into RAID10 to further increase
MU> performance.

I don't see why not.  You certainly will want redundancy and will need
to monitor the health of storage regardless of whether you store your
bits in polysilicon or magnetized rust.

Of course the failure modes and rates are completely different; disks
don't really care how much you write to them while SSDs will have
basically a fixed write quantity.  Once you hit that, your SSD is
generally out of warranty regardless of its age.

And, of course, there are all sorts of complicated interactions because
SSDs are still relatively new and things like TRIM and what happens to
read data after it has been trimmed are still kind of up in the air.

Of course, if cost really is no object, go all NVMe, use the latest
Intel datacenter grade SSDs, enjoy the incredible speed and drives rated
for multiple petabytes written.

One thing of importance is that the Cyrus mailbox index files get
rewritten a whole lot.  I don't know if anyone has tried to quantify the
write load of a mail spool for purposes of predicting SSD replacement
intervals.

MU> What do you think?

I think if I had the money I'd have done it already.  I have several
all-SSD machines, but not yet my Cyrus server.  For cost containment, I
would probably go with a mixed NVMe-SATA system, putting the hottest
bits on NVMe (and, of course, tmpfs for things which can disappear) and
the spool on SATA, still with plenty of redundancy.

 - J<