Print

Print


I shared the public part of LDM-135 (http://ls.st/64m) with
the mysql team at Facebook, below is Mark's reaction. I think
these are all very good points and we should pay close attention
to the feedback from them -- these who know Mark and his team
know how good they all are (facebook uses mysql very seriously,
they are the largest mysql user to the best of my knowledge)

I want to discuss some of the issues Mark brought up during
our qserv mtg on Thursday.

As some of you know, we will see that team next week at XLDB,
and we plan a face to face meeting with them on 9/20

Jacek


-------- Original Message --------
Subject: 	Re: title, abstract, bio needed for xldb-2013
Date: 	Tue, 3 Sep 2013 13:21:40 -0700
From: 	MARK CALLAGHAN
To: 	Jacek Becla, Domas Mituzas, Harrison Fisk



I read the report that you shared. It was interesting to read how you
decided on technology. My guess is that you will spend most of your time
responding to failures (bad CPU, RAM, disk) and performance stalls. The
other way to look at that is that this solution will distinguish itself
(look great) more from your ability to manage failures and performance
problems than from the underlying database software. A common problem
will be performance outliers where a few servers are slow but the source
of the problem could be network, disk, configuration, old software, too
much load, etc.

Do you plan on using MyISAM or Aria? AFAIK MyISAM doesn't have per-block
checksums which make it easier to detect corruption and that can be
extremely useful. Most (or all) of our corruption comes from things
under MySQL (not InnoDB but disks, disk firmware, RAID cards). I guess
you will see the same given the use of MyISAM is read-only after being
loaded.

Do you have a solution for aggregate performance monitoring to track
load per table, per user, etc? MariaDB has some of that but your notion
of table & user might be at a higher level.

Do you have a solution for per-access performance monitoring to track
which parts of the system contribute to latency for a specific query
submitted to the system?


On Fri, Aug 30, 2013 at 11:38 AM, MARK CALLAGHAN <[log in to unmask]
<mailto:[log in to unmask]>> wrote:

     I am in the Facebook Menlo Park office on Friday, September 20 if
     you want to meet there for 1 hour. Hopefully Domas will attend and I
     can invite 1 or 2 other people. Harrison tends to have many more
     meetings than Domas and I so I won't predict whether he can attend.


     On Thu, Aug 29, 2013 at 1:55 AM, Jacek Becla
     <[log in to unmask] <mailto:[log in to unmask]>> wrote:

         Mark,

         Yes we could get the discussion going at xldb, but I tend
         to be very very busy during the event, and I think having
         a quiet conversation with white board etc might be better.

         Thanks,
         Jacek



         On 08/27/2013 12:01 AM, MARK CALLAGHAN wrote:



             On Tue, Aug 27, 2013 at 2:10 AM, Jacek Becla
             <[log in to unmask] <mailto:[log in to unmask]>
             <mailto:[log in to unmask]
             <mailto:[log in to unmask]>>> wrote:

                  We briefly talked several times I believe. Our needs
             are somewhat
                  different, we are myisam shop, with read only data, 
complex
                  analysis involving full scans, joins, near neighbor
             search etc


                  I'd love to have my group sit down for a day with fb
             mysql team
                  and discuss! Maybe we will be able to collaborate? We
             have a small
                  team of pretty good developers here making mysql
             scalable for
                  our needs... Who should I talk to realize such meeting?


             If you get some of Domas, Harrison & Yoshinori at such a
             meeting then
             that would be good. Given that Jeremy, Domas and Harrison
             will be at
             XLDB do you want to set aside an hour to meet with them?

########################################################################
Use REPLY-ALL to reply to list

To unsubscribe from the QSERV-L list, click the following link:
https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=QSERV-L&A=1