Date
Attendees
Fabrice Jammes, Unknown User (kelsey), Andy Salnikov, Unknown User (npease), Fritz Mueller, John Gates, Vaikunth Thukral, Brian Van Klaveren, Andy Hanushevsky, Serge Monkewitz, Jacek Becla
Frossie Economou (for part of the meeting)
Discussion items
DM-3985
- hostname, dns issues. Solved by using NAT, and docker spy
ELK
- latest version (which has useful modules) incompatible with syslogng (which is used on IN2P3 cluster)
- ask IN2P3 to consider switching to new ELK and not using syslogng
Releases
- need a designated person to bless qserv / dax releases - Fritz
- planning to migrate Qserv Release Procedure to a tech note
- issues with readthedocs because of eups (readthedocs needs to build docs, needs setup.py)
- eups tags for qserv release - have two options:
- proper tags displayed, at the expense of extra recompilation if tag changes but code did not
- proper tags not displayed, but we skip recompilation if code did not change
- we opt for a.
- would be better to tag everything, yes, including 3rd party modules
- K-T/Mario would rather not tag 3rd party with monthly or weekly tags
- would be good to not hide the original tag, e.g. if we tag mariadb as 10.6.1.lsst1, we want to see that version, and m.2016.01 would hide it
- and eups manifest can hold only one tag per module
- Frossie will think how to satisfy both needs
- Frossie will think how to satisfy both needs
storage/IO updates
- hardware for czars
- on average we will have merge from 3 machines, 10 Gbits/machine
- disk speed (across all chars) should at least match network speed (so 40Gbits/sec)
- might consider SSD for the results, streaming will happen in small pieces, lots of random IO
- recovery from failures
- reserve % of IO for contingency behavior
- industry standard ~50%. Can't afford
- resource mgmt / throttling query throughput might help
- consider 3 replicas instead of 2. Disk are getting denser
- intermediate results
- small, even at 6 GB/ high vol query x 100 = <1 TB across 100+ machines
- Do a tour of LDM-141 storage/IO spreadsheet for the DB group
anaconda dependencies
- update requests, capture dependencies in .table file before removing anaconda
- flask now packaged together with 3 other dependencies. It is a bit hacky, there is a silent dependency...
prep for JTM
- all: should think what to discuss with the other 5 teams, will reuse the google sheet we used before.