Short update: using local puppet patches, I got the syslog forwarding working. Next step will be getting those patches into the production branch.

@Bstorm thank you! I presume the project will be created shortly? Let me know if there is anything I can do.

In T127717#6859176, @Bstorm wrote:

So central logging is a thing we've been very interested in for Cloud VPS projects. A strictly security-focused syslog server would have value even if it was just for cloud admins. If it's something more like a multitenant logging solution that somehow implements good multitenancy in Elastic's stack without their enterprise offering or a multitenant Grafana Loki system (which would be simpler on the log aggregation end but not-so-simple at the Grafana end without "enterprise" features since you'd need to have it set up as multi-org) that'd be quite a project to dig into! Loki is nice because it is not slated to stop being Apache 2 licensed, but it is also not deployed at Wikimedia at all yet (though Grafana obviously is).

Yes, Grafana Loki is one of the newer solutions. I am not experienced with Grafana Loki, but it may be the best solution for extracting metrics from logs, although I am not sure if Grafana Loki fits any high availability / scalability requirements. Elastic is more known, but with the license issues, it's better to let that slide for now (if the software is forked, we can reconsider!). What do you think?

@Andrew / @bd808, thank you for all of your comments. I wouldn't call myself the expert of WMCS and ELK, but I do have experience with syslog and log management, albeit with other tools (unfortunately nothing OSI-approved anymore, since Graylog has the SSPL license now as well).

As the original reporter of T127656, could you use my help here? Central logging sounds like a fun, but useful side project.

@MoritzMuehlenhoff understood. Patch set 4 will use the custom unit (with /run) on systems older than bullseye. On bullseye systems we'll be using the built-in unit.

In T252990#6408248, @ema wrote:

This still uses the legacy /var/run though, hence it does not fix the issue.

Correct, see this:

I have uploaded a new patch using /run on all servers (regardless of OS). However, what about removing the unit and using the built-in one? That reduces the likelihood of units conflicting with intended behavior from upstream. An apt-get source for nagios-nrpe=3.2.1-2 shows the following unit (debian/nagios-nrpe-server.service):

[Unit]
Description=Nagios Remote Plugin Executor
Documentation=http://www.nagios.org/documentation
After=var-run.mount nss-lookup.target network.target local-fs.target remote-fs.target time-sync.target
Before=getty@tty1.service plymouth-quit.service xdm.service
Conflicts=nrpe.socket

@Ciencia_Al_Poder What do you mean with internal URLs? purgeList.php fetches the URLS using Title->getCdnUrls(), which seems to work fine in 1.34.2:

The PID directory has been changed to /run since Buster: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=932353#10

In T260739#6403512, @Dzahn wrote:

@Southparkfan Thank you very much for this detailed ticket! Everything went fine going through the check boxes.

You're welcome.

Except I wanted to ask where you got the 308 from specifically.

I got that from T256276. However, https://gerrit.wikimedia.org/r/c/operations/puppet/+/612442/3/modules/dynamicproxy/files/domainproxy.lua says "ngx.HTTP_MOVED_PERMANENTLY", which is a 301. Either one is fine, though.

Definitely not working on this. Thanks for the help!

I feel my point was completely missed, regardless of the cause or fix: this change has a performance impact of roughly 28x. And since I have new questions, reopening.

It doesn't seem to be limited to broken citations, {{Special:WhatlinksHere/{{FULLPAGENAME}}}} also seems enough to reproduce this.

GDNSD supports HTTP health checks (or custom ones as well). Provided the DNS TTL is no longer than a few minutes, 'automatic failover' is possible.

The tasks regarding loss of PSU redundancy on cp303[2689] are normal priority, does this one need to be high priority?

(correct datacenter)

The same issue is occurring again at a high rate, and is causing problems. Also, there are users reporting they are silently being logged out of the site. I'm not sure how much they are related to each other, but I spotted this in the debug logs:

[session] Session "[50]CentralAuthSessionProvider<-:2:Southparkfan>[REDACTED]": Metadata merge failed: [Exception MediaWiki\Session\MetadataMergeException(
 /srv/mediawiki/w/includes/session/SessionProvider.php:195) Key "CentralAuthSource" changed]
#0 /srv/mediawiki/w/includes/session/SessionManager.php(629): MediaWiki\Session\SessionProvider->mergeMetadata(array, array)
#1 /srv/mediawiki/w/includes/session/SessionManager.php(498): MediaWiki\Session\SessionManager->loadSessionInfoFromStore(MediaWiki\Session\SessionInfo, WebRequest)
#2 /srv/mediawiki/w/includes/session/SessionManager.php(182): MediaWiki\Session\SessionManager->getSessionInfoForRequest(WebRequest)
#3 /srv/mediawiki/w/includes/WebRequest.php(700): MediaWiki\Session\SessionManager->getSessionForRequest(WebRequest)
#4 /srv/mediawiki/w/includes/session/SessionManager.php(121): WebRequest->getSession()
#5 /srv/mediawiki/w/includes/Setup.php(747): MediaWiki\Session\SessionManager::getGlobalSession()
#6 /srv/mediawiki/w/includes/WebStart.php(137): require_once(string)
#7 /srv/mediawiki/w/index.php(40): require(string)
#8 {main}

Is this the same as the recent behavior seen on various API appservers? https://ganglia.wikimedia.org/latest/?c=API%20application%20servers%20eqiad&m=cpu_report&r=hour&s=by%20name&hc=4&mc=2

I wonder why @RobH and @Cmjohnson are talking about 4TB disks. The current problems are caused by 4k (6TB) disks, and the accessoires link given by Cmjohnson mentions a non-4k 6TB disk.

I wonder why @RobH and @Cmjohnson are talking about 4TB disks. The current problems are caused by 4k (6TB) disks, and the accessoires link given by Cmjohnson mentions a non-4k 6TB disk.

Do we prefer a fallback that cannot be impacted by a Wikimedia outage of any kind? Conpherence is an option, but it is not off-site; a network outage affecting Wikimedia will render Conpherence useless.

In T141482#2509001, @Tgr wrote:

Possibly your farm is not sharing sessions, which is a requirement for CentralAuth. Session storage configuration has changed for 1.27, you might want to review that.

It could be possible that my configuration is not up-to-date. Like Wikimedia, we use redis for sessions (production 1.26 config):

$wgObjectCaches['redis'] = array(
	'class' => 'RedisBagOStuff',
	'servers' => array( '<IP>' ),
	'password' => $wmgRedisPassword,
);
$wgMainCacheType = 'redis';
$wgSessionCacheType = 'redis';
$wgSessionsInObjectCache = true;
$wgMessageCacheType = CACHE_NONE;
$wgParserCacheType = CACHE_DB;
$wgLanguageConverterCacheType = CACHE_DB;

Yes, it does.

Step 1 seems to be re-introducing the texvc package (whose deletion actually was unrelated when I created the issue, it just made this task more or less invalid, because the issue was kinda gone - since nobody (except users who still have that package installed) would be hit by it anymore). That is a blocker for this task.

@Andrew labvirt1012 lacks hyperthreading. Can you enable that?

mw1017 and mw1099 are the eqiad MW debug hosts, so those should not be in conftool-data.
mw1161-mw1169 were previous appservers, but have been converted to jobrunners in December, per rOPUPba0a47b56ded3dc748c436c2940f114389b312f2. Jobrunners are not present in conftool-data, so I guess everything is as expected?

@Kghbln yes, I am aware of that, and I spoke to @MoritzMuehlenhoff, and he said that in hindsight the mediawiki-math-texvc should never have been deleted. Perhaps it will be packaged again - but I'm not sure if Debian's policies allow that, since it would mean that it has to be introduced in the jessie (stable) suite again. Otherwise we'd still need to wait until Debian 9 has been released.

@Joe mw1306 seems to have the same IP as mw1091 (although that one shows up as mw1091 in Ganglia, whereas mw1090 shows up as mw1305...). So I think you need to fix that one as well.

Just tested this, this doesn't happen in Firefox 47.0 on Windows 8.1. Perhaps an upgrade to 47.0 might help, although I doubt it because I can't see in the 47.0 release notes anything related to this.

Yep, that change (and the maintenance script) worked perfect. Everything is displaying correctly now.

cp1044 has been decomissioned per T133614

In T130130#2353909, @awight wrote:

That would be great! The only additional benefit we might get from HHVM
would be to cut PHP cpu load by 50%, but that can be an optimization for
later. This card is concerned with the breaking language changes, so
Jessie is a perfect solution.

As of PHP 5.5 (Jessie has 5.6), there is an opcache extension included (only needs to be enabled in config), which caches compiled bytecode in the memory. Together with the other performance improvements since 5.4 (assuming paymentswiki runs on 5.3 now), CPU usage / response time should already decrease without using HHVM (so you get some more headroom before HHVM is really needed).

Jessie has PHP 5.6 actually.

I'm sure Legoktm is right :-)! But I failed to make it work, so someone else should take it over honestly...

It would be cool if this bug could be fixed (whatever the solution is).

@Paladox I've already upgraded to HHVM 3.12, which seems to work.

@Joe did you already install one of those servers? noticed https://ganglia.wikimedia.org/latest/?c=Application%20servers%20eqiad&h=mw1305.eqiad.wmnet&m=cpu_report&r=hour&s=by%20name&hc=4&mc=2 (I can't view which specs the new servers have, but this 'mw1305' seems to have the same cpu/ram specs as the old appservers?)

@Cmjohnson yeah, perhaps I have been a bit too fast by already doing the DNS part (despite that's the only thing I can do it seems) :-)

Just to learn how the process works, I've submitted a patch for the DNS adjustments. I noticed db1058 is referenced in the dhcpd and manifests/role/coredb.pp files in puppet but I have no idea how the latter one works, so I'll leave the puppet work to someone else.

Can this patch be backported to REL1_26, REL1_25 and REL1_23 too? The current security patch is only available in master, and these three branches still receive security support.

Yeah, this doesn't seem to be a Varnish problem:

@ori that is T124163 actually.

@Andrew mw1138 is not depooled (anymore), its CPU and network graphs show it is serving traffic. Looking at http://ganglia.wikimedia.org/latest/graph.php?r=day&z=xlarge&h=mw1132.eqiad.wmnet&m=cpu_report&s=by+name&mc=2&g=network_report&c=API+application+servers+eqiad it was idle for a few hours, but then it somehow got repooled..

I have a setup where I tell Varnish to redirect all /.well-known/acme-challenge traffic to one backend server (practically any server with a webserver should work), so I can use this (acme-tiny).