Re: Doppler zoning query


john mcteague <john.mcteague@...>
 

- From etcd I see 5 unique entries, all 5 doppler hosts are listed with
the correct zone
- All metron_agent.json files list the correct zone name
- All doppler.json files also contain the correct zone name


All 5 doppler servers contain the following two errors, in varying amounts.

{"timestamp":1432671780.232883453,"process_id":1422,"
source":"doppler","log_level":"error","message":"AppStoreWatcher: Got error
while waiting for ETCD events: store request timed
out","data":null,"file":"/var/vcap/data/compile/doppler/loggregator/src/
github.com/cloudfoundry/loggregatorlib/store/app_service_store_watcher.go
","line":78,"method":"
github.com/cloudfoundry/loggregatorlib/store.(*AppServiceStoreWatcher).Run"}

{"timestamp":1432649819.481923819,"process_id":1441,"
source":"doppler","log_level":"warn","message":"TB: Output channel too
full. Dropped 100 messages for app f744c900-d82d-4efc-bbe4-
004e94ffdfec.","data":null,"file":"/var/vcap/data/compile/
doppler/loggregator/src/doppler/truncatingbuffer/
truncating_buffer.go","line":65,"method":"doppler/truncatingbuffer.(*
TruncatingBuffer).Run"}

For the latter, given the high log rate of the test app, it suggests I need
to tune the buffer of doppler, but I dont expect this to be the cause of my
cpu imbalance.

On Tue, May 26, 2015 at 5:08 PM, John Tuley <jtuley(a)pivotal.io> wrote:

John,

Can you verify (on, say one runner in each of your zones) that Metron's
local configuration has the correct zone? (Look in
/var/vcap/jobs/metron_agent/config/metron.json.)

Can you also verify the same for the Doppler servers
(/var/vcap/jobs/doppler/config/doppler.json)?

And then can you please verify that etcd is being updated correctly? (curl
*$ETCD_URL*/api/v2/keys/healthstatus/doppler/?recursive=true with the
correct ETCD_URL - the output should contain entries with the correct IP
address of each of your dopplers, under the correct zone.)

If all of those check out, then please send me the logs from the affected
Doppler servers and I'll take a look.

– John Tuley

On Tue, May 26, 2015 at 9:26 AM, <cf-dev-request(a)lists.cloudfoundry.org>
wrote:



Message: 2
Date: Tue, 26 May 2015 16:26:30 +0100
From: john mcteague <john.mcteague(a)gmail.com>
To: Erik Jasiak <ejasiak(a)pivotal.io>
Cc: cf-dev <cf-dev(a)lists.cloudfoundry.org>
Subject: Re: [cf-dev] Doppler zoning query
Message-ID:
<CAEduAK4WmMfrhdhxWDfpR=
Ot0eM+yspsswqx4hG36Mte0bS9kg(a)mail.gmail.com>
Content-Type: text/plain; charset="utf-8"


We are using cf v204 and all loggregators are the same size and config
(other than zone).

The distribution of requests across app instances is fairly even as far as
I can see.

John.
On 26 May 2015 06:21, "Erik Jasiak" <ejasiak(a)pivotal.io> wrote:

Hi John,

I'll be working on this with engineering in the morning; thanks for
the details thus far.

This is puzzling: Metrons do not route traffic to dopplers outside
their zone today. If all your app instances are spread evenly, and all
are
serving an equal amount of requests, then I would expect no
major variability in Doppler load either.

For completeness, what version of CF are you running? I assume your
configurations for all dopplers are roughly the same? All app instances
per
AZ are serving an equal number of requests?

Thanks,
Erik Jasiak

On Monday, May 25, 2015, john mcteague <john.mcteague(a)gmail.com> wrote:

Correct, thanks.

On Mon, May 25, 2015 at 12:01 AM, James Bayer <jbayer(a)pivotal.io>
wrote:

ok thanks for the extra detail.

to confirm, during the load test, the http traffic is being routed
through zones 4 and 5 app instances on DEAs in a balanced way.
however the
dopplers associated with zone 4 / 5 are getting a very small amount
of load
sent their way. is that right?


On Sun, May 24, 2015 at 3:45 PM, john mcteague <
john.mcteague(a)gmail.com>
wrote:

I am seeing logs from zone 4 and 5 when tailing the logs (*cf logs
hello-world | grep App | awk '{ print $2 }'*), I see a relatively
even
balance between all app instances, yet doppler on zones 1-3 consume
far
greater cpu resources (15x in some cases) than zones 4 and 5.
Generally
zones 4 and 5 barely get above 1% utilization.

Running *cf curl /v2/apps/guid/stats | grep host | sort *shows 30
instances, 6 in each zone, a perfect balance.

Each loggregator is running with 8GB RAM and 4vcpus.


John

On Sat, May 23, 2015 at 11:56 PM, James Bayer <jbayer(a)pivotal.io>
wrote:

john,

can you say more about "receiving no load at all"? for example, if
you restart one of the app instances in zone 4 or zone 5 do you see
logs
with "cf logs"? you can target a single app instance index to get
restarted
with using a "cf curl" command for terminating an app index [1].
you can
find the details with json output from "cf stats" that should show
you the
private IPs for the DEAs hosting your app, which should help you
figure out
which zone each app index is in.
http://apidocs.cloudfoundry.org/209/apps/terminate_the_running_app_instance_at_the_given_index.html

if you are seeing logs from zone 4 and zone 5, then what might be
happening is that for some reason DEAs in zone 4 or zone 5 are not
routable
somewhere along the path. reasons for that could be:
* DEAs in Zone 4 / Zone 5 not getting apps that are hosted there
listed in the routing table
* The routing table may be correct, but for some reason the routers
cannot reach DEAs in zone 4 or zone 5 with outbound traffic and
routers
fails over to instances in DEAs 1-3 that it can reach
* some other mystery

On Fri, May 22, 2015 at 2:06 PM, john mcteague <
john.mcteague(a)gmail.com> wrote:

We map our dea's , dopplers and traffic controllers in 5 logical
zones using the various zone properties of doppler, metron_agent
and
traffic_controller. This aligns to our physical failure domains in
openstack.

During a recent load test we discovered that zones 4 and 5 were
receiving no load at all, all traffic went to zones 1-3.

What would cause this unbalanced distribution? I have a single app
running 30 instances and have verified it is evenly balanced
across all 5
zones (6 instances in each). I have additionally verified that
each logical
zone in the bosh yml contains 1 dea, doppler server and traffic
controller.

Thanks,
John

_______________________________________________
cf-dev mailing list
cf-dev(a)lists.cloudfoundry.org
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev


--
Thank you,

James Bayer

--
Thank you,

James Bayer
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <
http://lists.cloudfoundry.org/pipermail/cf-dev/attachments/20150526/31789891/attachment.html
------------------------------

_______________________________________________
cf-dev mailing list
cf-dev(a)lists.cloudfoundry.org
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev


End of cf-dev Digest, Vol 2, Issue 73
*************************************

_______________________________________________
cf-dev mailing list
cf-dev(a)lists.cloudfoundry.org
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev

Join cf-dev@lists.cloudfoundry.org to automatically receive all group messages.