[erlang-questions] How do you go to production?

Thu Apr 18 00:10:25 CEST 2019

Hi Pierre,

It's early days yet, so I'm still shopping for ideas and am much inspired to explore some presented so far on this thread.

My system has some 25 services bundled into one big Erlang Nitgrogen app with at least one, if not more, connected through  WAN (out-sourced Discourse). 

Some of the services are public with high read/low write; others are low read/low write controlled access but require maximum possible data availability and integrity. I'm exploring and intend to test various ways of factoring these services into containers.

I'd love to hear more about your experience with Kubernetes. Scares me a bit since my devops skills are so deficient.

Many thanks,

Lloyd

-----Original Message-----
From: "Pierre Fenoll" <pierrefenoll@REDACTED>
Sent: Wednesday, April 17, 2019 5:31pm
To: "Sölvi Páll Ásgeirsson" <solvip@REDACTED>
Cc: "Dieter Schön" <dieter@REDACTED>, "Lloyd R. Prentice" <lloyd@REDACTED>, erlang-questions@REDACTED
Subject: Re: [erlang-questions] How do you go to production?

Piping in since I'm not seeing our deployment procedure mentioned here (I may have missed emails cause Gmail seems to dislike this ML server). 
We push a git tag following semver (a plug-in that'd do something like Helm does: automatic decision on which part of major minor patch to bump would be interesting). 
This triggers GitlabCI jobs that runs tests and builds docker images at the same time (so QA/hitfixes don't have to wait for tests, cause skipping tests isn't easy). 
Docker images are built in stages. The release tarball is unpacked into the base OTP image and weights ~40MB compressed.
It's nice to keep this size small so kubernetes can pull it quick. 
We always keep at least two instances running in prod behind the load balancer so we can deploy progressively. 
This only handles 20 requests second though. 
Tagging is a manual action. 
We don't use relups and in fact I've never worked at a company that did. 
It seems canary deployments are enough most of the time. Easier than thinking of the state changes a relup involves. 
In this perspective I think that the "embedded mode" should be the default mode of development. It'd help catch things earlier and maybe allow for more aggressive optimisations. 
Also we're running elixir. 

On Wed 17 Apr 2019 at 22:42, Sölvi Páll Ásgeirsson <[ solvip@REDACTED ]( mailto:solvip@REDACTED )> wrote:
We do a couple of things:
- package some apps as RPMs for running as metal
- package things as docker containers to be scheduled by ECS
Generally, we just use ‘rebar3 as prod release’ and copy the resulting release into the package.  For RPMs, we use fpm.

I suspect it gets little attention because it’s extremely simple to make a prod ready release

Sent from my iPhone

On 17 Apr 2019, at 00:15, Lloyd R. Prentice <[ lloyd@REDACTED ]( mailto:lloyd@REDACTED )> wrote:

Hi Dieter,
Really helpful!  
From my poking around, Erlang deployment gets too little attention on the web or in the literature— particularly in this age of cloud deployment, containers, and distributed systems. So thanks again for your nuts-and-bolts overview.
I’d love to read how others have brought Erlang apps and websites into production and lessons learned.
Best wishes, 
Lloyd

Sent from my iPad

On Apr 16, 2019, at 4:24 PM, Dieter Schön <[ dieter@REDACTED ]( mailto:dieter@REDACTED )> wrote:

Hi Lloyd,
this week I am not in the office, so I can just try to do a brain dump.
The installation was quite simple, I installed a plain Ubuntu desktop,
and then extracted the erlang tar file into /opt. This is done by one or two shell scripts.

The erlang app is started on demand by another application, which is launched by the 
operator.
So I do not have to deal with systemd (which I also do not like very much) or the like.
My application could run as a server, but in the current use case there is no need for that.

For the footprint of the docker container I will have to come back next week.
I really have no idea how big it is. In former times I have used VirtualBoxes with snapshots
for similar tasks, but if the focus is quite narrow, a docker container is far more easier and quicker
than a complete VM.

Regards,
Dieter

On 16.04.19 21:54, [ lloyd@REDACTED ]( mailto:lloyd@REDACTED ) wrote:
Hi Dieter,

Many thanks for sharing your experience.

Could you please outline your installation procedure?  Also, the size of your Docker application image (RAM footprint) ? I've been considering LXC containers, but I'd like to test with Docker.

All the best,

Lloyd

-----Original Message-----
 From: "Dieter Schön" [ <dieter@REDACTED> ]( mailto:dieter@REDACTED )
 Sent: Tuesday, April 16, 2019 3:46pm
 To: "Erlang-Questions Questions" [ <erlang-questions@REDACTED> ]( mailto:erlang-questions@REDACTED )
 Subject: Re: [erlang-questions] How do you go to production?

 Hi,
here is a case from the other end of the spectrum.
I developed a protocol converter for a satellite test system, it is installed on 2 (in words: two)  PCs.
For development, I used one application and rebar3. Production is just "rebar3 as prod tar" (From a labelled git commit).
Testing was done on the two PCs, where I used one as the system under test and the other as test harness/data generator.
 I was also using wireshark and other third party tools, to prevent incestuous behaviour.
To test the installation procedure I used docker, which was really nice. I had a blank machine in two seconds, 
where I could load and execute the installation script from the host. I had turnaround cycles from several seconds.
What else.. the application is quite small. It fits into one erlang application. Apart from it, the release only contains the observer and sasl.
Unit testing was done in EUnit.

This was my first project where I used erlang, and the learning curve was quite gentle. 
Kind regards,
Dieter

Am So., Apr. 14, 2019 11:20 schrieb Max Lapshin [ <max.lapshin@REDACTED> ]( mailto:max.lapshin@REDACTED ):

Hi.
I'm developing Flussonic for almost 10 years and we have some practices for packaging, deploying, running, maintainig that are not well known, however they are rather good for us.
It is interesting how do other people solve this.
At first, we do not use releases. It is because when we have started our long path, using of releases was not very easy.  Couple of days ago I've tested it with rebar3 and it is really easier to use. Perhaps we would use releases today.
Next: we use our own fpm script replacement for packaging. I can boast that it seems to be the only existing implementation of rpm outside of original library, however I'd better never pass this path again. I've written it in pre-docker era and frankly speaking it is a traumatic experience. However, we are using it for debian and it is really convenient:
[ https://github.com/flussonic/epm ]( https://github.com/flussonic/epm )
Some time ago we have switched to systemd. I personally consider systemd a very badly designed thing that was created without any discussions with existing system adminstrators. For example, systemd doesn't offer config validation before launch. Another brilliant idea is to offer libsystemd for linking into application. Unknown library with unknown quality. What can go wrong if you link it into your erlang or java application?
Sorry, but no. We have systemd.erl:  [ https://gist.github.com/maxlapshin/01773f0fca706acdcb4acb77d91d78bb ]( https://gist.github.com/maxlapshin/01773f0fca706acdcb4acb77d91d78bb )

Use Type=notify in  youdaemon.service
After you manage to launch your erlang daemon, you need to collect statistics. We had to add some more linux-related tools to fetch: cpu usage, disk I/O usage, system ram usage (swap, etc), per-interface network statistics, udp errors count, nvidia card usage, etc. 
If this is worthy outsource, I think we can extract it. 
Our os_stat library is linked with our in-erlang pulsedb library. We try to maintain as less dependencies as possible, so we collect all ticks from monitoring tools inside erlang library pulsedb:  [ https://github.com/pulsedb/pulsedb ]( https://github.com/pulsedb/pulsedb )   (maybe should update public branch)
It can save several thousands metrics with one tick per 1-3 seconds.
Support is an important part of our business, because customers cannot just launch software, they often need help.  We have many people in support staff and I do not want to manage their own public ssh keys on customer's servers. So we have written an ssh proxy: system that login to customer server with one private key and allow support guy to use his own key:  [ https://github.com/flussonic/ssh-proxy ]( https://github.com/flussonic/ssh-proxy )
All these things are rather useless for development and many of them are not required for in-house development, however it is hard to live without them when you sell software.
What is your experience with such things that standard erlang lacks?

_______________________________________________erlang-questions mailing list[ erlang-questions@REDACTED ]( mailto:erlang-questions@REDACTED )[ http://erlang.org/mailman/listinfo/erlang-questions ]( http://erlang.org/mailman/listinfo/erlang-questions )
_______________________________________________
erlang-questions mailing list
[ erlang-questions@REDACTED ]( mailto:erlang-questions@REDACTED )
[ http://erlang.org/mailman/listinfo/erlang-questions ]( http://erlang.org/mailman/listinfo/erlang-questions )
_______________________________________________
erlang-questions mailing list
[ erlang-questions@REDACTED ]( mailto:erlang-questions@REDACTED )
[ http://erlang.org/mailman/listinfo/erlang-questions ]( http://erlang.org/mailman/listinfo/erlang-questions )_______________________________________________
 erlang-questions mailing list
[ erlang-questions@REDACTED ]( mailto:erlang-questions@REDACTED )
[ http://erlang.org/mailman/listinfo/erlang-questions ]( http://erlang.org/mailman/listinfo/erlang-questions )-- 

Cheers,
-- 
Pierre Fenoll
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20190417/1173f160/attachment.htm>