Re: SSH access to CF app instances on Diego

Guillaume Berche


The CAB minutes [1] mentionned you were still looking for feedback from the
community on the policy for altered instances, but this thread seems silent
for a while.

Not sure you had seen my email and suggestion below for a way to quarantine
the altered instances (beyond the per-space restart policy configuration).
Such quarantine request might be a good place to include option to ask for
the quarantine instances to be excluded from gorouter traffic.




On Fri, Jul 3, 2015 at 3:56 PM, Guillaume Berche <bercheg(a)> wrote:


please find my feedback to this thread

*short version:*
1- need preserve good CF experience with HTTP only (direct SSH flow is
still blocked and a pain in many organisations) => +1 to preserve "cf
files" or fine tune diego plug to have ssh over HTTP to work out of the box
2- default "recycle tainted containers by default" policy seems good to me
3- needs to be completed with more control of the recycling policy (UX
such as "quarantine" or GAE "lock/unlock" )
4- development use-cases need to be better supported (dev/prod parity) not
sure ssh/scp is the right path though

*long version:*

*1- cf files and ssh over HTTP*

As previously mentionned into [1], CF exposing apis over HTTP api made a
great job to be easily consummed through HTTP proxies that some companies
still use, making CF experience seemless to consumme public paas, or
private paas among corporate entities. It seems important to me to preserve
good CF experience with HTTP only.

If SSH interactive access, scp and port forwarding become the mainstream
solution to operate and troubleshoot apps (supporting "cf files",
replacement for the previous DEBUG and CONSOLE ports), it will be useful
for users behind such firewalls to be able to configure diego ssh plugin to
use HTTP/SOCKS proxies to reach public CF instances. As the diego ssh cli
plugin supports using the regular local host ssh binaries, this may
potentially be done by tweaking the .ssh config file to add flags
associated to host ssh.${domain} to go through proxies (possibly double
tunnels as described into [2]). However, for new users in such network
context, especially on windows operating system, the set up work before
using a CF public instance starts to add up?

*2- default "recycle tainted containers by default" seems good to me*

Given that apps deployed on CF comply to 12 factor apps, there instance
may be restarted at anytime (e.g. during a CF new release deployment or
stemcell upgrade). So the default policy "recycle tainted containers by
default" is not a surprise.

*3- need to be completed with more control of the recycling policy (UX
such as "quarantine" or GAE "lock/unlock" )*

There are some specific use-cases where the "recycle tainted containers by
default" policy would be problematic when running applications in

An application instance is malfunctionning (e.g. hanging) and an
interactive debugging is necessary. The app-ops ssh into the container and
starts taking some diagnostic steps (e.g sending kill -SIGTERM signals to
take thread dumps, or locally changes log levels).

If ever the ssh connection breaks/timeout, the "recycle tainted containers
by default, preventing the current diagnostc to complete.

Another similar use case: a production application is suspected to be
compromised by an attacker. App-ops need to capture evidences and
understand better how the abuse was done. There isn't enough information in
streamed logs, and there is a need to get into the container to inspect the
ephemeral FS and the processes and memory. This may require more than one
simultanenous SSH connection, and may span on multiple hours

In both use-cases above, while the application is 12 factor compliant and
the "recycle tainted containers by default" policy would be opted in on the
corresponding space, there would be a need to transiently turn the mode off.

In term of user experience, this may appear as an explicit user request to
"quarantine" the tainted app instances (or the whoe app) so that CF does
not attempt to restart them. Or it may appear as the google app engine

a call to a new "unlock" command to a CF app instance would be necessary
to get SSH access to it. CF then considers this instance as
"tained"/untrusted, as it may have deviated from the pushed content, and
does not act to it anymore (i.e. does not monitor its bound $PORT or root
process exit, which may be handy to diagnose it as wish). When the "lock"
command is requested on this instance, Cf destroys this tainted instance,
and recreates a fresh new "trusted" one.

*4- development use-cases need to be better supported (dev/prod parity)
not sure ssh/scp is the right path though*

I agree with James Myers that development use-cases should be better

First, CF should strive to support dev-prod parity [4]. However currently,
there is not anymore a version of CF that a developper can run on his
laptop (e.g. when doing offline development during commute) that would
behave like prod and embed buildpacks. There used to have "CF on a single
VM". Heroku or GAE have emulators. Cloud rocker [5] is close, but it still
takes 10s or more to have changes made on the app be reflected into a
running app.

There are some legitimate use cases during development for modifying
sources of the application and have those changes be taken in effect
immediately. Lots of app development framework supports those development
modes (even those that promote test-driven practices), and getting a fast
feedback is important. Having dev-prod parity means supporting these use
cases while preserving prod behavior (having the VCAP_SERVICES and
VCAP_APPLICATION and the buildpack processing applied on the same stack
(cflinux2)). Being able to run offline would be even better.

I however believe that providing SSH/SCP access to change the file system
to a running app instance may not be the appropriate response, given the FS
and the app instance is still ephemeral. Who would want to modify files
that could be lost at any time (e.g. a stemcell upgrade ) ?

I'd rather see value in further exploring the ideas layed out by James
Bayer into [5] e.g. as a form of a git repo populated with the
/home/vcap/app subdir, that developers could clone, push to, and have the
instance epheremal FS updated with pushed changes.

This may be combined with a cloudrocker mechanism as to work with a fully
offline mode when this is required.


On Thu, Jul 2, 2015 at 10:18 PM, James Myers <jmyers(a)> wrote:

I have to agree with Matt on this one. I feel that the recycling of
containers is a very anti-developer default. When you approach Cloud
Foundry from the perspective of running production applications the recycle
policy makes complete sense. However, I feel that this misses out on one of
the massive benefits/use cases of Cloud Foundry, what it offers to the
development process.

From a security stand point, if you can ssh into a container, it means
you have write access to the application in CloudFoundry. Thus you can
already push new bits/change the application in question. All of the
"papertrail" functionality around pushing/changing applications exists for
SSH as well (we record events, output log lines, make it visible to users
that action was taken on the application), and thus concerned operators
would be able to determine if someone modifying the application in question.

Therefore I'm lost on how this is truly the most secure default. If we
are really going by the idea that all defaults should be the most secure,
ssh should be disabled by default.

As a developer, I can see many times in which I would want to be able to
ssh into my container and change my application as part of a
troubleshooting process. Using BOSH as an example, CF Devs constantly ssh
into VMs and change the processes running on them in order to facilitate
development. BOSH does not reap the VM and redeploy a new instance when you
have closed the SSH session. Once again this is largely due to the fact
that if you have SSH access, you can already perform the necessary actions
to change the application through different means.

Another huge hindrance to development, is that the recycling policy is
controlled by administrators. It is not something that normal users can
control, even though we allow the granularity of enabling/disabling SSH
completely to the end user. This seems counterintuitive.

I feel that a better solution would be to provide the user with some
knowledge of which instances may be tainted, and then allowing them to opt
into a policy which will reap tainted containers. This provides users with
clear insight that their application instance may be a snowflake (and that
they may want to take action), while also allowing normal behavior with
regards to SSH access to containers.

To summarize, by enabling the recycling policy by default we not only
produce extremely unusual behavior / workflows for developers, we are also
minimizing the developer-friendliness of CF in general. This mixed with the
fact that as a user I cannot even control this policy, leads me to believe
that as a default recycling should be turned off as it provides the most
cohesive and friendly user experience.

On Mon, Jun 29, 2015 at 9:14 AM, John Wong <gokoproject(a)> wrote:

after executing a command, concluding an interactive session, or
copying a file into an instance, that instance will be restarted.

How does it monitor the behavior? Is there a list of commands
whitelisted? I am curious because I am trying to find out what the
whitelist contain. Also is it at the end of the bosh ssh APP_NAME session?
What if two users are there simultaneously?


On Mon, Jun 29, 2015 at 5:49 AM, Dieu Cao <dcao(a)> wrote:

I think with the CLI we could add clarifying messaging when using ssh
what the current policy around recycling is.
Eric, what do you think about calling it the "recycling" policy,
enabled by default? =D


On Sat, Jun 27, 2015 at 3:42 AM, Matthew Sykes <matthew.sykes(a)
Depends on your role and where your app is in the deployment pipeline.
Most of the scenarios I envisioned were for the tail end of development
where you need to poke around to debug and figure out those last few

For example, Ryan Morgan was saying that the Cloud Foundry plugin for
eclipse is going to be using the ssh support in diego to enable debug of
application instances in the context of a buildpack deployed app. This is
aligned with other requirements I've heard from people working on dev tools.

As apps reach production, I would hope that interactive ssh is
disabled entirely on the prod space leaving only scp in source mode as an
option (something the proxy can do).

Between dev and prod, there's a spectrum, but in general, I either
expect access to be enabled or disabled - not enabled with a suicidal

On Thu, Jun 25, 2015 at 10:53 PM, Benjamin Black <bblack(a)>


could you elaborate a bit on what you believe ssh access to instances
is for?


On Thu, Jun 25, 2015 at 9:29 PM, Matthew Sykes <
matthew.sykes(a)> wrote:

My concern is the default behavior.

When I first prototyped this support in February, I never expected
that merely accessing a container would cause it to be terminated. As we
can see from Jan's response, it's completely unexpected; many others have
the same reaction.

I do not believe that this behavior should be part of the default
configuration and I do believe the control needs to be at the space level.
I have have already expressed this opinion during Diego retros and at the
runtime PMC meeting.

I honestly believe that if we were talking about applying this
behavior to `bosh ssh` and `bosh scp`, few would even consider running in a
'kill on taint mode' because of how useful it is. We should learn from that.

If this behavior becomes the default, I think our platform will be
seen as moving from opinionated to parochial. That would be unfortunate.

On Thu, Jun 25, 2015 at 6:05 PM, James Bayer <jbayer(a)>

you can turn the "restart tainted containers" feature off with
configuration if you are authorized to do so. then using scp to write files
into a container would be persisted for the lifetime of the container even
after the ssh session ends.

On Thu, Jun 25, 2015 at 5:50 PM, Jan Dubois <jand(a)>

On Thu, Jun 25, 2015 at 5:36 PM, Eric Malm <emalm(a)>
after executing a command, concluding an
interactive session, or copying a file into an instance, that
instance will
be restarted.
What is the purpose of being able to copy a file into an instance
the instance is restarted as soon as the file has been received?

cf-dev mailing list

Thank you,

James Bayer

cf-dev mailing list

Matthew Sykes

cf-dev mailing list

cf-dev mailing list

Matthew Sykes

cf-dev mailing list

cf-dev mailing list

cf-dev mailing list

cf-dev mailing list

Join to automatically receive all group messages.