Partial service update

[cpuguy83]

It would be cool to be able to specify an update policy such that I only want to deploy an updated version of my service to a subset of the swarm (to e.g. manual verification that the service is OK).

ping @duglin

[duglin]

A little more info....

Let's say I have 10 instances of my service running at v1 and I want to upgrade to v2.

I ask for a rolling upgrade but only want to upgrade one instance so I can do some testing before I do the rest (this assumes that the normal health checking done by swarmkit isn't sufficient to know if things are really ok). Right now there's no way to ask for 1 to be upgrade and then wait before doing the rest.

One option that @stevvooe mentioned was to allow for us to specify the number (or %) of instances to upgrade and then the user could increase that amount over time until they're all done.

[kelseyhightower]

This is where it might make sense to decouple a task from a service. It would be nice to have a service map to one or more tasks based on something like labels to identify a set of containers that should receive traffic for a service.

If services and tasks where decoupled then you can use a canary pattern to achieve what you want. Run two "jobs" with N number of tasks to reach your target ratio, then have a single service that points to both tasks.

Another possible way to do this is to add the ability to pause a rolling update at a certain % and resume it latter. Pause and resume might provide the right clues to help a user reason about the current state of the task.

[cpuguy83]

@kelseyhightower I think pause/resume (with an automated pause) would achieve this quite well.

[resouer]

I think it's better to be called blue-green deployment which is a standard
function in any PaaS like CF etc, but really not sure it should be included
in swarmkit. It's more like a function build on swarmkit.

On Fri, Jun 24, 2016 at 9:33 PM, Brian Goff notifications@github.com
wrote:

@kelseyhightower https://github.com/kelseyhightower I think
pause/resume (with an automated pause) would achieve this quite well.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#1045 (comment),
or mute the thread
https://github.com/notifications/unsubscribe/ABn3lrD4WmRrT2C29Q0Gys4A2nIUbvOZks5qO9yogaJpZM4I8Q2g
.

[sukrit007]

really looking forward for blue-green kind of deploy. Currently, the only way i can think of doing using swarm mode is to bring up another service (which binds to a different port) and use something on top like cloud-loadbalancer on top for routing traffic to v1 or v2. But would love to see this as part of swarm kit.

[aluzzardi]

Agreed with @resouer

@kelseyhightower I think just the service concept is enough. You can implement blue/green, canary, session stickiness, etc by using services.

For instance, you could deploy service-1 and service-2 then have an haproxy service where you can explicitly do traffic split (e.g. "send 1% of traffic to service-canary, the rest to service-prod).

We could have some abstraction in SwarmKit itself but there are gazillions of use cases (a/b testing?).

The simple service concept allows you do mix and match so that your own code can decide how to split traffic.

[stevvooe]

Here is an example of blue-green deployment with swarm services: https://github.com/stevvooe/sillyproxy.

[pnickolov]

I propose that these can be looked as two separate issues:

1. per @cpuguy83 - ability to manually control the update process and effectively control when the replacement occurs (also see Proposal: support manual control of service update process / canary moby#26160 which seems to a duplicate and opened by me in the wrong project)
2. per @aluzzardi and @stevvooe - ability to change the distribution of traffic between the old (blue) and new (green) instances

Assuming random/round-robin load balancing, #1 also controls the traffic distribution but it is quite heavy (still, seems needed, see also moby/moby#26160 and moby/moby#26159)

#2 can provide a lightweight mechanism to control the traffic distribution (per @aluzzardi ) while not touching the number of instances, especially not destroying old instances that we may end up needing if we have to rollback.

Question: I am about to move moby/moby#26160 (manual control) and moby/moby#26159 (pause/unpause) here. If this issue remains about instances, it will make sense to merge at least moby/moby#26160 here. Agree?

cc/ @aaronlehmann
Thanks @stevvooe for connecting the the issues

[DoomTaper]

Any updates on this ? When will this feature be available ?