It "geneerally works if" you " rebuild docker hosts on a daily or more frequent basis."

Perhaps I'm misunderstanding, but needing to rebuild my prod env several times a day seems pretty "not ready for prime time" to me.

That's like when we'd say that Rails ran great in production in 2005, as long as you had a cron task to bounce fastCGI processes every hour or so.

So, can you elaborate on why rebuilding the containers is good advice?

The theory runs that attackers need time to accrue and compound their incomplete positions into a successful compromise.

But if you keep patching continuously, attackers have fewer vulnerabilities to work with. If you keep rotating keys frequently, the keys they do capture become useless in short order. And if you rebuild the servers frequently, any system they've taken control of simply vanishes and they have to start from scratch.

I'm not completely sold on the difference between repair and repave, myself. And I expect that sophisticated attackers will begin to rely more on identifying local holes and quickly encoding those in automated tools so that they can re-establish their positions after a repaving happens.

But it raises the cost for casual attackers, which is still worthy.

[0] https://medium.com/built-to-adapt/the-three-r-s-of-enterpris...

The rest: not so much.

Rebuilding continuously for security is not something I would recommend.

So that I understand, could you elaborate?

Particularly, do you mean "not recommend" as in "recommend against" or "not worth the bother"?

It's not crazy to periodically rotate keys, but attackers don't acquire keys by, you know, stumbling over them on the street or picking them up when you've accidentally left them on the bar. They get them because you have a vulnerability --- usually in your own code or configuration. Rebuilding will regenerate those kinds of vulnerabilities. Attackers will reinfect in seconds.

The win to rotating them is not so much because you'll be regularly evicting attackers you didn't know had your keys, but because when you do have a fire, you won't be finding out for the first time that you can't actually rotate them.

It also forces you to design things much more reliably which helps continuity in non-security scenarios.

After redeploying and realizing that Todd has to ssh in and hand edit that one hostname and fix a symlink that was supposed to be temporary so the new version of A can talk to B, that's going to get rolled in pretty quickly. Large operations not doing this tend to quickly end up in the "nobody is allowed to touch this pile of technical debt because we don't know how to re-create it anymore" problem.

After a bunch of harrowing experiences with clients, I'm pretty close to believing "using packages for critical infrastructure is a bad idea".

In that case, you might be interested in bosh: http://bosh.io/docs/problems.html (the tool that enables the workflow jacques_chester was describing). It embraces the idea of reliably building from source for the exact reasons you've mentioned.

- Any infrastructure with lots of data. Data just takes time to move; backups take time to restore.

- You're on bare metal because running node on VMs isn't fast enough.

- You're in a secure environment, where the plain old bureaucracy will get in the way of a full rebuild.

- Anytime you have to change DNS. That's going to take days to get everything failed over.

- Clients (or vendors) whitelist IPs, and you have to work through with them to fix the IPs.

- Amazon gives you the dreaded "we don't have capacity to start your requested instance; give us a few hours to spin up more capacity"

> Imagine an outage in one of the 3 datacenters you are running your infra in the same region. You need to move 1/3 of the capacity to the remaining 2 datacenters.

Oh, this is very different. If your provider loses a datacenter, and your existing infrastructure can't handle it, you're already SOL - the APIs for spinning up instances and networking is going to be DDOSed to death by all of the various users.

Basic HA dictates that you provision enough spare capacity that a DC (AZ) can go down and you can still serve all of your customers.

I used to work in the team that runs Amazon.com. All of the systems serving the site can be re-built within hours and nothing can serve the site that cannot be rebuilt within a very thing SLA. However, I understand that not all the companies have this requirement. This feature is only relevant when a site downtime hurting the company too much, so it could not be allowed.

Reflecting to your points:

- Lots of data -> use S3 with de-normalized data, or something similar

- Running a VM has 3% overhead in 2016, scalability is much more important than a single node performance

- High security environments are usually payment processing systems, downtime there can be a bit more tolerated, delaying transactions is ok

- Amazon uses DNS for everything, even for datacenter moves. It is usually done within 5 minutes

- This is a networking challenge, using something like EIP (where the public facing IP can be attached to different nodes) makes this a non-issue

- Amazon has an SLA, they extremely rarely have a full region outage, so you can juggle capacity around

Losing a dc out of 3 does not require work because you can't handle the load, it is required to have the same properties (same extra capacity for example) just like before. Spinning up instances should not DDOS anything, it is with constant load on the supporting infrastructure.

The last point I agree with.

> Lots of data -> use S3 with de-normalized data, or something similar

S3's use case does not match up with many different computing models (hadoop clusters, database tables, state overflowing memory), and moving data within S3 between regions is painful. Also, not all cloud providers have S3.

> Running a VM has 3% overhead in 2016, scalability is much more important than a single node performance

Not when you have a requirement to respond to _all_ requests in under 50ms (such as with an ad broker).

> High security environments are usually payment processing systems

Or HIPPA, or government.

> delaying transactions is ok

Not really. When I worked for Amazon, they were still valuing one second of downtime at around $13k in lost sales. I can't imagine this has gone down.

> Amazon uses DNS for everything, even for datacenter moves. It is usually done within 5 minutes

Amazon also implements their own DNS servers, with some dynamic lookup logic; they are an outlier. Fighting against TTL across the world is a real problem for DR type scenarios.

> EIP (where the public facing IP can be attached to different nodes) makes this a non-issue

EIPs are not only AWS specific, but they can not traverse across regions, and rely on AWS' api being up. This is not historically always the case.

> they extremely rarely have a full region outage, so you can juggle capacity around

Not always. Sometimes, you can. But not always. Some good examples from the past - anytime EBS had issues in us-east-1, the AWS API would be unavailable. When an AZ in us-east-1 went down, the API was overwhelmed and unresponsive for hours afterwards.

> Spinning up instances should not DDOS anything, it is with constant load on the supporting infrastructure.

See above. There's nothing constant about the load when there is an AWS outage; everyone is scrambling to use the APIs to get their sites backup. There's even advice to not depend on ASGs for DR, for the very same reason.

AWS is constantly getting better about this, but they are not the only VPS provider, nor are they themselves immune to outages and downtime which requires DR plans.

Exactly. Don't put data in Docker. Files go in an object store, databases need to go somewhere else.

OP's first point is 'don't put data in docker'. Docker is not for your data. But more to the point, if you're rebuilding your data store a couple of times every day, a couple of hours downtime isn't going to be feasible.

> You're on bare metal because running node on VMs isn't fast enough

In such a situation, you should be able to image bare metal faster than 2 hours. DD a base image, run a config manager over it, and you should be done. Small shops that rarely bring up new infra wouldn't need this, but anyone running 'bare metal to scale' should.

> bureaucracy

Isn't part of the infra rebuild per se.

> Anytime you have to change DNS. That's going to take days

Depends on your DNS timeouts, but this is config, not infra. Even if it is infra, 48-hour DNS entries aren't a best-practice anymore (and if you're on AWS, most things default to a 5 min timeout)

> Clients (or vendors) whitelist IPs, and you have to work through with them to fix the IPs

I'd file this under 'bureaucracy' - it's part of your config, not part of your prod infra (which the GP was talking about).

> Amazon gives you the dreaded...

Well, yes, but this is on the same order as "what if there's a power outage at the datacentre". Every single deploy plan out there has an unknown-length outage if the 'upstream' dependencies aren't working. "What if there's a hostage event at our NOC?" blah blah.

The point is that with upstream working as normal, you should be able to cover the common SPOFs and get your prod components up in a relatively short time.

I agree, but I (and the GP, from my reading) was not speaking about only Docker infrastructure.

> Isn't part of the infra rebuild per se.

I can see your point, and perhaps these points don't belong in a discussion purely about rebuilding instances discussion. That said, I have a very hard time focusing just on the time it takes to rebuilding capacity when discussing a DC going down; there's just too many other considerations that someone in Operations must consider.

When I have my operations hat on, I consider a DC going down to be a disaster. Even if the company has followed my advice and the customers do not notice anything, we're now at a point where any other single failure will take the site down. It's imperative to get everything taken down with that DC back up; and it's going to take more than an hour or two.

Many larger companies can't do this; my company has 70+ datacenters with tens of thousands of servers. We can't re-build our prod infra in minutes or hours. We are still doing devops right :D

Like I said, I know you aren't talking about my situation when you made your statement... I just get frustrated when people act like there are hard and fast rules for everyone.

I work at MindGeek, depending on the time of the year it would be fair to say we rank within the top 25 bandwidth users in the world. We are not even close to that amount of servers and we deal with some of the largest traffic in the world. What company is running in 70+ datacenters!? World's largest VPN provider? Security company providing all the data to the NSA?

Maybe it is just my broad assumptions but I would hope that the major big 10 that come to mind such as Google, Amazon, Microsoft, etc would be able to rebuild their production regions in hours.

I don't work for one of those two, but I do work for a very large CDN.

"I'd have thought I've had heard of them..."

One Wiki search later... yup. I've heard of them.

People running edge networks and therefore need servers local to everywhere in the world to keep latencies down. Maybe it's not as much of an issue for MindGeek (the parent company for a lot of video streaming sites). I would guess you guys need a lot of throughput but latency isn't so much of a problem. Or you simply don't need to serve some parts of the world where it might be illegal to distribute some types of content.

FWIW, Cloudflare has 86 data centers: https://www.cloudflare.com/network-map/

Short version: That's a video streaming websites which is rather simple, yet bandwidth intensive.

Outsourcing the caching and video delivery means MindGeek can do with little servers and a few locations.

Nonetheless the CDN you're outsourcing to does need a lot of servers at many edge locations.

Actually, if we think in terms of "top bandwith users in the world". It's possible that your company is far from being in the list. It's likely dominated by content delivery / ISP / and other providers, most of which are unknown to the public.

I am in distributed systems and try to work exclusively on hard problems. So when I say "simple", that is biased on the high end of the spectrum.

If you go to pornhub.com and looks at "popular keyword", you'll only find thousand or ten thousands videos. In a way, there is not that much content on pornhub.

All major websites have challenge. Pornhub is a single purpose website and a lot of the challenge is in video delivery, which can be outsourced to a CDN nowadays.

"simple" is maybe too strong a word. I am trying to convey the idea that it has limited scope and [some of] the problems it's facing are understood by now and have [decent] solutions.

That's not to say it's easy ;)

Edit: didn't notice someone else had already said this, opened this tab like 15 minutes ago

Sure, you want to be able to deploy quickly. But if there's no reason to, then don't.

And I would be very scared if Docker images had a 1 day uptime max

We also see them mostly in non-prod environments where we have greater container/image churn. We use AWS autoscale and Fleet so containers just get moved to other hosts when we terminate them. We have actually thought about scheduling a logan's run type job that kills older hosts automatically - it's in the backlog.

While I sincerely hope I'm wrong, I assume it's because you reset the clock on the probability something goes very wrong.

I am currently fighting an ongoing battle at work to point out that the plans for our Mesos cluster have not factored in that the first outage we have will be when someone fills up the 100gb OS SSD because no one's given any thought to where the ephemeral container data goes.

It also naturally rewards optimizing around time-to-redeploy, probably a lot of benefits there.

That's like "my car works fine as long as you spray WD-40 into the engine block every 48 hours..."

These are not necessarily container centric.

Also hosts shouldn't need to be rebuilt THAT often. Of course, your infrastructure should automatically nuke and replace failing hosts (and there's various ways to establish what's "failing"), and like any good infrastructure team you should be keeping your hosts up to date with security patches, etc, which is best done by replacing them with new versions (as then you can properly test and apply a CI lifecycle), but docker itself is stable enough these days that you don't need to be that aggressive with nuking.

Agree (especially with the comments about orchestrating with kube/mesos) with all the other points, though.

"you don't use it to store data" - What is wrong with Docker volumes specifically? What issues did you run into?

"wrap docker pulls with 'flock' so they don't deadlock" - I have never had a problem with docker pull, can you elaborate on this?

I got bit hard by this issue:

https://github.com/docker/docker/issues/20079

when CoreOS stable channel updated Docker. All my data volume containers broke during the migration and migration could not be reverted.

As for the second issue: when two systemd services that pull containers with the same base images would start simultaneously, they would deadlock pretty reliably. Had to flock every pull as a result. This might be fixed by now though.