Upgrading Executable on the Fly

https://nginx.org/en/docs/control.html#upgrade

by pantuza • 5 years ago

208 71 5 years ago

mholt5• 5 years ago

We did this for Caddy 1 too [1]. It was really cool. I am not sure how many people used this feature, so I haven't implemented it for Caddy 2 yet, and in the ~two years that Caddy 2 has been released, I've only had the request once. It's a bit tricky/tedious to do properly, but I'm willing to bring it over to Caddy 2 with a sufficient sponsorship.

[1] https://github.com/caddyserver/caddy/blob/v1/upgrade.go

JesseObrien• 5 years ago

Can you explain any of the technical details around this perchance? I'm super curious. I know that SO_REUSEPORT[1] exists but is that the only little trick to make this work? From what I've read with SO_REUSEPORT it can open up that port to hijacking by rogue processes, so is that fine to rely on?

[1] https://lwn.net/Articles/542629/

duskwuff• 5 years ago

You don't even need that. If the old server process exec()s the new one, it can pass on its file descriptors -- including the listening socket -- when that happens.

tyingq5• 5 years ago

it can open up that port to hijacking by rogue processes

That seems relevant if the process is using a non-privileged port that's >= 1024. If we're talking about privileged ports (<= 1023), though, only another root process could hijack that, and those can already hijack you many other ways.

fragmede• 5 years ago

If an attacker is already running rogue processes on your box, the minor details surrounding SO_REUSEPORT is the least of your worries. An attacker could just restart nginx, and won't care about lost requests.

khc• 5 years ago

poked around a bit of that from a previous job, here's what I remember:

1. there's a control process and worker processes

2. on upgrade, control process launches new worker processes from the new binary

3. requests are drained from old worker processes

4. most of the time nginx request handlers allocate from a per-request allocation pool, so requests mostly don't share memory

5. for the cases where there are global states, there's a separate shared memory pool that you need to allocate from (which is kind of hard to work if you are not using built-in nginx primitives)

eliaspro2• 5 years ago

I'm torn on this feature.

In the one hand, an application should never be able to replace itself with "random code" to be executed. I want my systems to be immutable. I want my services to be run with the smallest set of privileges required.

On the other hand, it encourages "consumer level" users to keep their software up-to-date, even when it wasn't installed from a distribution's repository etc.

So I think in general it's a good feature to have, as advanced users/distributions will restrict what a service/process is able to to anyways and won't have any downsides of not using this feature.

It should be optional, that's all!

mholt5• 5 years ago

an application should never be able to replace itself with "random code" to be executed.

To clarify: it doesn't, nor has it ever worked that way. You have to be the one to do that (or someone with privileges to write to that file on disk). Most production setups don't give Caddy that permission. And you have to trigger the upgrade too.

blibble• 5 years ago

if you can log into the machine and replace the nginx executable you are probably capable of running it too

zimbatm4• 5 years ago

If Caddy were to support systemd socket activation, this self-restart dance is not necessary as the parent process (systemd) is holding the socket for you. And for other systems, they can use https://github.com/zimbatm/socketmaster instead. I believe this to be more elegant and robust than the nginx approach as there is no PID re-parenting issues.

But I suspect that most Caddy deployments are done via docker, and that requires a whole container restart anyways.

mholt5• 5 years ago

Good point, and I'm not sure which deployment method is more popular.

In general I am personally not a fan of Docker due to added complexities (often unnecessary for static binaries like Caddy) and technical limitations such as this. All my Caddy deployments use systemd (which I don't love either, sigh).

tyingq5• 5 years ago

It's kind of fun to watch things go out of fashion and back in. We used to use inetd, mostly because memory was expensive, so it could spawn a service only when a request came in, then the spawned process would exit and give the memory back to the os. Then someone decided tcpd should sit between inetd and servers, for security and logging. Then, every service just ran as it's own daemon. Now I'm occasionally seeing posts like this reviving inetd.

nextaccountic• 5 years ago

If I'm running on containers, does this mean that I need to run systemd inside docker?

kevincox• 5 years ago

I think one caveat is that systemd stops the old instance before starting the new one. This will give you a latency blip as requests will be queued during that time. The nginx approach doesn't introduce any latency impact.

latchkey• 5 years ago

Someone did this for golang. It isn't perfect, but works for some basic use cases... https://github.com/jpillora/overseer

bragr• 5 years ago

I've implemented this a few times in a few languages based on exactly what nginx does. It works well, and it is pretty straight forward if you are comfortable with posix style signals, sockets, and daemons.

I'm not sure it is super critical in the age of containerized workloads with rolling deploys but at the very least the connection draining is a good pattern to implement to prevent deploy/scaling related error spikes.

wereHamster2• 5 years ago

Even with containerized workloads, you still have an ingress, or SPOF (or multiple, when using multicast), and the seamless restart is meant for exactly those processes. Nginx is often used (https://kubernetes.github.io/ingress-nginx/), or when you use AWS, GCS etc they provide such a service for you.

Not sure how the cloud providers do it though, maybe combination of low DNS TTL and rolling restart since they often have huge fleets of servers which handle ingress?

nullify885• 5 years ago

A container though should be immutable and ideally shouldn't have changes made to it. If the container were to die, it'd revert back to the old version? It looks to me like these seamless upgrades would be an anti pattern to containers.

With ingress you'd have a load balancer in front or have it routed in the network layer using BGP.

Grollicus• 5 years ago

If you really want to have no SPOF you'd probably build something like this:

Multihomed IP <-> Loadbalancer <-> Application

By having the same setup running on multiple locations you can replace the load banacers by taking one location offline (stop announcing the corresponding route). Application instances can be replaced by taking the application instance out of the load balancer.

xyzzy_plugh4• 5 years ago

Not sure how the cloud providers do it though

At least some I am familiar with operate at the packet level and can hand off live "connections" to a peer or hot standby, along with full session state.

Remember that with TCP or anything else, the abstract session is an illusion.

rootlocus• 5 years ago

How do the two processes listen to the same port?

loeg• 5 years ago

There’s an ioctl for this on FreeBSD and Linux — SO_REUSEPORT. You could also just leave the listening socket open when exec’ing the new httpd, or send it with a unix domain socket.

ctrlrsf• 5 years ago

Using socket option SO_REUSEPORT allows multiple processes to bind to same port.

VWWHFSfQ• 5 years ago

is this what it's actually doing though? It doesn't say the reuseport option to the listen directive is required for this.

riyadparvez• 5 years ago

Is there any restrictions on this option? Eg only children of the same parent process are allowed to bind to same port. Otherwise how does the packet distribution work? And how does the response from that port work?

markbnj• 5 years ago

This article on how haproxy uses SO_REUSEPORT goes into some more detail: https://www.haproxy.com/blog/truly-seamless-reloads-with-hap...

nullify885• 5 years ago

Once the USR2 signal is received the master process forks, the child process inherits the parents file descriptors including listen(). One process stops accepting connections creating a queue in the kernel. The new process takes over and starts accepting connections.

You can follow the trail by searching for ngx_exec_new_binary in the nginx repo.

krab4• 5 years ago

Just to add - Nginx normally spawns several worker processes that all process connections to the same port.

nullify885• 5 years ago

The systemv init script for nginx had an upgrade operation (in addition to start/stop/reload etc) which would send the signal. Worked like a charm.

moderation• 5 years ago

See Envoy Proxy's Hot Restart [0]

0. https://www.envoyproxy.io/docs/envoy/latest/intro/arch_overv...

monroeclinton2• 5 years ago

Also HAProxy, they both use UNIX sockets via ancillary messages+SCM_RIGHTS I believe.

https://www.haproxy.com/blog/truly-seamless-reloads-with-hap...

secondcoming• 5 years ago

I've always found the multi-process approach taken by both nginx and apache to be nothing but a hindrance when you have to write a custom module. It means that you may have to use shared memory, which is a PITA.

I don't know why they haven't moved on from it; it only really made sense when uni-core processors were the norm.

majke• 5 years ago

Just a shout out: it's super hard to do it for UDP / QUIC / H3. Beware.

(but I don't think nginx supports h3 out of the box yet)

krab4• 5 years ago

Why so? I thought UDP was stateless, making that process even easier. But I never implemented it.

TimWolla2• 5 years ago

UDP itself is stateless, but QUIC itself is stateful. Without knowing the background I would assume the issue to be that the incoming UDP packets will be routed to the new process after the reload and that new process is not aware of the existing QUIC connections, because the state resides in the old process. Thus it is not able to decrypt the packets for example.

petters• 5 years ago

How are quick/http3 servers usually upgraded? As you say, it seems tricky.

xyzzy_plugh4• 5 years ago

It's fundamentally identical. How do you think one hands off a TLS session? Or any other state?

It's only a problem if your state is tangled and impossible to serialize or bundle up to hand off.

UDP is perhaps the easiest because there's nothing to do in the basic case, for example with DNS.

jabedude• 5 years ago

Seems like a useful feature for a service manager like systemd to have for its managed services. It is already able to perform inetd style socket activation, I imagine this would be a welcome feature

TimWolla2• 5 years ago

This is already possible. You can configure whether you want inetd style socket activation (where systemd calls accept() and passes you the client socket)), or just systemd listening to the socket (where systemd passes you the listen socket and your binary calls accept()).

https://www.freedesktop.org/software/systemd/man/systemd.soc...

jasonjayr• 5 years ago

inetd style socket activation (iirc) forks a process for every connection.

So, simply replacing the binary on disk will cause all new connections going forward to use the new binary, while existing held connections (with in-memory references to the old binary's inode) will finish the operations. Once they are done and all references to that inode are gone, the blocks referencing the binary will be removed.

wahern• 5 years ago

inetd style socket activation (iirc) forks a process for every connection.

inetd supports both process-per-connection and single process/multiple connections using the "nowait" and "wait" declarations, respectively. The former passes an accept'd socket, the latter passes the listening socket.

tannhaeuser• 5 years ago

CGI (apache's mod_cgi) has supported this since almost the beginning of the web as well. Deploying a new CGI is as simple as replacing the CGI binary.

politelemon• 5 years ago

So if I understood correctly, would it be like this

cp new/nginx /path/to/nginx kill -SIGUSR2 <processid>

That does sound pretty neat if you're not running nginx in a container. I wonder if they've built a Windows equivalent for that.

bogomipz2• 5 years ago

I am curious does anyone know why Nginx uses SIGWINCH for this? I know Apache uses WINCH as well which makes me wonder if there was some historical reason a server process wound up using a signal meant for a TTY?

xyzzy_plugh4• 5 years ago

It's otherwise unused in the context of a daemon. SIGHUP is similarly popular and historically designed for use with a TTY.

bogomipz2• 5 years ago

That's what I suspected - that it would pretty much be guaranteed to not to be needed however I was not able to find any historical UNIX lore mentioning it specifically. Cheers.

monroeclinton2• 5 years ago

I've been working on something similar in a load balancer I've been writing in Rust. It's still a work in progress.

Basically the parent executes the new binary after it receives a USR1 signal. Once the child is healthy it kills the parent via SIGTERM. The listener socket file descriptor is passed over an environment variable.

https://github.com/monroeclinton/- (this is the proper url, it's called dash)

bob1029• 5 years ago

I've considered building something like this to allow for us to update customer software while it's serving users.

In my proposals, there would be a simple application-aware http proxy process that we'd maintain and install on all environments. It would handle relaying public traffic to the appropriate final process on an alternate port. There would be a special pause command we could invoke on the proxy that would buy us time to swap the processes out from under the TCP requests. A second resume command would be issued once the process is running and stable. Ideally, the whole deal completes in ~5 seconds. Rapid test rollbacks would be double that. You can do most of the work ahead of time by toggling between an A and B install path for the binaries, with a third common data path maintained in the middle (databases, config, etc)

With the above proposal, the user experience would be a brief delay at time of interaction, but we already have some UX contexts where delays of up to 30 seconds are anticipated. Absolutely no user request would be expected to drop with this approach, even in a rollback scenario. Our product is broad enough that entire sections of it can be a flaming wasteland while other pockets of users are perfectly happy, so keeping the happy users unbroken is key.

kayodelycaon• 5 years ago

https://en.wikipedia.org/wiki/Blue-green_deployment

DNS not required. You can use a load balancer to do the same thing. If you don't want a full second setup, do a rolling restart of application servers instead.

Edit: I forgot... you can do this with containers too.

dang• 5 years ago

One thread with one comment from a long time ago:

Upgrading an Nginx executable on the fly - https://news.ycombinator.com/item?id=8677077 - Nov 2014 (1 comment)

kokizzu2• 5 years ago

I use Overseer https://github.com/jpillora/overseer

Upgrading Executable on the Fly

Related Stories