We use Janus as a WebRTC SFU for projects in the education/gaming sector. Its general purpose approach to WebRTC has been a good foundation to help us build custom solutions.
Did you write your own plugin or are you using VideoRoom or something? The thing that turned me off of Janus is that VideoRoom seemed way too high level and made tons of schema assumptions, but all of the important SFU functionality seemed intertwined into the codebase of that plugin, so if I were to actually want to build a "custom solution" it looked like I would have to maintain an annoying-to-merge fork of that thick layer :/.
Janus itself doesn't make any assumptions about a specific use-case. All the functionality outside of the core RTC stuff is implemented with plugins. The default "Video Room" implementation is just a Lua script [0]. Mozilla has written their own SFU plugin (in Rust) [1] for game networking that powers Mozilla Hubs [2].
No: the vast majority of the VideoRoom functionality is written in C, and it is where all of the actually-hard-to-do video SFU stuff -- like quality control feedback and SVC support, particularly in an end to end encryption context with all the codec-specific workarounds--is commingled together with the notion of "rooms" (which is a really awkward and specific high-level abstraction with a schema that you have to abuse for basic use cases).
Yes: if you don't want any of the complex video functionality, you can easily write your own Janus plugin, and that maybe sounds reasonable for some trivial game "SFU" where you are just going to move around some data channel packets... but at that point you can (and I argue should) just use libwebrtc (I do this, and I helped one of my friends do this for his product in a weekend: people act like it is hard to compile but it really isn't).
(Even more so: the Lua script you linked to looks more like a demo/example of a way to use the Lua plugin to get some functionality vaguely similar to the VideoRoom plugin, and it is notably ridiculously long and contains a lot of codec-specific knowledge, while not having anywhere near the actual functionality of the actual real C VideoRoom plugin. It is as if Janus is just a super low-level WebRTC library in the form of a framework, with an explicitly monolithic plugin doing everything.)
At that point why wouldn't I just use libwebrtc? The reason to get an off-the-shelf SFU is because all the hard work is in handling all the codec-specific workarounds, being able to handle keyframe request sharing, responding to RTCP bandwidth feedback to do SVC layer switching, and now doing all of this while most of the state is encrypted due to insertable streams... this is all hard stuff that people keep learning more about and for which the state of the art is a moving target due to browser changes.
I would expect 100% of applications doing anything at all with video to want all of that functionality, but only some small number to have a "room" concept that maps to the idea of the specific schema imposed by the VideoRoom plugin. It is thereby strange that all of that general video functionality is commingled together in a 8k line C file with all of the high-level room abstraction... the answer with Janus is always "write your own plugin", but either you are doing something so trivial that Janus doesn't seem to be doing anything but the lowest level WebRTC layer, or, as far as I can tell, you have to fork the VideoRoom plugin and then hope you can merge changes from upstream back into your plugin.
Am I wrong here? Like, I would love to find out I am wrong here ;P. (Which is why I was asking the OP about if their "custom solution" was a fork of VideoRoom: to see if they told me something I don't know.) But when I skim through that C file (or even the Lua file! though that demo very notably seems "incomplete" vs. the "real" C copy) I see tons of code referencing all of the codec-specific negotiation and stuff that I would explicitly be using Janus to get, so I can't not use or fork the VideoRoom plugin without losing the purpose of the platform as I would be reimplementing all of the hard parts myself (again, unless you are doing something so trivial--broadly speaking, something that doesn't involve video--that you frankly should be using libwebrtc or one of its various alternatives, such as Pion).
Not sure what you were expecting: at the very foundation of WebRTC is SDP, which implies negotiation, and with endpoints supporting potentially different codecs, negotiation is very much important whether you like it or not. That's why the VideoRoom plugin does need to take that into account. I won't get into the discussion of how complex a fork is to maintain: I always hope people contribute back what they add (assuming it's generic enough to fit the project and not customer-specific), rather than keeping it to themselves.
That said, the vast majority of people don't really need to write their own plugin, or even customizing existing ones. What we foster a lot is leveraging existing plugins as much as possible, maybe combining them at an application level, and not reinvent the wheel, and it seems to work for most (it certainly does for us, for our own applications).
On the Lua demo, it is indeed a bit more limited than the C counterpart (we clearly didn't invest as much time on it), but I'd disagree on the "incomplete" part. All the relevant parts are there, and most importantly, it's supposed to be much easier to extend and modify than the C version. There's at least one big company we're aware of that's using it in production and is very happy with it.
If you want low-level video routing functionality without the added layer of room-like logic, i.e. something that handles WebRTC and then tells you "now, here is your incoming video flow, do whatever you want with it", you might want to check out Kurento.
However our WebRTC stack has the minimum of congestion control features (plain REMB, no simulcast), and it doesn't implement SVC or newer toys like insertable streams, nor does it completely abstract you from the grunt work that WebRTC leaves up to the user (like signaling, setting up a TURN server, or having a minimum of understanding about ICE in order to troubleshoot when problems arise).
Like I said, Janus doesn't do anything by default except for some core RTC stuff and push packets around. You have to implement your own use-case via a plugin. It sounds like the Video Room plugin that's implemented in C is not exactly what you want. You can write your own plugin (maybe based on it) yourself.
If you want to implement E2E via insertable streams then you can start right here:
We write our own services to coordinate janus rooms over multiple instances. That service takes care of translating from our own domain specific entities to janus room. It's also responsible for sharding rooms to healthy and available Janus instances. Instead of treating janus rooms as thick/permanent, we prefer to treat them like "just-in-time", rooms get destroyed when not used and get provisioned when requesetd, making it very flexible.
We have made some modifications to the Janus video room plugin but only to fix bugs or provide specific functionality we needed for certain gaming use-cases. Out of the box, the plugin is already very featureful. The key is to always stay up to date with master, the project is fairly active and so keeping up to date is very important.
We have debated writing our own plugin and that's certainly a possibility in the future. (There is a Duktape JS layer available) To be completely transparent, if we were to reach that need, I would re-evaluate the solution with more recent alternatives, like pion.ly.
I recommend NOT implementing plugins in JavaScript with the Ducktape JS layer. I strongly recommend writing plugins in Lua for maximal compatibility and performance.
Why's that? We wrote it to be functionally identical to the Lua plugin (the code base is the same), so engine and language apart they should behave pretty much the same way. Is there any known issue or limitation you're aware of?
Answering for myself, not for bbeausej. We are currently using coturn. It is quite easy to setup, you have to dig a little into the configuration parameter and that's it - I would recommend it.
Oh well, you probably should use the rest api for generating credentials on the fly and think about scaling (depending on your needs).