Telemetry getting stuck when there is many same subscriptions

[atruebin]

I need to get telemetry from multiple drones. So I start mavsdk_server for each, connect to them and trying to get telemetry. For some minor number of drones it works, but for more (approximately 10) it gets stuck (none of generators yields). So as for many same tasks for one drone.

The simpliest way to reproduce is to edit telemetry.py example like this:

for i in range(n):
  asyncio.ensure_future(print_battery(drone))

It makes no sense to make multiple identical tasks for one drone of cource, but it behaviours the same as one for each drone.

If n=8 it works, if n=12 gets stuck. But if you run one script with n=8, and one more with n=8 (for other mavsdk_server of cource), they both work. And if you use each subscription from Telemetry plugin once (there more than 12) it works as well.

So I can guess the problem is related to grpc, but don't know how to solve it. Any ideas?

[JonasVautherin]

Just one thing: in MAVSDK, you can subscribe to a callback only once in C++ (that's something that should be improved, but that's how it is now). So if you subscribe twice to the position against the same mavsdk_server, the first one will be unsubscribed when the first one is registered.

This is not a limitation in MAVSDK-Java and MAVSDK-Swift, because on the client side we handle it. But for Python we don't. I started looking into it a while ago, you can find my work here: #60.

I feel like this is not your root issue, but this is the reason why your reproducible example fails. Is that correct?

[atruebin]

Well, maybe this is the reason why experiment with one mavsdk_server fails.

But yes, my goal is multiple drones, so there should be many mavsdk_servers and each one has only one subscription of a type. In this case unsubscribing not happening at all, right?

When you start many mavsdk_servers and create System instance for each, the only thing that intersects is some global variables and global limitations, as far as I understand. This is why I think of gRPC, cause there is one server for the whole script, yes?

[JonasVautherin]

Well, maybe this is the reason why experiment with one mavsdk_server fails.

Yep, probably

so there should be many mavsdk_servers

Yes, but you are starting them manually, each listening on a different gRPC port, right?

and each one has only one subscription of a type. In this case unsubscribing not happening at all, right?

Correct 👍

the only thing that intersects is some global variables and global limitations, as far as I understand.

They should not intersect at all. You have multiple System objects that are each connected to a different gRPC server. Just like there is nothing shared between your Firefox connected to DuckDuckGo and your Safari connected to ProtonMail: it's twice HTTPS, but those are two very separate connections, clients and servers.

How do you start the mavsdk_server instances?

[atruebin]

With subprocess.Popen([mavsdk_server_path, '-p', f'500{d_id}', f'udp://0.0.0.0:150{d_id}'], stdout=subprocess.PIPE).

[JonasVautherin]

Hmm, so they should not share anything with each other if {d_id} is indeed different everytime 🤔. And then are you sure that each of your System() connects to the right instance? How do you do that part?

[atruebin]

Right after starting mavsdk_servers:

d = mavsdk.System(mavsdk_server_address='localhost', port=int(f'500{d_id}'))
await d.connect(system_address=f'udp://0.0.0.0:150{d_id}')
drones.append(d)

[JonasVautherin]

Can you try without setting the argument in connect()?

d = mavsdk.System(mavsdk_server_address='localhost', port=int(f'500{d_id}'))
await d.connect()
drones.append(d)

I don't think it makes a difference, but it's useless anyway if you set mavsdk_server_address above.

[atruebin]

Tried, the result is same.

[JonasVautherin]

I don't get how the MAVSDK instances would interfere with each other 🤔. How do you start the simulated drones? All in the same gazebo instance? Could it be that MAVLink stops being sent from the other side, rather than it not being received on the MAVSDK side?

[atruebin]

This is actualy real drones connected to router. And I just tried turn on 10 ones, and it started working!

Experementaly it turned out that the problem occures when there are 10 or more inactive subscriptions (not connections!). When 10 drones turned on, script works with any number subscriptions for working drones and 9 subscriptions for not working, and with 10 subscriptions for not working - gets stuck.

This definetly not problem on the other side, cause starting 2 scripts simultaneously for more then 10 inactive subscriptions in total - works.

[JonasVautherin]

Sorry, what do you mean by "inactive subscriptions"?

[atruebin]

I mean subscription to stream of the mavsdk_server that connected to empty UDP port (there is no data).

Imagine I have working script, telemetry comes from all drones. I turn off one drone, subscriptions for that server become 'inactive' (so there become more than 10 'inactive' subscriptions in total) and all telemetry of all drones in script gets stuck. I turn it on again, it continue working.

[JonasVautherin]

What if you start mavsdk_server from a shell instead of using subprocess.Popen? Does subprocess.Popen somehow link them together? Like could it be that one being stuck would block all the others there? Not sure how subprocess.Popen works 🤔

[atruebin]

Quickly tried one shell mavsdk_server and others from script. Shell one gets stuck as well.