Media Streaming Mesh is a new concept for supporting RTP-based real-time media applications in Kubernetes.
The goal of Kubernetes is to provide a "platform for automating deployment, scaling, and operations of application containers across clusters of hosts". Most applications deployed in Kubernetes today are web-based, and so much of the effort around networking in kubernetes is optimised for web applications. One example of this is the service mesh architecture (exemplified by Istio), where applications communicate with each other via web proxies rather than directly over IP.
Media Streaming Mesh will enable developers of real time media applications based on RTP to focus on their business logic whilst the Media Streaming Mesh infrastructure facilitates real-time connectivity for microservices.
Today's service meshes generally only support TCP-based applications (and in fact are optimised for HTTP-based web applications). Any support for UDP that is added to service meshes is likely to be focussed on enabling QUIC (since HTTP/3 runs over QUIC).
Real-time applications generally run over UDP rather than TCP. Media Streaming applications typically rely on RTP (the Real-time Transport Protocol) - which runs on top of UDP, and hence RTP will be the initial focus of Media Streaming Mesh.
RTP enables measurement of loss and jitter as it carries sequence numbers and timestamps in the packet header and we will monitor these in Media Streaming Mesh.
One challenge with RTP is that it often runs on ephemeral UDP ports which are assigned by a TCP-based control channel such as SIP or RTSP. However proxying these TCP-based protocols will enable us to implement URL/URI-based routing and to map the UDP ports dynamically.
Service meshes bring many benefits to web applications such as:
- "Layer 7" URL based routing
- flexible load-balancing
- support for canary deployments
- failure detection
- statistics/metrics export
- pod-to-pod authentication
- data encryption
Our goal is to extend all these to real-time media streaming applications, whilst also enabling additional capabilities such as:
- data protection using Forward Error Correction, sending data over multiple paths, and/or NAK-based mechanisms
- stream fan-out to multiple clients by replication and/or unicast to multicast conversion.
Many cloud-native applications involve a mixture of 'east-west' traffic between microservices (generally within the same cluster) and 'north-south' traffic between the application and external entities.
This will be equally true for real-time media applications.
For example in real time collaboration there might be traffic between clients and the cloud-based infrastructure, but also traffic within the cloud - e.g. for speech to text transcription or for meeting recording.
Equally for broadcast media applications there might be multiple camera feeds into a news-room where one feed is selected, various data (e.g. breaking news) is overlaid, and then the resulting stream is sent out for broadcast.
Our original demo implementation relied on a simple Go-based proxy that ran as a pod sidecar (plus a micro-CNI that creates IPtables rules to direct traffic into the sidecar, and a mutating webhook that injects the sidecar proxy into labelled pods).
We have now started implementing our longer-term architecture, and will iteratively enhance this implementation.
The implementation consists of:
- An RTSP control plane that runs as a per-cluster kubernetes service - written in Golang, and leveraging the https://github.com/aler9/gortsplib library.
- An RTP/RTCP data plane that runs as a per-node daemon-set - our prototype implementation is written in Golang, but our next step will be to re-implement it in asynchronous Rust using Tokio/Tonic/etc.
- An stub sidecar which sends control plane traffic to the per-cluster control plane over gRPC, and which is also able to convert interleaved RTP data over RTSP to native RTP over UDP. Written in asynchronous Rust using Tokio/Tonic/etc.
- A micro-CNI and mutating webhook as noted above.
Longer term our expectation is that we'll implement SPIFFE/SPIRE for pod to pod authentication, OPA for workload authentication and RTP over QUIC for node to node encryption.
We intend to split the control plane into a "controller" that takes care of mapping streams onto the cluster, and a "control plane" entity into which we can plug RTSP and other protocols (such as SIP, RIST, WebRTC, AMWA NMOS etc.)
We also intend to make the data-plane pluggable. The most likely architecture is a configurable chain of WASM filters with a simple filter interface consisting of input and output packet vectors. This will enable easy contribution of enhanced features such as FEC, NAK-based error correction, sRTP encryption, congestion control etc.
To improve performance the data-plane proxy will be enhanced to use DPDK for network I/O, and the stub will use eBPF to intercept data plane traffic from the app at the TCP/UDP socket layer and forward it to the RTP proxy using AF_XDP at the proxy.
In order to keep footprint light one key will be to deploy only the required components for the service being implemented. So for example we only deploy a single control plane instance per cluster and a single RTP proxy per node. Remote edge nodes can be deployed without a local control plane - instead relying on a cloud-located control plane instance.
The RTP proxy is deployed with a co-located stub sidecar that acts as an ingress for control plane traffic to enable north/south flows. The same stub sidecar is also co-located with media pods to proxy east/west control plane traffic, and also (as noted above) may be used to interwork between interleaved RTP data over RTSP to native RTP over UDP. Similar interworking functions will be required for RIST.
We're looking for potential users of Media Streaming Mesh to help us define the solution, and for developers to help us create it!
Please do join our Slack channel.