Hello all,
Last month, I sent out a survey on Discord asking for orchestrator feedback to help gauge sentiment around some design points for Livepeer’s next-generation capacity management platform, which we are calling the Control Center.
The results of the survey were very insightful and I wanted to share them with the community, as well as offer some additional background for some of the questions.
Overall, we had 9 respondents, many of whom also manage multiple orchestrators. With just 100 orchestrators in the current set, this was a good sample, representing a variety of opinions, although there were some clear areas of agreement.
The first question was more of a gut check: How interested are you having a service manage your Livepeer infrastructure for you? The range was:
1: “Not interested, I need full control”
5: “Very interested - I’ll supply hardware if someone else does the rest”
There’s certainly room for interpretation among those ranges, and most said they would prefer to retain more control, rather than less. This is somewhat understandable - orchestrator control has been the “Livepeer way” since the beginning, but let’s drill down into exactly what they might have meant by full control.
This next question was about the environment their hardware was running under: do orchestrators own their hardware? Are they leasing machines, whether bare metal of VM?
The motivation for this question was to better understand what kinds of GPU sandboxing that orchestrators might be able to run, and in what packaging. For example, Kata (Confidential Containers) requires a datacenter-grade bare metal machine. Should we ship a whole VM image, or stay with containers? Etc.
As a result of this spread, we are using gVisor for the sandboxing component since it offers the broadest compatibility at the moment with acceptable tradeoffs.
Are you comfortable with a service that would manage the Livepeer capabilities on your machines? Capabilities would update, stop or be swapped automatically based on network demand.
Yes: 100%
No: 0%Can this service check your GPU, CPU, memory, storage, and network performance for app compatibility?
Yes: 100%
No: 0%Are you comfortable publishing your hardware specifications, uptime, benchmark results, and application-probe performance?
Yes: 100%
No: 0%
Even though a significant number of orchestrators prefer to have “full control” they are unanimous in wanting a service to manage the Livepeer capabilities on the machine, including compatibility and performance checks, and publishing those results.
This is an encouraging result, as it recognizes that capacity management is one of the weak spots of the network right now, and orchestrators are still committed to the network’s history of performance transparency.
Can this platform manage public access for Livepeer jobs through a tunnel? You would not have to manage DNS, TLS certificates, proxies, open ports and ingress configuration.
Yes: 66.7%
No: 33.3%
On tunneling: the idea is that the service would be able to shield orchestrators from direct traffic, reduce the amount of configuration required on their part, and minimize their exposure. A majority were receptive to this idea, but a few weren’t. That may have depended on a few perceived use cases, eg wanting a direct connection for CDN-like use cases.
Can the platform collect host logs, application logs, metrics, and other per-job diagnostic information?
Yes: 88.9%
No: 11.1%
Logging and metrics, however, was an easier sell. We’ve collected logs in the past for Stream Diffusion workloads, which has been invaluable in helping diagnose issues and ensure network stability. Some orchestrators have also historically made their metrics dashboards available for scraping. Orchestrators have signaled a willingness to continue doing this.
Are you comfortable having a service redeem payments on your behalf? You would not need to manage a redemption wallet.
Yes: 77.8%
No: 22.2%
Payment redemption is another bit of heavy-handed configuration that we can relieve orchestrators of: they would no longer have to fund a wallet and keep it hot for redemptions. There is also less exposure overall: without wallets on the orchestrator machine, there is less to lose.
There is certainly a tradeoff in trusting a third-party to redeem payments on your behalf, which shifts the trust model slightly (although there are ways to work around this aspect), which may explain the small amount of reluctance here.
Are you comfortable dedicating participating machines exclusively to Livepeer?
Yes: 77.8%
No: 22.2%
Knowing that a machine is Livepeer-exclusive allows us to better allocate the workloads on it and have more performance predictability, which is better for the network and its customers.
Would you participate in host monitoring to help detect sandbox breakouts or machine compromise? Information that could be collected includes process activity, host logins, network connections, or security events.
Yes: 66.7%
No: 33.3%
Being able to monitor the host may allow us to identify anomalous patterns that could compromise the security of the network. This helps us ensure that the network is running as expected, which protects both orchestrators and customers.
Are you comfortable having the platform update host packages, GPU drivers, or operating system, and reboot machines when necessary?
Yes: 22.2%
No: 77.8%
Host control is where it gets a little bit spicy. The reluctance is understandable. This question stems from a few things:
-
The GPU-driver-app compatibility matrix can be very broad and complicated. The right types of GPU support isn’t always easy for operators to set up, nor is it easy for developers to debug. Anything we can do to minimize variability for everyone involved is a good thing.
-
With AI-assisted attacks accelerating, we may not want to wait for operators to apply security patches. The operators would not need to bear the added responsibility of monitoring and responding to these attacks. This is especially more challenging when everyone is using slightly different setups.
-
Operators can sometimes fall behind on updating their boxes, which potentially means they fall out of the compatibility window for some Livepeer apps. This loses revenue for the operator, and the network loses capacity, which has downstream impact with customers. We can take the onus for this off operators.
Shipping a VM image, instead of a container, is perhaps an alternative here, but is less compatible with existing orchestrator setups.
Host access This is more of a drill-down on the precise types of host access that folks would be willing to allow.
Are there any types of applications, workloads or specific software sources you would not allow on your machine? If YES, what would you not allow?
Yes: 66.7%
No: 33.3%Response Summary
- Restricted Content and Activities: Exclude specific geographic locations, illegal content, non-consensual deepfakes, adult content, and hacking activities.
- Prohibited Workloads: Block crypto mining, botnet runner images, workloads from anonymous or unaudited sources without a review process, and anything that burns GPU cycles for other networks instead of Livepeer demand.
- Security and System Access: Prevent workloads that require privileged host access, weaken sandbox boundaries, or create legal exposure under EU law, requiring operators to have a way to automatically exclude or review such categories before deployment.
These concerns give a sense of the kinds of guardrails or knobs that might be needed. While ideally orchestrators could be hands-off with respect to network management, they need to also be comfortable with the workloads they are running for Livepeer.
Would you contribute a machine to a pilot of the observability and allocation platform?
Yes: 66.7%
No: 33.3%
All in all, a very encouraging set of results. We’ll be sharing more about the network management service (Control Center) this week.


