Maintenance
How scheduled and emergency maintenance work on Meshive, what happens to billing, earnings and uptime during a window, and what the platform handles for you (package updates, driver updates, IP changes).
Scheduled maintenance — zero uptime impact
Section titled “Scheduled maintenance — zero uptime impact”If you need to power down, swap parts, or do any planned work, request a maintenance window from the Maintenance tab on the machine’s detail page at least 48 hours in advance.
Within that window:
- Meshive stops scheduling new pods onto your machine.
- Running pods keep serving — Meshive does not stop them for you. Clients whose pods are affected are notified of your window (with your title and description) so they can wind down or move their work in time. If you must power down while a pod is still running, coordinate with the client first.
- Your uptime is not penalized for any downtime during the window.
- Pods and volumes on your machine are not billed to clients while the window is open, so you earn nothing from them. Details: Billing and earnings during a window.
When you finish, end the window from the same tab (or let it expire). Either way, the machine stays unlisted afterwards: it won’t take new client workloads, and clients can’t start their stopped pods on it, until you switch it back to Listed on the machine card. Re-list once the hardware work is done and the card shows the machine as healthy. Pods and volumes still on the machine are billed again — and earn for you again — the moment the window ends, even while the machine is unlisted; re-listing only decides whether it takes new workloads. Need more time? Extend the window before it runs out — once it ends, downtime counts against your uptime again under the normal rules, even while the machine is still unlisted.
You can also fill in a title and description for the window. These are surfaced to every client whose pod is affected, so a one-line “swapping the PSU, back by 3pm KST” goes a long way toward keeping the relationship friendly. If a client still wants to discuss timing, use the message feature on the machine page to negotiate the window before you confirm it.
Emergency maintenance — reduced uptime impact
Section titled “Emergency maintenance — reduced uptime impact”Hardware does not always wait 48 hours. When you need to intervene immediately, toggle Emergency Maintenance on the dashboard before you start work.
In emergency mode, downtime counts against your uptime only for the hours when a client workload was actually running on the machine, and even then at 1/10 of the normal rate — i.e., the uptime hit is reduced to 10% of an unannounced outage. If no client pod was running, an emergency window has no effect on your uptime rate. It is still better than nothing, so always toggle it on if you cannot wait.
An emergency window ends the same way: the machine stays unlisted until you switch it back to Listed.
Emergency windows are billing-exempt in the same way as scheduled ones: no client billing and no earnings from resources on the machine while the window is open (details below). The uptime rule above is unchanged.
Billing and earnings during a window
Section titled “Billing and earnings during a window”While a maintenance window is open, clients aren’t billed for what is on your machine, and you earn nothing from it. This applies to scheduled and emergency windows alike. A window is open from its start until it actually ends — either it reaches its planned end time, or you end it early.
| On the machine | While the window is open |
|---|---|
| Pods, including dedicated serverless replicas | Not billed |
| Volumes on the machine. On a storage node, the network volumes stored on it | Not billed |
| Pods on other machines that mount a network volume on your storage node | Still billed — only the volume itself is exempt |
| Serverless Tasks | Billed as usual |
| Assets storage | Billed as usual |
| External (vast.ai) serverless replicas | Billed as usual |
What else to know:
- No commission is taken, because nothing is billed.
- Pods are not stopped. Meshive does not stop or pause Pods during a window. They keep running, unbilled, unless your work interrupts them.
- Billing and earnings resume the moment the window ends, even if the machine stays unlisted. Done early? End the window from the Maintenance tab to start earning again right away — Pods and volumes are billed again from that moment.
- Extending a window keeps resources unbilled until the new end time.
- Uptime is unchanged. Downtime inside a scheduled window doesn’t count against your uptime. In an emergency window, downtime counts only while a client workload is running, at 1/10 of the normal rate (how downtime is defined).
Package updates — handled by Meshive
Section titled “Package updates — handled by Meshive”OS-level package updates are managed by the platform. The agent applies them automatically when:
- The machine is idle (no active pods), or
- The number of pending updates crosses an internal threshold, at which point we coordinate a safe time.
You do not need to run apt update yourself. Doing so manually is unnecessary and can conflict with the agent.
Driver updates — operator-pinned, pod-aware
Section titled “Driver updates — operator-pinned, pod-aware”NVIDIA driver versions are pinned cluster-wide by Meshive’s operations team after we verify a version is stable on production hardware.
When a new driver rollout reaches your machine, the agent will:
- Wait for any active GPU pods to finish (or be re-scheduled).
- Apply the new driver.
- Bring the machine back into the pool.
This means you never have a driver mismatch with the rest of the network, and active client workloads are never killed by a driver upgrade.
IP changes — automatically reconnected
Section titled “IP changes — automatically reconnected”If your home or ISP IP address changes, the agent automatically detects it and reconnects via Tailscale. There is no need to log in and update anything. Dynamic-IP hosts are first-class citizens on Meshive.