tech-news · 19 November 2026
Tailscale Caches the Map Before the Network Goes Missing
Tailscale's opt-in netmap cache lets clients restart into a usable tailnet when the control plane is slow or unreachable, with clear limits around stale state, storage and mobile support.
Tailscale has introduced an opt-in client-side netmap cache that lets a machine restart into a usable tailnet before it can reach the Tailscale control plane. The 17 November announcement matters because a temporary control-plane problem, captive portal, filtered connection or simply poor Wi-Fi no longer has to turn a routine reboot into a complete loss of access to familiar peers.
This is not an offline mode and it does not remove Tailscale's coordination service from the design. It is a carefully limited fallback: save the last useful map locally, use it as a warm starting point, and keep trying to replace it with current information.
What changed this week
A Tailscale client normally receives a network map, or netmap, from the coordination server. That map contains the local node's settings and the peer information it is allowed to see, including endpoints, keys, routes, DNS details, capabilities and each peer's preferred DERP region. Access policy is already reflected in the per-node map, so two devices in the same tailnet do not necessarily receive identical views.
Until now, that working state was principally held in memory. Existing connections could survive a coordination gap, and a running client could still know enough to make some new connections, but a restart or profile switch removed that advantage. The restarted client first had to contact control and fetch enough state to proceed.
With caching enabled, the client writes relevant netmap state to persistent storage and loads it during startup or a profile switch. Tailscale says it divides the cache into chunks and rewrites only changed pieces, reducing write volume and storage wear. A fresh full map also causes old cached items to expire.
The interesting part is what happens after loading. Cached peer endpoints and the remembered home DERP region can bootstrap an encrypted WireGuard connection through DERP. A restarted client has generated a new short-lived DISCO key, however, and peers cannot initially use that key to negotiate a direct path. Tailscale's solution is to carry the new DISCO public key inside the already encrypted relayed tunnel. Once peers exchange that material, they can probe endpoints and may establish a direct connection without first involving the coordination server.
That is more than saving a configuration file. It preserves just enough trusted state to rebuild the discovery path while keeping peer traffic encrypted end to end.
The current documentation says the feature requires Tailscale client version 1.104 or later. It remains opt-in through the cache-network-maps node attribute in the tailnet policy file. The same documentation provides a status check and says mobile clients and ephemeral nodes are not supported. An independent summary also highlights the practical boundary: this is a routine resilience feature to test on unreliable links, not an urgent migration.
Why it matters
For a homelab, the awkward failure mode is easy to recognise. A remote server, subnet router or workstation restarts at precisely the moment its internet path can carry some traffic but cannot reach Tailscale control reliably. The hardware is healthy, the local network is healthy and known peers may be reachable, yet remote administration waits on coordination. A warm netmap gives the client a credible route back into service sooner.
The same pattern appears in field devices, temporary sites, remote development machines and agent workers. Automated systems often reboot without a person nearby to clear a captive portal or diagnose selective filtering. Letting a previously enrolled node reconnect to known peers can preserve access to logs, queues or management services while normal coordination recovers.
There are meaningful trade-offs. Cached state ages. It cannot describe peers added after the last refresh, and moved endpoints, changed DNS settings or new routes may be wrong. Node-key expiry eventually ends communication until the client contacts control again. Tailscale also cautions that durable connectors such as subnet routers and exit nodes may gain less from caching if they rarely restart, while incurring extra disk activity. Small systems using wear-sensitive flash deserve the same caution.
Policy behaviour also needs precise interpretation. The cache contains only peers the node was permitted to reach when that map was issued. If policy changes while one side remains connected to control, enforcement on that updated side still applies. A fully isolated group using old maps cannot learn every new administrative fact immediately, which is why caching is a continuity aid rather than a replacement control plane.
What to watch next
The first useful evidence will come from real failure tests, not ordinary connected startups. Operators should compare cold and warm boot times while selectively blocking control access, then verify which peers reconnect through DERP, which become direct, and how quickly fresh policy and endpoint changes arrive after control returns.
Tailscale has said caching may become the default after successful testing. Before that happens, watch for mobile support, clearer measurements across large tailnets, reports about write volume on embedded storage, and edge cases involving expired keys or rapidly moving peers. The feature's value is straightforward, but its quality will be measured by how predictably it fails when the saved map no longer matches reality.