Skip to content

v1.9.0

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 21:11
· 5 commits to main since this release
Immutable release. Only release title and notes can be modified.
35e40e3

Sable 1.9.0 adds a DHCP server. Sable can now hand out addresses on your
network, keep reservations, and publish each device's name in a zone it
serves, so a device's name resolves the moment it gets its lease. Every node
of a cluster serves DHCP from its own slice of each range, and there's an
import for reservations from dnsmasq, Pi-hole, OpenWrt, AdGuard Home,
Technitium and UniFi. DHCP is new and off until you set up a scope, so
nothing changes until you turn it on.

Sable 1.9.0 also supports UniFi Network 11, so UniFi Host Sync keeps
publishing your devices after your gateway updates.

Upgrading from 1.8.2

  • Update every node in a cluster first. Sable won't turn DHCP on while a
    member runs an older release, because 1.8.2 can't load a configuration
    with a [dhcp] section.
  • Port 67 needs the right to bind a low port and to see broadcasts. The
    systemd service already has CAP_NET_BIND_SERVICE. In Docker, use host
    networking with the new compose.dhcp.yaml, which keeps only
    NET_BIND_SERVICE; Docker's bridge doesn't pass broadcasts. A web update
    can't add the image's new file capability to an older container: pull the
    1.9.0 image.
  • The DNS Administrator and Operator roles now have dhcp.read and
    dhcp.write, and Auditor has dhcp.read. dhcp.read lists every lease
    with the device's hardware address. Check API tokens and users with those
    roles.
  • The database gains a lease table on first start, which 1.8.2 ignores. To
    go back to 1.8.2, remove the [dhcp] section first; a backup taken with
    DHCP set up won't load on 1.8.2.
  • Cluster clocks must agree within 30 seconds. A replica on 1.8.2 with a
    wrong clock shows as offline, as before, because it can't answer the
    primary's clock check; the primary's log still names the clock. Update
    every node first.
  • Nothing else is needed. Turning DHCP on is your choice; until then Sable
    answers DNS as before.

DHCP

  • Scopes serve DHCPv4 (RFC 2131) on any interface with an address in the
    subnet. DHCP → Set Up fills a scope in from the network: the router
    Sable found, a pool that leaves out Sable's own address and the router's,
    and Sable as the DNS server. A scope can have several ranges, a lease
    time, DNS, NTP and search domains, static routes (option 121, RFC 3442),
    PXE boot settings, IPv6-only preferred (option 108, RFC 8925) and custom
    options.
  • Before an offer, Sable pings the address to check nothing already uses it
    (RFC 2131 section 2.2). Before Turn On, it listens up to 8 seconds for
    another DHCP server on each of the scope's networks and names the one that
    answers.
  • Each lease's name is published as an A record and, with Reverse on, a
    PTR record. The label comes from the reservation, else the device's FQDN
    option (option 81, RFC 4702) or host name (option 12). A record you added
    by hand wins over a reservation, a reservation over a UniFi name, and a
    UniFi name over a dynamic lease. Sable never renames a device with a -2
    suffix, and a lease never takes wpad or isatap.
  • Relayed requests are answered when the relay's address is inside a scope,
    and the relay agent information (option 82, RFC 3046) is echoed back.
    A renewal a device sends straight to Sable takes the scope that holds its
    address (RFC 2131 section 4.3.2), so it works behind a relay that passes
    on broadcasts only, and on a node whose own address is a /31 or /32.
  • The DHCP page, in the sidebar's System group, lists leases,
    reservations and scopes. Reserve turns a lease into a reservation, and
    reservation edits apply without a restart.
  • The scope dialog keeps rarely changed fields (DNS servers, lease time,
    the scope ID) under Advanced Options. Typing a subnet fills in the
    range and router without overwriting what you typed, and a new scope with
    no ID takes one from its name. Errors name the field in the form's words.
  • Import brings reservations from dnsmasq, Pi-hole 5 and 6, OpenWrt,
    AdGuard Home, Technitium and UniFi. Preview lists what it would add and
    skip, with the reason, and saves nothing. The same import is
    POST /api/v1/dhcp/import, with dry_run=true to preview.
  • The JSON API covers scopes, reservations, leases and takeover. New MCP
    tools get_dhcp_status, list_dhcp_leases, set_dhcp_reservation and
    remove_dhcp_reservation are off until you add them to the MCP tools,
    and a token needs the new dhcp.read or dhcp.write permission as well.
  • The DHCP guide walks through moving from a router or another server,
    with a test device first and a way back.

UniFi

  • UniFi Host Sync works with UniFi Network 11. Network 11 gave every network
    a new ID and kept the old one as legacy_id, so after a gateway updated,
    existing mappings matched nothing and the sync published no hosts. Mappings
    now match either ID, with nothing to redo. A mapping whose network the
    controller no longer reports shows Missing on the UniFi page.

Clustering

  • Turn DHCP off before you create or join a cluster.
  • Each node leases only from its own slice of every range, so two nodes that
    lose touch never hand out the same address. Every node answers every
    reservation. The scope dialog splits a range evenly across nodes and warns
    when a slice is smaller than the scope's active leases.
  • Leases travel on the cluster heartbeat, so every node holds all of them
    and the primary publishes every name.
  • Take Over DHCP Ranges on an offline node's Cluster page moves its
    ranges to the primary, which renews its devices on the addresses they
    hold. Promote This Replica takes over the old primary's ranges in the
    same step. A node that comes back after a takeover stops serving those
    ranges.
  • A node whose database is wiped or restored from a backup takes its leases
    back from the other nodes, including leases made or released since the
    backup.
  • A node whose clock is more than 30 seconds off keeps serving its own
    ranges. Both Cluster pages show it as Clock out of step with the
    offset, both logs name the clock, and its ranges can't be taken over until
    its clock is set. After 5 minutes out of step, alerts send Node clock out
    of step
    , and Node clock back in step once it syncs again.

Insights

  • DHCP leases name devices on the Devices tab, after your own names and
    UniFi's. The device drawer shows a device's lease, with Reserve
    Address
    .

Console

  • With security off, the console now refuses cross-site requests that change
    something. Same-site and API requests still go through.

Diagnostics

  • GET /api/v1/debug/pprof/{profile} returns this node's heap, allocs,
    goroutine or threadcreate profile for go tool pprof, or as text with
    debug=1 or debug=2. gc=1 collects garbage before a heap profile. It
    needs metrics.read, like /metrics.

Known gaps

  • DHCPv4 only. DHCPv6 is not included.
  • The scope for a relayed request comes from the relay's address (giaddr)
    alone. Link selection (RFC 3527) and subnet selection (RFC 3011) aren't
    built.
  • Relays tried: ISC dhcrelay 4.4.3 with option 82, and dnsmasq 2.91's relay.
    No hardware router's relay has been tried.
  • While a node's clock is out of step, the primary doesn't publish names for
    the leases it grants. Once the clock is set, a lease it granted while
    behind shows as ended until the device renews.
  • Not yet tried: Operator, Auditor and DNS Administrator sessions in the
    console, Safari and Firefox, two replicas restored at once, clock skew on
    a PostgreSQL cluster, and promoting a replica whose clock is out of step.

Fixed

  • Stopping Sable no longer waits up to five seconds on idle connections.
  • SQLite writes from one process queue for the write lock instead of failing
    with SQLITE_BUSY under load.
  • A node that started without working upstream DNS retried its DNSSEC trust
    anchor refresh at once, logging thousands of failures a second and holding
    shutdown for the whole shutdown_timeout. It now waits between failures,
    from 5 seconds doubling to 5 minutes, and stops as soon as Sable does.
  • A request still running at shutdown, such as a backup download to a slow
    client, gets at most half of the remaining shutdown budget before it is
    closed and logged, so the DNS listeners stop on time and the DNS cache is
    saved.

Dependencies

  • Added github.com/insomniacslk/dhcp for DHCPv4 messages.
  • Updated templ to 0.3.1118, pgx to 5.13.0, modernc.org/sqlite to 1.62.0,
    and golang.org/x/crypto, mod, sys and term.