The collector rewrite: MikroDash asks your routers a lot less
This release rebuilds how MikroDash gets data out of your routers. Nothing about
the pages changes, but almost everything behind them does, and the result is a
dashboard that is lighter on your hardware and quicker to fill in.
Why it mattered. MikroDash used to have every collector run its own timer and
ask the router for whatever it needed, whenever it felt like it. Two collectors
that wanted the same table asked twice. A page you were not looking at kept
asking anyway. A router nobody had open still ran the full set. RouterOS limits
how many things can talk to it at once, so all of that competed for the same
narrow pipe, and on a busy device it showed as pages that filled slowly.
What it does now. Collectors say what they need rather than going and getting
it. One reader per router works out what is due, asks once, and hands the answer
to everyone who wanted it. Where RouterOS can push updates instead of being
polled, it does. And nothing runs unless something is actually watching it.
Measured on a four router fleet, sitting idle: 270 commands a minute down to
81. The router being viewed went from 95 to 47.
Added
- Device names and IP addresses in more places. MikroDash now reads the ARP
table, which is what connects an IP address to a device. WiFi Clients shows each
client's address; Connections can name a device whose DHCP lease is filed under
a different address, and shows its MAC either way; Network Topology can locate
a neighbour that does not announce an address of its own. - Reverse DNS as a last resort for naming. A device with a fixed address and
no DHCP lease, typically a printer or a server, is looked up by name and appears
as that name rather than as a bare MAC. - The Stream / Poll switch does something now. It has been in the device
dialog since the rewrite and nothing read it. Fourteen menus can now be pushed
by the router instead of polled, and the switch chooses per device. Both modes
honour the same interval slider, so choosing Poll never silently means slower. Collector-Architecture.md, a plain description of how the collector layer
works and why. It is checked by the test suite, so it cannot quietly go out of
date.
Changed
- Collectors stop when nobody is looking. Close a page and its collector
stops; open it and it starts. This used to be a hand written list of which page
meant which collector, and it disagreed with reality five separate times, each
one a dashboard card that silently stopped updating. - Routers you are not viewing cost almost nothing. A device kept online only
for alerting runs the handful of collectors the alert rules need, not all of
them. One kept online only so the Devices page can say "up" runs none at all. - Two collectors wanting the same table now cost one read. Connections and
Bandwidth share the connection table; Interfaces and the traffic graph share one
channel where they used to open two. - Interface rates are read from a live channel rather than measured by
repeatedly asking. That alone removed 52 commands a minute per router. - Ping honours the interval you set. RouterOS silently ignores a ping interval
above five seconds; a 30 second setting delivered a ping every five. Above five
seconds MikroDash now polls instead, so the setting means what it says.
Fixed
- The Devices page showed no WAN RX/TX. Three separate causes, including one
that hid because most routers use the default interface name and matched by
accident. - Dashboard cards that stopped updating after you visited the page they share
a collector with and navigated away. - A closed browser tab left collectors running until the connection timed out.
- The DHCP and Connections pages could lose their device names on a router
where the lease table is keyed by a different address than the one in use. - The ARP poll interval setting was saved, validated and read by nothing.
Internal
- Three layer split: acquisition, derivation and views. Every collector's
rows-to-payload step is now a plain function that can be tested without a
router. - 58 static self checks, all failing in both directions: an unrecorded gap fails,
and so does a recorded gap that has since closed. - The behavioural guidelines were rewritten after they were measured causing
harm, and the risk appetite is now stated rather than implied.