Adaptive weight controller for mwan3-nft (OpenWrt 25.12, nftables). It adjusts load-balancing weights at runtime from link quality. It does not patch mwan3 and does not run any active probes. Is ucode and shell implementation of https://github.com/AlexStarc/mwan3-autobalancer

How it works

Reads the latency and packet loss that mwan3track already collects (via ubus call mwan3 status).
Smooths the samples (EWMA) and computes a quality factor per link, relative to that link's own baseline latency. A single lost ping barely changes anything.
Scales each member's configured uci weight by its quality factor and rounds the result to a small total (default 20, i.e. 5% steps).
If the result differs enough from the live chain (hysteresis), rewrites the numgen ... vmap rule of mwan3_policy_<name> in one atomic nft transaction (flush + add in a single batch). All other rules, including last-resort, are re-added unchanged. Both numgen inc and numgen random are supported.

Safety

Observe mode is the default: decisions are logged, nothing is changed.
Level-triggered: if mwan3 rebuilds the chain (ifup/ifdown), the next pass or the hotplug poke restores the adjusted weights.
Unrecognized chain shape, unknown member or missing data: the chain is left untouched.
On stop (SIGTERM) the original uci weights are restored. procd respawns the daemon after a crash.

Requirements

mwan3-nft, ucode with the fs, uci, ubus and uloop modules (ucode from OpenWrt 25.12 or newer).
check_quality enabled on the balanced interfaces. To keep mwan3's own up/down decisions unchanged, neutralize its thresholds:

for i in wan1 wan2; do          # your interface names
	uci set mwan3.$i.check_quality='1'
	uci set mwan3.$i.failure_latency='100000'
	uci set mwan3.$i.recovery_latency='99999'
	uci set mwan3.$i.failure_loss='100'
	uci set mwan3.$i.recovery_loss='99'
done
uci commit mwan3 && mwan3 restart

Usage

/usr/sbin/mwan3-autobalancer once       # one-shot calculation printout, changes nothing
/usr/sbin/mwan3-autobalancer restore    # restore original uci weights
/etc/init.d/mwan3-autobalancer start    # run the daemon (needs enabled=1)

Recommended rollout: enabled=1, keep observe=1, watch logread | grep -i autobalancer, then set observe=0.

Configuration (`/etc/config/mwan3_autobalancer`)

Option	Default	Meaning
`enabled`	`0`	Start the daemon
`observe`	`1`	Log decisions only, do not touch the chain
`policy`	`balanced`	mwan3 policy name
`interval`	`30`	Seconds between passes (minimum 5)
`total`	`20`	Sum of integer weights in the chain (step = 100/total %)
`hyst`	`1.0`	Apply when a target is off by at least this many steps

With numgen random finer steps are fine, e.g. total=100, hyst=5. With numgen inc keep total small, because connections are handed out in runs of up to total.

Model constants (baseline drift, loss threshold, smoothing) are at the top of model.uc.

ubus

Object mwan3_autobalancer:

poke: trigger an early pass (called from the hotplug script after mwan3 rebuilds the chain)
status: current mode, policy and last decision

ToDo

LuCI page.

Statistics collector (optional)

ab-stats.sh appends one CSV row per run: chain weights, per-link status/latency/loss, interface byte counters, connection counts per mark, daemon RSS and CPU ticks.

cp ab-stats.sh /root/ && chmod +x /root/ab-stats.sh
(crontab -l 2>/dev/null; echo '* * * * * /root/ab-stats.sh >/dev/null 2>&1') | crontab -
/etc/init.d/cron enable && /etc/init.d/cron restart

Data goes to /tmp/ab-stats.csv (RAM, trimmed to the last 8000 rows, lost on reboot).

Thanks

To idea thakns https://github.com/AlexStarc

README.md

mwan3-autobalancer