classic: per-link weights (Moblin connection priorities) via the IPs file

[ziggy6792]

Why

On a bonded setup the links rarely cost the same: free venue Wi-Fi or Ethernet next to a metered phone tether. Classic mode balances purely on capacity (window / (in_flight + 1)), so an expensive link carries traffic even when the cheap one has room. This PR lets the user say "prefer this link": the preferred link carries what it can while it is healthy, and the others take over as it congests or dies, exactly as they do today.

It is a copy of Moblin's "connection priorities", the only open per-packet bonding scheduler with priorities we could find (BELABOX srtla and belaUI have none), so it behaves the same as what users already know from Moblin.

Semantics (as Moblin)

• A weight per link, integer 1..10 (Moblin's slider range). Missing = 1.
• Normalised on every load so the lowest link is 1: w - lowest + 1 (it subtracts, it does not divide), as in SrtlaClient.swift updateConnectionPriorities.
• In classic selection the link's get_score() is multiplied by the weight while its window is above 20 000; between 20 000 and 10 000 the multiplier fades linearly from the weight to 1; at or below 10 000 it is 1. This is RemoteConnection.swift score(). The thresholds carry over 1:1 because WINDOW_DEF/WINDOW_MAX/WINDOW_MULT are the same constants.
• So a congested preferred link loses its preference before it drops packets, and a dead one is skipped as before (timed-out and stall-gated links are still excluded).

How to use it

The weight goes in the IPs file, one link per line:

192.168.0.15 10
192.168.0.2

That is the file the sender already reloads on SIGHUP, so weights are hot-reloadable with no new flag or RPC. Rules:

• values above 10 are clamped to 10;
• a 0 or an unparsable weight is logged as a warning and read as 1. The IP still counts, so the reload guard behaves as before.

Scope

• Classic mode only. get_score() itself is unchanged, and enhanced mode ignores weights (a test covers this).
• Weight 1 is exactly neutral, so an IPs file without weights schedules as it does today. Bare-IP lines are fully backward compatible.

Observability

• New gauge srtla_send_link_weight{ip=…}.
• The weight is in the stats topic, in the added uplink … (weight N) line, and in the reload / SIGHUP log lines.

Tests

14 new tests in src/tests/link_weight_tests.rs:

• the three multiplier bands;
• weight 1 is neutral in every band;
• normalisation;
• unweighted links still share evenly;
• classic prefers the weighted link while it is healthy, the preference fades as its window shrinks, and a small window hands traffic to the other link;
• timed-out and stall-gated weighted links are still skipped;
• enhanced mode ignores weights;
• reload re-weights existing links;
• the metrics export.

cargo +nightly test --workspace passes (549 tests), and cargo +nightly fmt --check and clippy --all-targets -D warnings are clean on this branch (rebased onto c86a3e1; the only conflict was the src/tests/mod.rs module list).

Bench evidence

Two-uplink LAN bench: Ethernet en7 + Wi-Fi en0, a 1080p30 file source through an adaptive encoder, v4.1.0 + this patch, classic mode, receiver irlserver/srtla 7fa6985.

• Link A capped to ~1.7 Mbit/s for 60 s, B uncapped (preferred A = weight 10, other link 1). B's share of the bytes:

  	before the cap	during the cap	last 30 s
  weights 1/1	4.2 %	96.4 %	5.6 %
  weight 10 on A	4.6 %	93.4 %	0.3 %

  Traffic spilled to B within 0.9 s of the cap and came back to A 4.7 s after it ended (5.5 s with weights 1/1). No stream damage in either mode.

• Preferred link unplugged 2 / 8 / 12 / 8 s, then the other link for 8 s: hitless with and without weights. There was no SRT session END and no corrupt packets, and the survivor carried 5.1–7.2 Mbit/s.

• Both links capped below the stream: it still aggregates with weights on, splitting 49.5 % / 50.5 % for 600 s with no damage.

On this Ethernet+Wi-Fi bench, classic mode already sends most traffic to the wired link. With a real metered tether, the idle trickle on the expensive link is where the saving shows up.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added optional link weights (1–10) to IP list entries. In classic mode, weights influence link selection based on the link’s available window; enhanced mode ignores them.
  • Added per-link weight reporting in metrics. When weights are omitted, they default to 1; configured weights are normalized so the lowest is 1.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Repository guideline files applied to this review (2)

RELEASING.md — configured
AGENTS.md — configured

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: irlserver/srtla_send/.coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: 09342244-bb92-421d-beeb-b95aba6ba097

📥 Commits

Reviewing files that changed from the base of the PR and between c86a3e1 and ca2d033.

📒 Files selected for processing (12)

• README.md
• crates/srtla-core/src/connection/mod.rs
• crates/srtla-core/src/selection/classic.rs
• crates/srtla-core/src/test_helpers.rs
• src/metrics.rs
• src/sender/connections.rs
• src/sender/mod.rs
• src/sender/reload.rs
• src/stats.rs
• src/tests/link_weight_tests.rs
• src/tests/mod.rs
• src/tests/sender_tests.rs

Included review availability: This review used your included allowance. Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

BIND_IPS_FILE entries can include optional weights. The sender normalizes and assigns these weights to links. Classic selection applies a window-dependent weight factor, while enhanced selection ignores weights. Per-link weights are also exposed in metrics.

Changes

Weighted link selection

Layer / File(s)	Summary
Weight model and IP parsing crates/srtla-core/src/connection/mod.rs, src/sender/reload.rs, README.md	Connection weights are bounded and normalized. IP entries accept optional weights, with missing or invalid weights defaulting to 1 and values above 10 clamped. Reload parsing returns weights aligned with valid IPs.
Weight assignment and reload wiring src/sender/connections.rs, src/sender/mod.rs, src/tests/sender_tests.rs	Startup, queued updates, and SIGHUP reloads pass weights alongside IPs. Existing and new connections receive the weight for their IP; missing weights default to 1.
Classic selection and weight reporting crates/srtla-core/src/selection/classic.rs, crates/srtla-core/src/test_helpers.rs, src/stats.rs, src/metrics.rs, src/tests/*	Classic selection uses weighted scores. Stats and metrics expose each link’s weight. Tests cover normalization, selection, reload assignment, enhanced-mode behavior, and metrics.

Priority: ⬇️ Low

Estimated code review effort: 3 (Moderate) | ~25 minutes

Change: Feature

Sequence Diagram(s)

sequenceDiagram
  participant IpFile as BIND_IPS_FILE
  participant Sender as Sender
  participant Connections as Connection manager
  participant Selector as Classic selection
  IpFile->>Sender: provide IP entries and optional weights
  Sender->>Connections: pass parsed IPs and normalized weights
  Connections->>Connections: assign weights to existing and new links
  Selector->>Connections: read link scores and weights

Loading

Suggested reviewers: datagutt

Merge Risk: ⚪ Minimal · up to ca2d0

No actionable merge-blocking issue is established for the weighted-link behavior. The change is mergeable subject to normal build and test checks.

Security Architecture Review

Security architecture risk: 🔵 Low · up to ca2d0

The preferences are bounded and do not override link eligibility or grant additional authority. No introduced security weakness was established. Live reload retains existing partial-failure behavior, while configuration permissions and recovery after external process interruption remain unverified.

Retained concerns
No architecture-level concerns identified.

Security review details

Security Blast Radius

• inferred — The demonstrated effect is redistribution of a sender’s traffic among its configured, eligible uplinks to the configured receiver. Weight assignment does not itself choose a new receiver, create an additional authority, or admit an otherwise ineligible link.

Trust Boundaries and Controls

• observed — The inspected weight source is the local IP configuration file, not peer feedback. Parsing supplies bounded preferences, and core scoring independently clamps them. Configuration-file access permissions were not supplied, so this establishes the control flow rather than proving who can modify the file in production.

Resilience and Maintainability Implications

• observed — The event loop consumes a pending update once and awaits its application serially. Housekeeping retries existing connection objects; it does not reconstruct configured IPs whose connection creation failed. A subsequent accepted reload can attempt those absent IPs again. No automatic restoration of the consumed configuration snapshot was found in the inspected lifecycle.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: per-link weights for classic-mode scheduling configured through the IPs file.
Docstring Coverage	✅ Passed	Docstring coverage is 90.20% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 51 functions across 11 files. (1 skipped: 1…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Autopilot could not be updated. Open Coding to check access and billing.