perf: cut redundant Sparkplug payload work and bound topology-driven renders

[bluegrassiot]

Closes #161

What this does

Two lanes, both aimed at the same failure: throughput collapsing once Sparkplug traffic gets dense.

Core lane: receive, alias and metric work

• Alias enrichment reused the envelope the payload pipeline had already decoded instead of reparsing the protobuf payload and reading the application message payload segment a second time for every message.
• The payload pipeline now reads the raw payload only on failure paths, so a successful decode skips a second payload read.
• Metric merges index the existing snapshots instead of rescanning the list for every incoming metric, keeping previous order and last-write-wins on duplicate names.

UI lane: bounded render coalescing

• A burst of topology events drove one render per event on the nodes view and the page shell. Both now signal through a RefreshCoalescer that holds a single pending slot, so renders are bounded by the coalesce window rather than by event rate.
• The broker thread only requests a refresh. The dispatcher reconciles the active tab or the nodes selection state and renders in the dispatched refresh, so the state a render reads is not mutated from the broker thread.
• A refresh queued before disposal never runs, and a dispatch that throws releases the pending slot.

Tests

• Core: alias resolution from the decoded envelope, plugin decoder fallback when TypedPayload is null or a foreign type, and in-place metric merge ordering including concurrent node data.
• UI: coalescer unit tests plus page and view tests for burst coalescing, a blocked dispatcher backlog, dispatch failure and disposal.

Validation

• dotnet build MqttProbe.NoMaui.slnf: 0 warnings, 0 errors.
• Core tests: 1905 passed, 1 skipped (pre-existing platform skip).
• UI tests: 1801 passed, 1 skipped (pre-existing platform skip).
• Integration tests: 61 passed.
• Coverage: 89.4% line coverage, against the 80% gate.
• Full format check, comment check, blocking-awaits check, file-length check and the security scan all pass, along with the 361 gate-script tests (2 skipped).

Benchmark

Matched isolated runs on the Release Web host against a local Mosquitto with a headless 50 ms emulator and the same workload (about 528 to 2030 broker-delivered msg/s). At a cap of 50000 the whole-app process CPU was 72 to 82% lower, with no limiter drops. Node and detail interactions still succeed and both views are visible at the end of the run. A complete metrics scrape dropped from 5.7 to 11.4 seconds down to about 1.2 seconds. That is the whole scrape operation, not just a render.

Process CPU as a percentage of one core, by view and load:

Load	Nodes view	Browser view
1000	83.6 to 20.8	110.5 to 19.8
2000	132.8 to 36.5	168.7 to 34.0

Limits of this evidence

• Higher capacity and continuous detail-value liveness are not established.
• At a cap of 1000, throughput was unchanged and drops still appeared near the limit.
• The original freeze was not fully reproduced, and no maximum capacity claim is made here.

Notes

• Base branch is release/v1.0.10, not main.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; both previous findings are addressed and no new actionable issues were found.

Summary

The PR reduces repeated Sparkplug payload reads and parsing, indexes metric merges, and coalesces topology-driven renders.

• Sparkplug alias lookup reuses the payload already decoded for each message.
• Node and device metric updates keep their order without rescanning for every value.
• Topology bursts refresh the page and nodes view through one pending slot.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Topology event] --> B[Request refresh]
    B --> C{Pending slot free?}
    C -->|No| D[Use existing pending refresh]
    C -->|Yes| E[Wait for coalesce window]
    E --> F[Dispatch with generation]
    F -->|Callback starts| G[Release slot and refresh latest state]
    F -->|Task fails or is canceled| H[Release slot only if generation still matches]

Loading

_{Reviews (2) · Last reviewed commit: "fix(ui): recover coalesced refreshes aft..." · Reviewed by Greptile}