A snapshot: read at commit 707a04c7ce8 on 2026-10-18T00:06:01Z by majordomus 0.29.0. Nothing on this page is read again. See it live

I2488

The HTTP stack is chosen by measurement: the same routes and several hundred open streams on tiny_http and on axum

Where it stands

devtask.issue

Readiness: ready no dependency and no gate holds it; nobody has started it
Status: READY
Can start: startable
Milestone: development-runtime
Milestone status: ACTIVE
Wave: 2
Waiting on: none
Holds back: none
Behind it in all: 0
Evidence: 0 of 2

Moves

plan.transition

What it is

Objective: Decide which HTTP stack the shared server keeps, from a measurement of the two candidates on the same routes, a streamed route and long-lived subscribers, and record the decision so that the repository has one stack and knows why.
Why: ADR 0142, part 3, and step U6 of the owner's UI-performance pack, which the caching work deferred because a cached capability already answers in under 6 ms at p95. That measurement did not include what a streamed UI adds: a connection held open per page. The server runs each request on a thread of its own, at most 64 at once (http/server.rs MAX_HANDLERS), writes a response whole, and has no compression. The owner asked on 2026-10-12 for a top-notch architecture; the pack says the choice is made on same-route benchmarks and that two stacks are not kept.
Now: tiny_http 0.12, synchronous, four accept threads, a thread per request, 64 handlers at most, WebSocket upgrades on threads of their own, no compression, Content-Length always.
Wanted: The bench harness runs one system target on both stacks behind the same router and executor: a cached JSON route, a server-rendered screen, a streamed response, and N idle WebSocket subscribers (N in 64, 256, 1024) while the first three are timed; p50, p95, p99, memory and threads, on a release build on an idle host. The numbers and the cost of the move (the adapter over Context::execute, blocking handlers on a bounded pool, the deadline and admission rules of I2156 and the remote-caller guard carried over) are written as a decision record. If the decision is to move, the move is its own issue and removes tiny_http in the same change.
Done when: unknown: the record declares no `completion`
Risk: unknown: the record declares no `risk`

Acceptance

One bench target times the same three routes on both stacks, with and without N idle subscribers, and its results are recorded with the commit, the host and the build profile
The decision record states the numbers it was decided from and what the move costs, and proposes one stack
No second HTTP stack is left in the default build

Validation: bin/majordomus plan validatemajordomus bench system.http --profile quick --no-write
Evidence required: http_stack_benchmark_resultshttp_stack_decision_record
Evidence present: unknown: the record carries no `evidence` entries

Scope

Touches: apps/majordomus-cli/src/benchapps/majordomus-cli/src/httpapps/majordomus-cli/Cargo.toml.ai/repo/adrsdocs/PERFORMANCE.mdtest/cases.ai/repo/project/issues/I2488.yaml
Not: unknown: the record declares no `non_scope`
Depends on: I2486
Priority: p1
Profile: deep-work
Parallel safe: true
Owner: unknown: the record declares no `owner`

The work so far

Branches: none
Merged: none
Commits: 0
Trunk: master
Sessions: none
Created: 2026-10-12
Updated: 2026-10-12
Started: unknown: the record declares no `started_at`
Verified: unknown: the record declares no `verified_at`
Completed: unknown: the record declares no `completed_at`

Synchronisation

Adapter: scripts/github-sync
Repository: unknown: the derived plan's header carries no repository
External issue: unknown: a live GitHub state; this executable makes no network call, so only `scripts/github-sync --check` can answer it
State there: unknown: a live GitHub state; this executable makes no network call, so only `scripts/github-sync --check` can answer it

Findings

Level	Code	Message	Reproduce
WARN	`scope_conflict`	shares test/cases with I2652; they may not run concurrently	`majordomus plan validate`