Performance

The product budget is p95 < 200 ms for list/board endpoints at 100k items per org, and search < 300 ms. This page gives the current numbers and what keeps them there.

Baseline (September 2026)

Measured locally against the compose stack seeded with the perf organization:

Script	Load	p95	Target
list.js	20 req/s	15 ms	< 200 ms
board.js	5 req/s	77 ms	< 200 ms
search.js	10 req/s	20 ms	< 300 ms
detail.js	10 req/s	6 ms	< 200 ms
mcp.js (list_ready_work)	1 req/s	50 ms	< 300 ms

What guards the budget

• Nightly k6 run against compose - perf/*.js, driven by .github/workflows/perf.yml (cron, plus workflow_dispatch). It builds the branch, boots the deploy stack, seeds the perf organization, runs five load scripts whose thresholds are the budget, and uploads the summaries plus a pg_stat_statements top-10 dump as the perf-results artifact. A failed threshold fails the job. See perf/README.md for running it locally.
• The perf dataset - dotnet run --project backend/src/Api -- aictiq-seed-perf (or docker compose run --rm api aictiq-seed-perf) seeds an organization perf with 20 projects × 5,000 items (100k), 400k history rows, 50k comments, sprints, labels and one default team per project, in about 2½ minutes. Deterministic; skips if the organization already exists.
• Query-count tests - backend/tests/IntegrationTests/Perf/QueryCountTests.cs locks the number of SQL queries the hot endpoints issue (via a DbCommandInterceptor wired up by ApiTestContext.CreateAsync(countQueries: true) and asserted with AssertAtMost/AssertSameAs in QueryCount.cs). An N+1 that sneaks into a page assembly fails the suite long before the nightly does.
• pg_stat_statements - preloaded by the compose Postgres and created on first init of a fresh volume (deploy/postgres/init-statements.sql).

How the heavy paths work

• Board columns are paged. Every matching card's placement is read as a narrow (id, state_id, board_column_id) projection so counts and WIP stay exact, but full cards are built only for the first take of each column (default 50, the SPA asks for 30); expand=columnId:n grows one column as it is scrolled.
• Subtasks load on demand. A card's subtasks are fetched only when the card is opened, not for every card on the board.
• Key lookups use an index. work.items (organization_id, project_key, number) serves every lookup that addresses an item by its human key: item detail and the MCP tools that filter on project_key.