Skip to content

Second-pass performance hardening

Motivation

A second read-only critique found remaining latency and deploy-loop costs after the first performance batch. The highest-impact residuals were repeated subscription enumeration on /api/me, large Storage listing defaults, hidden-tab frontend timers, oversized initial Vite chunks, mutable terminal toolchain refs, and avoidable quick-deploy provider-registration checks.

User-facing change

Dashboard startup, large-result navigation, and background-tab behaviour should be lighter. Fast deploy iterations should also avoid unnecessary provider checks and rebuild drift. No API response schema changed.

Code-level diff summary

  1. /api/me now caches visible subscription discovery for 60 seconds (ME_SUBSCRIPTIONS_TTL_SECONDS) and caps the list at 100 subscriptions (ME_SUBSCRIPTIONS_LIST_LIMIT).
  2. Storage monitor usage defaults now cap per-container scans at 10,000 blobs, and result blob listing caps at 5,000 rows by default (STORAGE_RESULT_BLOB_LIST_LIMIT).
  3. Request detail inspector body capture is opt-in by default via REQUEST_DETAIL_CAPTURE_ENABLED=true, avoiding mutation-body buffering in normal runs.
  4. The frontend Vite build now splits React, TanStack Query, xterm, BLAST submit/results, and terminal chunks explicitly.
  5. The high-frequency BLAST jobs/history/log timers touched in this batch pause in hidden tabs, and BLAST job rows/details are memoized where the render path was repeatedly rebuilding styles or rows. Other long-running operation pollers keep their existing lifecycle semantics.
  6. Terminal toolchain builds pin kubectl, ttyd, and the sibling elastic-blast-azure ref instead of fetching floating latest/master refs.
  7. Quick deploy provider registration has a one-hour marker cache and an explicit SKIP_PROVIDER_REGISTRATION=true fast path.
  8. Docker build contexts exclude generated web/api/test artifacts, and the frontend Docker build pins the Node base image.

API / IaC diff

No route schema changes and no Bicep changes in this second-pass batch. Docker and deploy-script behaviour changed for deterministic/cached builds.

Validation evidence

  • uv run pytest -q api/tests/test_me_route.py api/tests/test_storage_data.py api/tests/test_request_metrics_detail.py — 45 passed.
  • uv run pytest -q api/tests — 1373 passed.
  • uv run ruff check on touched Python files — clean.
  • cd web && npm run build — passed. The manual chunk split reduced the former single large app bundle into route/vendor chunks; two lazy route chunks still exceed Vite's 500 kB warning threshold and remain candidates for deeper route-level lazy loading.
  • bash -n scripts/dev/quick-deploy.sh scripts/dev/terminal-base-image.sh scripts/dev/acr-build-access.sh — passed.
  • curl -fsI https://github.com/tsl0922/ttyd/releases/download/1.7.7/ttyd.x86_64 — passed.
  • git fetch --depth 1 origin f4b8b734a82285a18a2ca9aadcbe02759d13f903 against dotnetpower/elastic-blast-azure — passed.
  • az bicep build --file infra/main.bicep --stdout /tmp/elb-main-bicep-build.json — passed with only the Azure CLI Bicep version-upgrade warning.
  • git diff --check — clean.