Collect browser-side worker round-trips, computation times, main-thread
operations, and feature counts, batch them client-side, and expose as
Prometheus histograms via a new POST /api/perf endpoint.
The jq filter required pods to be younger than 60 seconds AND ready,
but pods often take longer than 60s to become ready (alembic migrations,
image pull, etc.), causing the check to never match. The image tag check
is sufficient to confirm the new version is running. This aligns the API
verify-deploy with the frontend verify-deploy which has no age filter.
R-tree building, hex grid generation, and percentile sorting now run
off the main thread, eliminating 20s+ UI freezes on large datasets.
The old bundled HexgridHeatmap.js is replaced by a typed worker +
main-thread client with dual R-trees (worker for grid gen, main
thread for synchronous click queries).
Listing stream fires immediately on auth without waiting for POI
fetch. POI distances are not needed for initial rendering and are
only computed when user selects POI metric or sets travel filters.
This saves ~200-500ms on initial load and keeps the stream on the
cached Redis path.
New GET /api/poi/distances/bulk returns all POI distances keyed by
listing ID, allowing the frontend to fetch distances separately
from the listing stream and keep the stream on the cached path.
Decision IDs are now loaded inside the streaming generators after
the metadata message is yielded, eliminating a blocking DB query
from the pre-stream path (~100-200ms improvement to TTFB).
Subsequent batches use the normal batch_size (default 50). This
reduces server-side time-to-first-property by ~10x since only 5
features need to be serialized before the first yield.
Overrides LimitRange defaults (500m CPU) which caused kernel CPU
throttling during streaming requests. API gets 2000m CPU limit,
celery-beat gets 200m.
Measured baseline: 877ms TTFP cold, 334ms warm, with CPU throttling
confirmed (500m limit, 10-12 throttle events per streaming request).
Plan covers: CPU limit bump, frontend waterfall elimination, adaptive
first batch, deferred decision filtering, POI decoupling, and
Server-Timing headers.
The X button aborts the in-flight fetch via AbortController,
which was already wired up but had no UI trigger. Works for
both desktop and mobile views.
Move drag handler from outer card to details section only.
Swiping on photos now navigates the carousel, swiping on
the details area below triggers like/dislike/skip actions.
Rightmove API stores photos under the 'photos' key in the response,
but the GeoJSON export and detail API were only checking 'images'.
This key mismatch caused all listings to fall back to the single
photo_thumbnail. Now checks both keys with fallback.
- Use maxSizeUrl instead of url for photo URLs (highest available
resolution from Rightmove)
- Remove 5-photo cap in GeoJSON export — return all available photos
- Apply same fix to both streaming and model-based export paths,
and to the listing detail API endpoint
- Card now fills available height with photo carousel in top half
and property details in bottom half
- Details section shows: price, beds/sqm/price-per-sqm, agency,
available date, all POI distances, and price history summary
- Fix DialogTitle accessibility warning in ListingDetailSheet and
MobileBottomSheet (add sr-only Drawer.Title)
SwipeCard now shows all available photos (up to 5) using
embla-carousel instead of just the thumbnail. Includes prev/next
arrow buttons and dot indicators. The photo area uses touch-action:
pan-x so carousel swipes don't trigger card swipes.
The age filter rejected pods older than 60s, but by the time the
verify step runs after build+publish+deploy, the new pods are often
already past that window. The image tag check is sufficient to
identify the correct deployment.
Same pattern as the frontend pipeline: build pushes to a staging
tag (build-N) in parallel with tests. Tests split into unit and
integration shards using a shared venv from the workspace. Publish
step uses skopeo to copy the manifest to final tags (:N, :latest,
:builder) only after all tests and the build succeed. Named the
final Dockerfile stage 'production' so target skips the test stage.
Build step now pushes to a staging tag (build-N) in parallel with
test shards. A new publish step uses skopeo to copy the manifest to
the final tags (:N and :latest) only after all tests pass. This is
a server-side manifest copy with no layer re-upload.
- Add onTap callback to SwipeCard using useDrag's tap detection
- Wire through SwipeReviewMode to open ListingDetailSheet on tap
- Fix color overlay misalignment: add relative to card container so
the absolute overlay positions within the rounded card, not the
full-width outer wrapper
Split the monolithic "build and test" kaniko step into a DAG:
tests run in 4 parallel shards (vitest --shard) alongside the
Docker image build, gated by a shared npm ci step. The kaniko
build now targets the named 'production' stage to skip the
in-Dockerfile test stage.
- Replace pip with uv for 10-25x faster dependency installation
- Add BuildKit cache mounts for apt and uv caches across builds
- Add non-root appuser for improved container security
- Add HEALTHCHECK directive for container orchestration
- Add curl to runtime for healthcheck support
- Remove libgl1 (unused), add syntax directive for BuildKit
- Rewrite csv_exporter.py to use stdlib csv.DictWriter instead of pandas DataFrame
- Rewrite notifications.py to use aiohttp direct Slack webhook instead of apprise
- Switch opencv-python to opencv-python-headless in pyproject.toml
- Move httpx from dev to prod dependencies
- Remove pandas and apprise from mypy ignore_missing_imports
- Remove watchdog (unused) and tqdm from pyproject.toml dependencies
- Replace tqdm.gather() with asyncio.gather() + logger.info() in
image_fetcher, floorplan_detector, and route_calculator services
- Replace tqdm progress bar with logger.info() in listing_repository
- Remove tqdm from mypy ignore_missing_imports overrides
Backend: include first 5 photo URLs from additional_info in GeoJSON
streaming response, with fallback to photo_thumbnail.
Frontend: replace single thumbnail with swipeable embla-carousel on
compact cards. Remove window.open on card tap so clicking opens the
detail bottom sheet instead of navigating to Rightmove.
Requests to Rightmove API previously had no explicit timeout, causing
hung connections to block workers indefinitely. Add a configurable
request_timeout (default 30s) to ScraperConfig and apply it to all
aiohttp sessions. Also retry on TimeoutError in addition to
ThrottlingError for all API query functions.
- Update test mocks from _get_disliked_ids to _get_user_id_safe
- Fix decision service test method names (clear_decision -> remove_decision, etc.)
- Fix positional vs keyword arg assertion in set_decision test
- Add decision_filter param to non-streaming listing_geojson endpoint
