Stop saving a trip rebuilding every route's geometry

[jjasloot]

Saving a RouteInstance imported from Träwelling took seconds before the list came back with its delay display. Almost none of it was the save.

What was happening

PUT /api/routes/instances commits the trip and then, because an import carries a TraewellingStatusId, computes station suggestions before it answers. The page only repaints after that answer, so everything below sat between the click and the delay display.

The dominant cost: the matcher's spatial index was one object fingerprinted on the route count and the station count. Importing a trip onto a new route bumps the route count, so the very next save found the index stale and rebuilt it — streaming every route's geometry out of the database (12,809 routes, 9.9M coordinates, 382 MB, per the measurements already recorded in MatcherIndex.cs) and running Douglas-Peucker over it, inside the request. MatchStopsAsync, the lookup that triggered this, reads only the station tree and never touches a route segment.

The 30-minute idle timeout meant the first import of any session paid it too.

Changes

461530b — split the index. Route segments and station points are now counted and cached separately, so matching a calling pattern builds only the station tree. FindStationsForTripAsync takes the route's endpoints off the geometry already on its stack instead of asking the route index, which leaves FindStationTripsAsync (the visit-dating and backfill lookups) as the only caller that builds the expensive half.

Three smaller costs on the same path:

• MatchStopsAsync asked the database whether names agreed once per stop, and a calling pattern runs to dozens of stops. Geometry pass first, then one batched query — which also replaces the separate name lookup it did afterwards.
• The save loaded the route with every trip it has ever had, joined across properties and maps, to find the one trip being edited — repeating the route's geometry column on every result row. It now checks ownership, loads that one instance, and reads the geometry only when a duration needs computing.
• The reload the screen waits on had the same joined shape. AsSplitQuery() on both instance reads.

c2355ad — build the index before the backfill queue needs it. Opening the dating queue used to fire the first station request into a cold index: an unexplained spinner, and a request long enough to be at a proxy's mercy. POST /api/stationbackfill/prepare now answers whether the index is warm and otherwise queues the build on MatcherWarmupService, with the page waiting on StationsHub. Because the builder streams rows and the count is known up front, the page shows a determinate bar rather than a spinner.

The queue / hosted-service / hub arrangement mirrors the Träwelling resync, which solves the same problem. Two details worth knowing: the page subscribes before it asks, because asking is what starts the build and connecting after would lose the event of a build that finished in between; and starting the queue is a one-shot claim, so the finished event, the watchdog and a ready answer cannot each load the first station. Preparing is an optimisation, never a gate — any failure falls through to the request that builds the index itself.

511d803 — bound the Träwelling call. The one upstream call after the save was holding the response open for something the user had already got, and rate limited it sleeps its whole window — up to a minute. It is now skipped outright when the limiter is already blocked, and given five seconds otherwise, cancelled through the token the limiter's delay and the HTTP call already honour.

That costs the suggestions for that trip, and they do not come back — they are computed at import and never stored. So the result distinguishes "nothing to offer" from "could not ask", and the page says so in a snackbar with no action button, which clears itself.

The inbox-row delete moved ahead of the fetch: with the request's own token now reaching that call, a client going away mid-fetch would have skipped the delete and left an imported trip sitting in the inbox looking unimported.

Verification

CI green on each commit — backend builds clean and 248 tests pass (243 before; seven added across the three commits). The production frontend build was also run locally against Node 24, so strictTemplates covers the new component and template code.

Not verified: actual timings. This environment has no production data and no .NET SDK, so the size of the improvement is reasoned from the code and the measurements already recorded in it, not from a stopwatch. The mechanism is certain; the magnitude is inference.

Known gaps, left deliberately

• GET /api/station/{id}/candidates (the single-station visit dialog) still hits a cold build unannounced. Rarer path — one station at a time rather than a queue — so it was scoped out.
• The route index still exists only because there is no usable spatial index on Routes.LineString. Precomputing route↔station proximity into a join table, the way RouteRegionsService already derives Route.Regions from geometry, would remove it entirely. Larger change, not attempted here.

🤖 Generated with Claude Code

https://claude.ai/code/session_012EGxqy7LB2FpCSns4JiNgZ

---

Generated by Claude Code