Serve the Data Library from any server, and let a server shape it (#1807)

A server can answer the library with createLibraryHandler, a fetch handler, and shape what it offers: hide entries (listed: false) or route browser-direct ones through itself (fetch: proxied). View-bounded entries ask for a grid-snapped box, and GeoParquet is read on the server within it. Applications can pass their own tile and data fetchers (createApplicationTools({library: {fetchTile, fetchData}})), and with library.requireCatalog a catalog that does not answer offers nothing and is asked again later. Tile paths have one spelling: integer z/x/y without leading zeros, within the grid and the entry's zooms. A server's index and partition reads are bounded and counted against its subrequest limit.
This commit is contained in:
Sameh Khamis authored and GitHub committed 2026-11-17 23:34:00 -08:00
1 parent 7b4c84a717
commit 9554d6ea71
15 files changed
+1510 -395

No files matched your search

+47
View File
@@ -167,6 +167,53 @@ code. It reaches only the manifest's host, refuses redirects, caps bodies,
caches per `cacheSeconds` and never returns upstream error text. The browser
names an entry id, never a URL, so the proxy cannot be pointed anywhere else.
## Serving the library from another server
The proxy is also a fetch handler, `createLibraryHandler` from
`gods-eye-view/sources/library` (`src/library/proxy.js`): a `Request` in, a
`Response` out, for a server other than the dev server (a Worker, say). Its
options decide what the server offers:
- `listed(entry)` false hides an entry: `/catalog` marks it `listed: false`,
Browse leaves it out entirely, and no route serves it.
- `proxied(entry)` true routes a `browser-direct` entry through the handler:
`/catalog` marks it `fetch: "proxied"` and the browser asks the server for
its tiles and data. A view-bounded entry then asks `/data/<id>?bbox=…` for
the camera's box snapped out to a power-of-two grid (`libraryViewBox`), so
nearby views share one answer; any other box is refused.
- `keep: false` keeps nothing in the process's memory, for a server with a
shared cache in front. Answers carry `Cache-Control: public,
max-age=<cacheSeconds>` (`private` for a keyed entry).
- `partitionCache` is the map of decoded GeoParquet partitions (at most four
are kept); pass one map per process to cap them there.
- `fetchIndex(entry)` reads a partitioned entry's index, for example the
handler's own `/index/<id>` through a shared cache. A `/data` request never
builds an index itself: without a complete or partial one it answers 503
`index_unavailable`.
- `subrequestLimit` caps the upstream requests one request may make (a host's
per-request limit); a request that would need more answers 503
`subrequest_limit` instead.
Every resource has one spelling: tile paths are `/tile/<id>/<z>/<x>/<y>` with
integers without leading zeros, `x` and `y` below `2^z` and `z` within the
entry's zoom range; a path with an empty segment or a trailing slash is
refused; a view box is accepted only in `libraryViewBox`'s exact spelling. A
shared cache keyed on the path (and a `/data` request's `bbox`) therefore
holds each answer once.
Same-site gates and rate limits stay with the server that mounts it. An
application can also fetch the library's tiles and proxied datasets itself,
to keep them in its own cache: `createApplicationTools({ library: { fetchTile,
fetchData } })`. `fetchTile` is the imagery spec's tile fetch, told which
entry asks; `fetchData` answers a proxied dataset URL.
`library: { requireCatalog: true }` makes the server's `/catalog` an
allowlist: the library offers only what it names, and while it does not
answer (an error status, a timeout, a page that is not the catalog) it offers
nothing at all, rather than every entry the browser could fetch itself. It is
asked again after 30 seconds, then at doubling intervals up to ten minutes,
and the viewer's saved entries come back once it answers.
## Operator switches
| Variable | Effect |