Malformed links when setting root-path through uvicorn

[sunu]

When root-path is set through uvicorn's --root-path flag, some of the links returned from stac-fastapi-pgstac are malformed and contain the root path twice.

Steps to reproduce

To reproduce this behavior locally, apply the following diff to the docker-compose file:

diff --git a/docker-compose.yml b/docker-compose.yml
index 2dcafa9..ae2628c 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -89,7 +89,8 @@ services:
   app-nginx:
     extends:
       service: app
-    command: bash -c "./scripts/wait-for-it.sh database:5432 && uvicorn stac_fastapi.pgstac.app:app --host 0.0.0.0 --port 8082 --proxy-headers --forwarded-allow-ips=*"
+    container_name: stac-fastapi-pgstac-nginx
+    command: bash -c "./scripts/wait-for-it.sh database:5432 && uvicorn stac_fastapi.pgstac.app:app --host 0.0.0.0 --port 8082 --proxy-headers --forwarded-allow-ips=* --root-path=/api/v1/pgstac"
     environment:
       - ROOT_PATH=/api/v1/pgstac

This sets the --root-path flag.

Now start the app-nginx service: docker compose up --build app-nginx nginx

Inspect the links returned from the collections endpoint:

curl -s http://0.0.0.0:8080/api/v1/pgstac/collections | jq '.links'
[
  {
    "rel": "root",
    "type": "application/json",
    "href": "http://0.0.0.0:8080/api/v1/pgstac/"
  },
  {
    "rel": "self",
    "type": "application/json",
    "href": "http://0.0.0.0:8080/api/v1/pgstac/api/v1/pgstac/collections"
  }
]

You would notice that the href in the second link entry is malformed and contains the root path /api/v1/pgstac twice.

Probable cause

There was a refactoring into how root path is handled by uvicorn in Kludex/uvicorn#2213
This was released as part of uvicorn 0.26.0 (https://github.com/encode/uvicorn/blob/master/docs/release-notes.md#0260-january-16-2024)
This was integrated into stac-fastapi-pgstac in 7135059 and released as part of the 4.0.0 release in #196

After uvicorn 0.26.0 and stac-fastapi-pgstac 4.0.0, all requests going through uvicorn go to /{root-path}/requested-path. This also explains behavior in developmentseed/eoapi-k8s#195 (comment)

Possible fix

One way to get around this is of course to not set root path through uvicorn's --root-path flag and use the ROOT_PATH env var instead. But I don't think that's a good solution.

Here's a patch for how to handle this during link generation:

diff --git a/stac_fastapi/pgstac/models/links.py b/stac_fastapi/pgstac/models/links.py
index 6697488..5d72de7 100644
--- a/stac_fastapi/pgstac/models/links.py
+++ b/stac_fastapi/pgstac/models/links.py
@@ -51,7 +51,30 @@ class BaseLinks:
     @property
     def url(self):
         """Get the current request url."""
-        url = urljoin(str(self.request.base_url), self.request.url.path.lstrip("/"))
+        # base_url is of the form http{s}://host{:port}/{root_path}
+        base_url = self.request.base_url
+        path = self.request.url.path
+        # root path can be set in the request scope in two different ways:
+        # by uvicorn when running with --root-path
+        # or by FastAPI when running with FastAPI(root_path="...")
+        # When root path is set by uvicorn, request.url.path will have the root path prefix.
+        # eg. if root path is "/api" and the path is "/collections",
+        # the request.url.path will be "/api/collections"
+        # We need to remove the root path prefix from the path before
+        # joining the base_url and path to get the full url to avoid
+        # having root_path twice in the url
+        if self.request.scope.get("root_path") and not self.request.app.root_path:
+            # self.request.app.root_path is set by FastAPI when running with FastAPI(root_path="...")
+            # If self.request.app.root_path is not set but self.request.scope.get("root_path") is set,
+            # then the root path is set by uvicorn
+            # So we need to remove the root path prefix from the path before
+            # joining the base_url and path to get the full url
+            root_path = self.request.scope["root_path"]
+            if path.startswith(root_path):
+                path = path[len(root_path) :]
+        path = path.lstrip("/")
+
+        url = urljoin(str(base_url), path)
         if qs := self.request.url.query:
             url += f"?{qs}"

We deployed an image with this patch applied for IFRC Montandon and it worked quite well. @vincentsarago let me know if this looks ok and I'd love to create a PR

[vincentsarago]

@sunu did you try with 4.0.3? we pushed some changes in #195

[sunu]

@vincentsarago yes and this is reproducible with the latest from main

[vincentsarago]

looking at the docker-compose you copy/pasted I think you shouldn't have both --root-path in the uvicorn cmd and also have ROOT_PATH in the environment

also you should hit the :8080 url but the one from nginx :8082

[sunu]

@vincentsarago starlette respects the root path set by uvicorn, so the ROOT_PATH env var doesn't really affect anything if --root-path is already set through uvicorn. But still, commenting out ROOT_PATH env var produced the same behaviour.

Also both hitting the api through nginx (:8080) and directly hitting uvicorn (:8082) produce the same behaviour:

$ curl -s http://0.0.0.0:8082/collections | jq '.links'
[
  {
    "rel": "root",
    "type": "application/json",
    "href": "http://0.0.0.0:8082/api/v1/pgstac/"
  },
  {
    "rel": "self",
    "type": "application/json",
    "href": "http://0.0.0.0:8082/api/v1/pgstac/api/v1/pgstac/collections"
  }
]

curl -s http://0.0.0.0:8080/api/v1/pgstac/collections | jq '.links'
[
  {
    "rel": "root",
    "type": "application/json",
    "href": "http://0.0.0.0:8080/api/v1/pgstac/"
  },
  {
    "rel": "self",
    "type": "application/json",
    "href": "http://0.0.0.0:8080/api/v1/pgstac/api/v1/pgstac/collections"
  }
]