fix: Address some issues in the split pdf logic

[awalker4]

This is a copy of the fixes that went into the release/0.25.x branch, updated slightly for the v2 SDK. The main branch is already retrying split requests, but we do not have an updated httpx timeout value, and the retries are not logged.

To verify

Same as the linked pr, we can mock arbitrary errors from the server and verify the behavior.

Apply this patch to your local unstructured-api, and do make run-web-app. Whenever we send a request starting at page 3, the server will sleep for a bit, and then return a 502 error.

diff --git a/prepline_general/api/general.py b/prepline_general/api/general.py
index 2ca56270..0624076c 100644
*** a/unstructured-api/prepline_general/api/general.py
--- b/unstructured-api/prepline_general/api/general.py
***************
*** 812,817 ****
--- 814,822 ----
      )
  
  
+ global num_bounces
+ num_bounces = 2
+ 
  @elasticapm.capture_span()
  @router.post(
      "/general/v0/general",
***************
*** 835,840 ****
--- 840,854 ----
      # For new parameters - add them in models/form_params.py
      form_params: GeneralFormParams = Depends(GeneralFormParams.as_form),
  ):
+ 
+     global num_bounces
+     page_num = form_params.starting_page_number or 1
+     if num_bounces > 0 and page_num == 3:
+         num_bounces -= 1
+         time.sleep(10)
+         raise HTTPException(status_code=502, detail="BOUNCE")
+ 
      # -- must have a valid API key --
      if api_key_env := os.environ.get("UNSTRUCTURED_API_KEY"):
          api_key = request.headers.get("unstructured-api-key")

Use the main branch to make a client request using layout-parser-paper, which should default to 2 page splits, with the second split starting at page 3. Because the httpx timeout is not specified, this split will raise a ReadTimeout error, and get stuck in a retry loop. This is not logged, so it will appear to be a client hang.

from unstructured_client import UnstructuredClient
from unstructured_client.models import shared, operations

filename = "_sample_docs/layout-parser-paper.pdf"

s = UnstructuredClient(
    server_url="http://localhost:5000",
)

with open(filename, "rb") as f:
    files=shared.Files(
        content=f.read(),
        file_name=filename,
    )

req = operations.PartitionRequest(
    shared.PartitionParameters(
        files=files,
        strategy="fast",
        split_pdf_page=True,
    ),
)

resp = s.general.partition(req)

if num_elements := len(resp.elements):
    print(f"Succeeded with {num_elements}")

Finally, make the same request against this branch. The bad split should be retried for the allotted 2 times, with logging, and then the doc will succeed.

[JOSHMT0744]

Really like these changes :)
On test_decorators.py, if endpoint url isn't "/general/v0/general", this could cause issues..?

[awalker4]

Really like these changes :) On test_decorators.py, if endpoint url isn't "/general/v0/general", this could cause issues..?

Thanks for the review! For that test - I was able to change the logic that required that check, so I can remove that now.

[jordan-homan]

does the client user have to know to set this environment variable? if so should the default go the other way around maybe? just thinking about how it might be challenging waiting up to 30 minutes without realizing why it's hanging so long

[awalker4]

We've had some users get stuck here. For very large pdfs, a hi_res call can sit for 30 minutes, at which point they get a httpx.ReadTimeout, the doc gets retried, and the same thing happens again. This should be rare, but we wanted the timeout to account for the worst case. And we added the env var so we could ask users to bump the timeout manually if this still wasn't resolved.

There is a http timeout parameter for the SDK now, but due to some gaps in the hook context, we don't have access to that param from down here. Longer term, that should be the value that users can configure, and we can probably have a much saner default. The env var here is a bit of a bandaid, so I'm keeping it out of the readme, etc.