Limits are not honoured when using httpx-retries without Custom Transport

[yks0000]

httpx.Client or httpx.AsyncClient Limits are not honoured when using httpx_retries

Client Code part of SyncHttpSessionManager:

import httpx
from httpx_retries import Retry, RetryTransport

retries = Retry(
            total=3,
            backoff_factor=0.2,
            status_forcelist=[500, 502, 503, 504],
            allowed_methods=["HEAD", "GET", "POST", "PUT", "DELETE", "OPTIONS"],
        )

transport = RetryTransport(retry=retries)
session = httpx.Client(
            transport=transport,
            http2=True,
            limits=httpx.Limits(max_connections=100, max_keepalive_connections=100, keepalive_expiry=60),
            headers={
                "User-Agent": f"Python/{__version__}",
                "Content-Type": "application/json",
            },
        )
return session

In SDK Interface:

class SDKClient:
  def _init__():
      self.http_sync_client = SyncHttpSessionManager()
  
  def authorize(request):
       return self.http_sync_client.post(....)

Usage

for i in range(5):
    start_time = time.monotonic()
    response = client.authorize(request=request)
    logger.info(f"Time taken for response={(time.monotonic() - start_time) * 1000:.2f} ms. Sleeping for 6 (default+1) second to test keep-alive. ")
    time.sleep(6) 

We intentionally put a sleep above and can see in logs that connection are getting closed after 5 seconds Limits.keepalive_expiry is not respected:

2025-05-09 17:43:39,184 - level=INFO logger=[client] Time taken for response=1747.09 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:43:46,998 - level=INFO logger=[client] Time taken for response=1809.17 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:43:54,755 - level=INFO logger=[client] Time taken for response=1754.75 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:44:02,629 - level=INFO logger=[client] Time taken for response=1869.00 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:44:10,402 - level=INFO logger=[client] Time taken for response=1769.34 ms. Sleeping for 6 (default+1) second to test keep-alive. 

when sleep is 4 seconds, the first call took 1784.65, however other re-used the connection:

2025-05-09 17:50:33,642 - level=INFO logger=[client] Time taken for response=1784.65 ms. Sleeping for 4 second to test keep-alive. 
2025-05-09 17:50:37,946 - level=INFO logger=[client] Time taken for response=300.42 ms. Sleeping for 4 second to test keep-alive. 
2025-05-09 17:50:42,265 - level=INFO logger=[client] Time taken for response=314.88 ms. Sleeping for 4 second to test keep-alive. 
2025-05-09 17:50:46,571 - level=INFO logger=[client] Time taken for response=300.44 ms. Sleeping for 4 second to test keep-alive. 
2025-05-09 17:50:50,884 - level=INFO logger=[client] Time taken for response=310.29 ms. Sleeping for 4 second to test keep-alive. 

As a work-around, created a custom HTTPTransport with Limits

transport = httpx.HTTPTransport(
            limits=httpx.Limits(max_connections=100, max_keepalive_connections=100, keepalive_expiry=60)
        )
transport = RetryTransport(retry=retries, transport=transport)

Again, added wait for 6 seconds again, but this time keep-alive is honored.

2025-05-09 17:55:09,961 - level=INFO logger=[client] Time taken for response=1754.35 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:55:16,292 - level=INFO logger=[client] Time taken for response=325.68 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:55:22,705 - level=INFO logger=[client] Time taken for response=408.59 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:55:29,005 - level=INFO logger=[client] Time taken for response=296.86 ms. Sleeping for 6 (default+1) second to test keep-alive. 
2025-05-09 17:55:35,409 - level=INFO logger=[client] Time taken for response=400.65 ms. Sleeping for 6 (default+1) second to test keep-alive. 

Should Limit passed to client be preferred or this is expected behavior?

Opened an issue on httpx-retries as well will-ockmore/httpx-retries#17

[will-ockmore]

As noted on the linked issue on https://github.com/will-ockmore/httpx-retries , this is currently expected behaviour for any custom transport passed to httpx.Client.

Perhaps we could add a docs issue to clarify this as it's probably a bit confusing. Unless the behaviour itself is unintuitive, and would be better off fixed; but it's not clear to me that it is intuitive to override transport behaviour (and limits is definitely transport behaviour) from the client. Both approaches can be confusing in different scenarios!