[Bug]: Timeout behavior for long-running agent responses

[benedict-khoo-sap]

What happened?

Observed behavior

When response chunks from AgentExecutor.execute take more than several seconds to produce, a 503 timeout error is received on the A2A SDK client when using certain transport protocols and streaming/polling configurations.

Methodology

For my tests, I used

• v0.3.22 of the SDK
• 10s delay before the first response chunk
• 10s delay after the first two response chunk
• JSONRPC and HTTP+JSON
• Streaming and non-blocking polling via tasks/get

Code to reproduce results: https://github.com/benedict-khoo-sap/a2a-python-long-running-response-demo

Summary

Here is a summary table of the results observed:

Server Type	Mode	10s delay before the first response chunk	10s delay after the first two response chunk	all response chunks sent quickly
JSONRPC	Streaming	❌ Timeout error after several seconds	⚠️ First 2 chunks received, then timeout error after several seconds	✅ Works fine
JSONRPC	Polling	❌ Timeout error after several seconds	✅ Works fine	✅ Works fine
HTTP+JSON	Streaming	✅ Works fine	✅ Works fine	✅ Works fine
HTTP+JSON	Polling	❌ Timeout error after several seconds	✅ Works fine	✅ Works fine

Potential related issue

#545 may be the cause of the issue in JSONRPC implementations, but HTTP+JSON with polling has the same timeout error if the first response chunk takes too long to return

Relevant log output

Code of Conduct

• I agree to follow this project's Code of Conduct

[guglielmo-san]

Hi @benedict-khoo-sap , I believe that your issue is due to the default timeout value the httpx client is using.
You can fix this passing an httpx_client in the ClientConfig of the ClientFactory, with a custom timeout value.

client_factory = ClientFactory(
        config=ClientConfig(
            httpx_client=httpx.AsyncClient(timeout=20),
            streaming=not force_polling,
            polling=True,
            supported_transports=[
                TransportProtocol.jsonrpc,
                TransportProtocol.http_json,
            ],
        ),
    )

[benedict-khoo-sap]

Hi @benedict-khoo-sap , I believe that your issue is due to the default timeout value the httpx client is using. You can fix this passing an httpx_client in the ClientConfig of the ClientFactory, with a custom timeout value.

client_factory = ClientFactory(
        config=ClientConfig(
            httpx_client=httpx.AsyncClient(timeout=20),
            streaming=not force_polling,
            polling=True,
            supported_transports=[
                TransportProtocol.jsonrpc,
                TransportProtocol.http_json,
            ],
        ),
    )

@guglielmo-san
Ah, this works. Thank you! Feel free to close this issue, but just one oddity I'm wondering about:

If the problem was on the client's side, then why did the client not timeout when the server type was HTTP+JSON (streaming)?

[guglielmo-san]

There is a difference between the HTTP+JSON and JSONRPC implementation.
While in HTTP+JSON the timeout is set to infinite

a2a-python/src/a2a/client/transports/rest.py

Line 149 in 2acd838

modified_kwargs.setdefault('timeout', None)

In JSONRPC it ueses on the timeout of the httpx_client

a2a-python/src/a2a/client/transports/jsonrpc.py

Lines 162 to 164 in 2acd838

	modified_kwargs.setdefault(
	'timeout', self.httpx_client.timeout.as_dict().get('read', None)
	)

We are going to fix this inconsistency in v1.0.0
For now I improved the error handling for timeout exceptions (PR #690)