Symfony version(s) affected

7.0.3

Description

I currently make extensive use of the Async AWS project across varying platforms, and ever since I've upgraded symfony/http-client from version 7.0.2 to 7.0.3 I've seen an influx of issues that I believe were handled accordingly previously, or at least gracefully retried.

I believe that #53506 may have inadvertently caused a regression, or caused unintended side-effects elsewhere in the package.

In terms of real world effects this means I see around 25 exceptions being thrown per 4 million requests, where I previously saw none. It's not a lot, but it's there, and it's noisy.

Please find a snippet of a stack trace below;

        "class": "Symfony\\Component\\HttpClient\\Exception\\TimeoutException",
        "message": "Idle timeout reached for \"https://sts.eu-west-1.amazonaws.com/\".",
        "code": 0,
        "file": "/var/www/html/vendor/symfony/http-client/Chunk/ErrorChunk.php:56",
        "trace": [
          "/var/www/html/vendor/symfony/http-client/Response/AsyncResponse.php:70",
          "/var/www/html/vendor/symfony/http-client/Response/CommonResponseTrait.php:144",
          "/var/www/html/vendor/symfony/http-client/Response/AsyncResponse.php:97",
          "/var/www/html/vendor/async-aws/core/src/Response.php:147",
          "/var/www/html/vendor/async-aws/core/src/Result.php:69",
          "/var/www/html/vendor/async-aws/core/src/Result.php:137",
          "/var/www/html/vendor/async-aws/core/src/Sts/Result/AssumeRoleWithWebIdentityResponse.php:112",
          "/var/www/html/vendor/async-aws/core/src/Credentials/WebIdentityProvider.php:116",
          "/var/www/html/vendor/async-aws/core/src/Credentials/WebIdentityProvider.php:53",
          "/var/www/html/vendor/async-aws/core/src/Credentials/ChainProvider.php:54",
          "/var/www/html/vendor/async-aws/core/src/Credentials/CacheProvider.php:38",
          "/var/www/html/vendor/async-aws/core/src/AbstractApi.php:146",
          "/var/www/html/vendor/async-aws/sqs/src/SqsClient.php:722",

Apologies for truncation of the stack trace, but this should give some insight.

In the example above you see that an idle timeout was reached for https://sts.eu-west-1.amazonaws.com/. It's not like I've got a super restrictive timeout configured for the HttpClient, as a matter of fact, even bumping it to 30 seconds didn't alleviate this issue.

Reverting back to 7.0.2 completely made this issue go away. I'm seeing no failed requests (due to incorrect STS credentials), which I believe is due to the fact that the HttpClient retries idle timeouts accordingly up until 7.0.2.

Of course I could try and handle this by creating a custom credential provider that I'd use with the AsyncAWS packages, but that would kind of defeat the purpose of using the RetryableHttpClient in conjunction with a retry strategy.

How to reproduce

This is nigh impossible to reproduce. As I mentioned, I'm only seeing a handful of exceptions over millions of requests, which makes it incredibly hard to pin down.

Possible Solution

I would suggest temporarily reverting #53506, and seeing which side-effects these changes have in the larger ecosystem of the package. It's clear that an exception now bubbles up, which was retried gracefully previously.

Additional Context

No response