Memory leak when creating many HttpClients

[blaugold]

Observed on:

Dart SDK version: 2.19.6 (stable) (Tue Mar 28 13:41:04 2023 +0000) on "macos_arm64"
Dart SDK version: 3.0.2 (stable) (Tue May 23 08:26:58 2023 +0000) on "macos_arm64

The following program has a memory leak:

import 'dart:io';

void main() async {
  final pendingWork = <Future<void>>[];
  while (true) {
    if (pendingWork.length >= Platform.numberOfProcessors) {
      await Future.any(pendingWork);
    } else {
      final work = asyncWork();
      pendingWork.add(work);
      work.then((_) => pendingWork.remove(work));
    }
  }
}

Future<void> asyncWork() async {
  final client = HttpClient();
  try {
    final request =
        await client.openUrl('GET', Uri.parse('https://google.com'));
    final response = await request.close();
    await response.drain();
  } finally {
    client.close();
  }
}

Why do I create a new HttpClient for each request? I noticed the memory leak when using Isolate.run to parallelize work, where each Isolate needs its own HttpClient. But the multiple Isolates don't seem to be the problem, as the code above shows.

In the actual app I'm using package:http, which has the same issue, but I tried to make the repro as minimal as possible.

[a-siva]

@blaugold are you running your app connected under devtools when you observe the memory leak? Wanted to confirm if what you are seeing is a duplicate of flutter/devtools#3290

[blaugold]

@a-siva I noticed the issue in production where we use an AOT build and replicated it running the repro locally with JIT but never using devtools, so I don't think its a duplicate.

[a-siva]

@blaugold could you give us some information on what tools you are using to observe the memory leak. I tried your test application above by running it for a long time and I did not observe any leaks while monitoring the Dart heap size or the native RSS.

[ GC isolate   | space (reason)           | GC# | start | time | used (MB)   | capacity MB | external| used (MB)     | capacity (MB) | external MB |  buffer | new  | old   ]
[              |                          |     |  (s)  | (ms) |before| after|before| after| b4 |aftr| before| after | before| after |before| after| b4 |aftr| (MB) | (MB)  ]
[ main         ,    Scavenge(   new space), 27201, 9836.36,   5.1,  32.0,   1.2,  32.0,  32.0, 15.8, 1.5,   10.3,   10.3,   30.8,   30.8,   0.7,   0.7,   1,   1, -30.8,    0.0, ]
[ main         ,    Scavenge(   new space), 27202, 9837.03,   5.3,  32.0,   1.3,  32.0,  32.0, 16.0, 1.6,   10.3,   10.3,   30.8,   30.8,   0.7,   0.7,   1,   1, -30.7,    0.0, ]
[ main         ,    Scavenge(   new space), 27203, 9837.73,   5.5,  32.0,   1.3,  32.0,  32.0, 15.8, 1.6,   10.3,   10.3,   30.8,   30.8,   0.7,   0.7,   1,   1, -30.7,    0.0, ]

[blaugold]

That's interesting. We first noticed the memory issue in our Cloud Run deployment, which started to run out of memory after a certain amount of time, after we started creating many more HttpClients than before. Then I tried to replicate the issue locally on my MacBook using Activity Monitor:

The highlighted process starts at around ~50 MB and slowly accumulates more memory use.

[blaugold]

Here are a few more infos for the process:

[blaugold]

Seems like this is the same issue as #53113. When I replace https://google.com with http://google.com the code above does not exhibit the described leak.

[a-siva]

Seems like this is the same issue as #53113. When I replace https://google.com with http://google.com the code above does not exhibit the described leak.

Thanks for that clarification, I will dup this issue into that one.

[a-siva]

Marking as a duplicate of #53113 and closing.