flaky: `parallel/test-http2-*` on Windows

[Trott]

• Version: v11.0.0-pre (master)
• Platform: win2008r2-vs2017,
• Subsystem: test http2

• test file - parallel/test-http2-pipe
• CI job windows/17290/COMPILED_BY=vs2017,RUNNER=win2008r2-vs2017,RUN_SUBSET=0
• CI worker test-rackspace-win2008r2-x64-3
• output:

not ok 251 parallel/test-http2-pipe
  ---
  duration_ms: 0.323
  severity: fail
  exitcode: 1
  stack: |-
    (node:6548) ExperimentalWarning: The http2 module is an experimental API.
    events.js:167
          throw er; // Unhandled 'error' event
          ^
    
    Error: read ECONNRESET
        at TCP.onread (net.js:658:25)
    Emitted 'error' event at:
        at emitErrorNT (internal/streams/destroy.js:82:8)
        at emitErrorAndCloseNT (internal/streams/destroy.js:50:3)
        at process._tickCallback (internal/process/next_tick.js:63:19)
  ...

@nodejs/build @nodejs/platform-windows @nodejs/testing @nodejs/http2

[refack: edited info and added link to test file]

[apapirovski]

This happens as a result of the recent change to destroy the socket properly but it might be revealing an actual bug rather than just being a flaky test. That test does a pretty basic pipe operation, it shouldn't ever trigger an ECONNRESET.

ping @nodejs/http2

[apapirovski]

This also just happened in parallel/test-http2-client-upload. Same root cause. It's definitely a bug.

https://ci.nodejs.org/job/node-test-binary-windows/17277/COMPILED_BY=vs2017,RUNNER=win2016,RUN_SUBSET=3/tapResults/

[Trott]

For posterity (since the link provided by @apapirovski will eventually 404):

not ok 231 parallel/test-http2-client-upload
  ---
  duration_ms: 0.220
  severity: fail
  exitcode: 1
  stack: |-
    (node:2492) ExperimentalWarning: The http2 module is an experimental API.
    events.js:167
          throw er; // Unhandled 'error' event
          ^
    
    Error: read ECONNRESET
        at TCP.onread (net.js:658:25)
    Emitted 'error' event at:
        at emitErrorNT (internal/streams/destroy.js:82:8)
        at emitErrorAndCloseNT (internal/streams/destroy.js:50:3)
        at process._tickCallback (internal/process/next_tick.js:63:19)
  ...

Host: test-azure_msft-win2016-x64-2

The fact that it's happening on both Azure and Rackspace, win2008 and win2016, seems to suggest (to me at least) a bug in Node.js or a dependency like libuv.

[apapirovski]

Pretty sure it's because we use setImmediate to destroy rather than having some more certain method of confirming that everything has been sent and we're truly done. Hopefully someone else from the http2 team will have a look soon. I'm not on Windows and it's not a problem on macOS as far as I can tell.

[mcollina]

The bug on pipe happens on a mac too.

[apapirovski]

@mcollina I can't seem to replicate on mine :( This is what I run it with: tools/test.py -j 96 --repeat 960 parallel/test-http2-pipe.

[Trott]

I also can't replicate on macOS nor do I see instances looking through macOS failures in Ci. It may very well be a bug on macOS too, but if so, we're definitely seeing it with much greater frequency on Windows in CI.

[apapirovski]

I'm testing a possible fix in https://ci.nodejs.org/job/node-stress-single-test/1845/ and https://ci.nodejs.org/job/node-test-commit-windows-fanned/18023/

[mcollina]

=== release test-http2-pipe ===
Path: parallel/test-http2-pipe
(node:37337) ExperimentalWarning: The http2 module is an experimental API.
events.js:167
      throw er; // Unhandled 'error' event
      ^

Error: read ECONNRESET
    at TCP.onread (net.js:658:25)
Emitted 'error' event at:
    at emitErrorNT (internal/streams/destroy.js:82:8)
    at emitErrorAndCloseNT (internal/streams/destroy.js:50:3)
    at process._tickCallback (internal/process/next_tick.js:63:19)
Command: out/Release/node /Users/matteo/Repositories/node/test/parallel/test-http2-pipe.js
[01:05|% 100|+ 938|-  22]: Done

On my system.