Sleep 1

We should use Wait-UntilTrue and use flags to sync threads.

@daxian-dbw I also don't like sleep timeouts-based synchronization in tests.

Thank you both for the comments. The Sleep 1 is not really necessary, so I removed it.

Would it be better to have this repeated call in a finally block? Or is there a scenario where an error occurs and we don't want to do it?

This is a good question and I'm not really sure if we should put it in a finally block. Don't know about other exceptions, but if the NewArgumentException is thrown from 4 lines above, maybe we don't want to reset the output? I'm inclined to keep it as is given the original code is not in a finally block and we haven't yet heard any issues regarding that.

Okie doke -- thought I'd comment since we call it at the end of 3/4 branches of the code.

This scenario seems to be completely broken, however I don't think this is the right solution. My understanding is that the pattern is that EndInvoke is always called after BeginInvoke, and the output buffer is not released until EndInvoke is called. However, this line will throw a "pipeline stopped" exception, and the output buffer is never released.

psAsyncResult.EndInvoke();              // Stop throws System.Management.Automation.PipelineStoppedException

I think we should catch the pipeline stopped exception and ensure EndInvoke() always releases the output buffer.

try
{
    psAsyncResult.EndInvoke();
}
catch (System.Management.Automation.PipelineStoppedException) { }
finally
{
    if (OutputBuffer) { ... }
}

...
return psAsyncResult.Output

Then the pattern is:

$ps = [powershell]::Create()
$ps.AddScript('1..10 | % { sleep 1; "Output $_" }')
$async = $ps.BeginInvoke()
sleep 3
$ps.Stop()
$output = $ps.EndInvoke($async)
$async = $ps.BeginInvoke()
...

Talked to @PaulHigin offline and the concern is more like "if an terminating exception is thrown from $ps.EndInvoke, there should be a way to get the partial output". However, I found that's already the case now:

$ps.Commands.Clear()
$ps.AddScript("'Hello World'; throw blah") > $null
$a=$ps.BeginInvoke()
$b = $ps.EndInvoke($a)
$b
> Hello World

$ps.AddScript("'Hello World'; [System.IO.FileInfo]::new(`$null)") > $null
$a = $ps.BeginInvoke()
$b = $ps.EndInvoke($a)
$b
> Hello World

As for the PipelineStoppedException thrown from EndInvoke due to $ps.Stop() being called previously, the cleanup is not needed as it's already done in Stop(), and we have to do the cleanup in Stop because with the current behavior, a developer will never call $ps.EndInvoke after calling $ps.Stop because that will raise an exception.

If we really want the user be able to get the partial results when $ps.Stop() is called, then the partial results probably should be returned somehow from the Stop method. The discussion of this would be a separate issue.

If we really want the user be able to get the partial results when $ps.Stop() is called,

I think it could be in EndInvoke() as option (and perhaps timeout option too).

@daxian-dbw You have an error in your example script.

PS C:\> $ps = [powershell]::Create()
PS C:\> $ps.AddScript('"Hello!"; throw "blah"') > $null
PS C:\> $a = $ps.BeginInvoke()
PS C:\> $b = $ps.EndInvoke($a)
Exception calling "EndInvoke" with "1" argument(s): "blah"
At line:1 char:1
+ $b = $ps.EndInvoke($a)
+ ~~~~~~~~~~~~~~~~~~~~~~
+ CategoryInfo          : NotSpecified: (:) [], MethodInvocationException
+ FullyQualifiedErrorId : RuntimeException

PS C:\> $b
PS C:\>

I feel this is fundamentally broken because if a terminating error occurs (including stop) then any data generated up to that point is lost ... a major sin in my book.

I feel that updating PowerShell.Stop() to throw away the output buffer only compounds this problem, while breaking the existing contract of BeginInvoke/EndInvoke pattern.

Some ideas to remedy:

1. Catch pipeline stop exception and insert into Error stream as non-terminating error. But this doesn't address a general terminating exception.
2. Catch all terminating exceptions (in EndInvoke) and insert into Error stream, marking the error as terminating in some way.
3. Create a new "EndInvoke" exception class that is thrown, which wraps the actual terminating error and also provides the output buffer as a parameter.

You have an error in your example script.

Good catch. Yes, I see the problem with throw now. It's weird though [System.IO.FileInfo]::new($null) works as you expected in my second example, which makes it so inconsistent. Please feel free to open an issue for your concern on this behavior (along with the ideas for the potential fix).

I feel that updating PowerShell.Stop() to throw away the output buffer only compounds this problem.

Base on the code below

it looks to me setting OutputBuffer = null does not throw away the output buffer. The PSDataCollection object referenced by OutputBuffer gets passed to CoreInvokeAsync via the parameters output and asyncResultOutput, then output is used by the Worker as the output stream and asyncResultOutput is assigned to PowerShellAsyncResult.Output, which will be used to return the results from EndInvoke.

So, even though OutputBuffer is set to null in Stop() and EndStop, it won't affect the output stream used by the worker, and theoretically the partial results are still available in psAsyncResult.Output when getting to EndInvoke.

while breaking the existing contract of BeginInvoke/EndInvoke pattern.

Having a pair of BeginXXX and EndXXX is the pattern only when you want the execution to complete, either to success or failure.

However, when Stop is called after BeginInvok, I don't think it's a must-have for the user to call EndInvoke(), because

1. that's not something the user can do today and we won't be able to change their existing habit;
2. there is a good chance that the user just doesn't care about the execution anymore after calling Stop;
3. it would be a breaking change if we suddenly tell the user you have to call EndInvok again after Stop to clean up some state.

My concern is that calling PowerShell.Stop() should not lose data, and the change you suggest does this (let me know if I am incorrect here). I admit that EndInvoke() is also broken and loses data but I feel we should fix this rather than compound the problem. I also respectfully disagree that Stop() should work the same as EndInvoke(). My reading of the code is that that was not the intention. Perhaps we should submit this to the committee to get their view.

My concern is that calling PowerShell.Stop() should not lose data, and the change you suggest does this (let me know if I am incorrect here)

Offline @daxian-dbw Dongbo showed me that this is in fact incorrect (I didn't look at the code closely enough).
The issue of EndInvoke() losing data can be treated separately and Dongbo will create an issue to track it. I have no objection to this change.

Thanks!

#8005 is opened to track @PaulHigin's concern on the behavior of PowerShell.EndInvoke.