#ensure: issues (was: Re: [Pharo-project] Pharo by Example vol 2: new chapter available)

On 3/3/2010 6:32 PM, Eliot Miranda wrote:

No, it is /broken/.  What exception handlers are in effect when an
unwind is run from another process?  Its not just tat process identity
isn't being preserved (which I would argue is important but not
essential) its that unwinds are being run in an entirely arbitrary
context which has nothing to do with their running normally.  e.g.

givethAndTakethAway: aCollection
     [[aCollection do: [:k| mydict at: k put: self thing].
       self doStuff] ensure: [aCollection do: [:k| mydict removeKey: k]]
        on: KeyNotFoundError
        do: [:ex| ex proceedWith: nil]

will remove whatever keys in mydict are in aCollection when run
normally, but when run if terminated from another process will break if
any keys in aCollection have already been removed from mydict.

Yes, but my claim was that this a feature, not a bug, for a value of "feature" which means that the terminator (<- I really wanted to use that :-) is capable of controlling the exception context of the ensure block.

Running unwinds in another process is /broken/, period.

The reason why I'm not convinced of that, period or not, repetitions or not, is because we've seen no evidence of it ever being a problem. If it's not a problem, neither in our production environments that have triggered all the problems and the fixes for process, semaphore, mutex, termination etc. handling over the last years, nor in messages or bug reports that I've seen on this or other lists, I'm just not buying the strong claim you are making.

Where is the *evidence* to support your claim? Not the theoretical discussion of what could possibly go wrong, but rather the actual, day-to-day things that people really get bitten by.

You could do this better by saying e.g.

      p terminateOn: Error do: [:ex| ex return "skip it"]

and have terminateOn:do: et al install the exception handler at the
bottom of stack.

Well, not really. In a realistic situation you'd probably want to handle multiple conditions, timeouts etc. and stitching the stack together for that is likely way beyond your average Squeaker :-)

[BTW, I don't get why you call this "better"; it would seem that if you call this "better" because it's using in-process termination that your reasoning is cyclical]

   There are obviously issues if the unwind actions refer to process
   local state; but then again perhaps providing said context for the
   duration of the unwind is the right thing, along the lines of:

   Process>>terminate

     self isActiveProcess ifFalse:[
       Processor activeProcess evaluate: [self unwind] onBehalfOf: self.
     ].

   This for example would guard errors during unwind as initially
   intended while ensuring that process relative state is retrieved
   correctly.


Yes, but what if I do something like

         [processMemo add: Processor activeProcess.
          self doStuff]
                ensure: [processMemo remove: Processor activeProcess]
?

It would work fine, why? Did you not notice that I used #evaluate:onBehalfOf: in the above for precisely this reason?

It may have facets but that doesn't mean that running unwinds in other
than the current process isn't badly broken, and isn't relatively easy
to fix ;)

I don't buy this, and claiming it repeatedly doesn't make it any more real unless you put up or shut up. I think you're vastly underestimating the resulting fallout (does that sound like something I've said before? ;-)

For example, here is an interesting one (and please use VW for reference here if you'd like; I'd actually be interested in finding a *small* download myself to try some of this stuff): One of the issues in Levente's example of terminating a process halfways through an unwind block is that if the unwind block is ill-behaved you can really be in for trouble. Let's say you have this:

[
 "just to have something to interrupt"
 [(Delay forMilliseconds: 100) wait. true] whileTrue.
] ensure:[
 "this is the actual killer"
 [(Delay forMilliseconds: 100) wait. true] whileTrue.
].

If you wish to support completing a well-behaved ensure block that has been started, it seems tricky to see how this process would actually be terminated, and how you'd find out that it didn't. In Squeak it's simple - the fact that the process sending #terminate is still in this code tells you it's not finished so one can (for example) associate a timeout with #terminate and catch that. Or, one can just interrupt that process and close the debugger. Both are quite reasonable options to deal with the invariant.

Under the assumption that you'd like to support the well-behaved case, what do you do when using in-process termination in a non-well-behaved case? How do you find out that a termination request didn't finish, and what do you do in practice to kill it?

Also, what are the priority implications of the different schemes? One of the more interesting issues here is that in cases like on our servers I could see how one of the lower priority processes times out due to starvation, is then terminated but this may never complete if it's run from the priority it's at instead of the terminator process. Does for example VW bump the priority of the terminated process? How does that interact when the terminated process isn't well-behaved and runs wild?

I think there are numerous interesting aspects that we have somewhat reasonable answers in Squeak today that I don't even know what the implications would be if you'd go for in-process termination.

Cheers,
 - Andreas

On 3/7/2010 3:17 PM, Igor Stasenko wrote:

My 2 cents.

Process termination should run in the same context as process itself.

Just FYI, Eliot and I had a very good discussion about this on friday. We both agree that it is advantageous to run unwind blocks inside the terminating process. We also both agree that the resulting consequences and subtle breakages are extremely hard to predict.

Cheers,
 - Andreas

Some simple analogy:
imagine a black box with two buttons 'start' and 'stop' and big red
warrantly label, which covers the bolts and
where is written: warranty will void if you attempt to disassemble the box.
You pushing start, and it is starting making noise, you can hear it
does something inside,
but you can't see, because its boxed.
You pushing stop - and it stopping, but you notice that it doesn't
stops immediately and it takes a
few seconds to complete stop.

So, all control you having from outside is a start and stop buttons.
Nothing else.
Unless.. take your wrench (or reflection hammer) and break inside to
see what it does.
But once you do that, the warranty will be void!

That's the point. If you think of Process as an object, not as a
complex conglomerate of different pieces (contexts/sheduler), then its
obvious, that you have to honor the encapsulation rules and do not
break inside unless it is necessary.

Making unwinding be handled outside of process, breaking
encapsulation, because you leaking object references from one process
to another. And we know very well, what this means - this immediately
leads to various security issues, and/or inability to contaminate the
growth of garbage in image, which can't be collected, because some
other process got a reference to an object which were not for its
eyes.

I could continue with other examples, but i think the above made my
point: an outside process usually having minimal knowledge of what is
going on inside another process. And because of that, it is false to
assume that running the code in scope of different process won't have
unwanted and unpredictable side effects. So, a best way to avoid these
issues is to keep thinking of it as a black-box, with which you
communicating using standard protocol, and at the same way be free
from being responsible for handling exceptional situations - all
exceptions should be handled inside a process where they appear.

On 7 March 2010 22:37, Eliot Miranda<[hidden email]>  wrote:

On Wed, Mar 3, 2010 at 9:29 PM, Andreas Raab<[hidden email]>  wrote:

On 3/3/2010 6:32 PM, Eliot Miranda wrote:

No, it is /broken/.  What exception handlers are in effect when an
unwind is run from another process?  Its not just tat process identity
isn't being preserved (which I would argue is important but not
essential) its that unwinds are being run in an entirely arbitrary
context which has nothing to do with their running normally.  e.g.

givethAndTakethAway: aCollection
     [[aCollection do: [:k| mydict at: k put: self thing].
       self doStuff] ensure: [aCollection do: [:k| mydict removeKey: k]]
        on: KeyNotFoundError
        do: [:ex| ex proceedWith: nil]

will remove whatever keys in mydict are in aCollection when run
normally, but when run if terminated from another process will break if
any keys in aCollection have already been removed from mydict.

Yes, but my claim was that this a feature, not a bug, for a value of
"feature" which means that the terminator (<- I really wanted to use that
:-) is capable of controlling the exception context of the ensure block.

Running unwinds in another process is /broken/, period.

The reason why I'm not convinced of that, period or not, repetitions or
not, is because we've seen no evidence of it ever being a problem. If it's
not a problem, neither in our production environments that have triggered
all the problems and the fixes for process, semaphore, mutex, termination
etc. handling over the last years, nor in messages or bug reports that I've
seen on this or other lists, I'm just not buying the strong claim you are
making.

Where is the *evidence* to support your claim? Not the theoretical
discussion of what could possibly go wrong, but rather the actual,
day-to-day things that people really get bitten by.

Evidence will be hard to come by.  In Squeak people will have coded around
the limitation.  I don't have access to the server applications in VW which
might or might not be affected by this.  But I can say that VW implements
unwinds on termination both preserving process identity and exception
context, and that if it wasn't considered important that effort would not
have been made in the first place.

You could do this better by saying e.g.

      p terminateOn: Error do: [:ex| ex return "skip it"]

and have terminateOn:do: et al install the exception handler at the
bottom of stack.

Well, not really. In a realistic situation you'd probably want to handle
multiple conditions, timeouts etc. and stitching the stack together for that
is likely way beyond your average Squeaker :-)

[BTW, I don't get why you call this "better"; it would seem that if you
call this "better" because it's using in-process termination that your
reasoning is cyclical]

Better means it's a form which would work with either approach.

   There are obviously issues if the unwind actions refer to process
   local state; but then again perhaps providing said context for the
   duration of the unwind is the right thing, along the lines of:

   Process>>terminate

     self isActiveProcess ifFalse:[
       Processor activeProcess evaluate: [self unwind] onBehalfOf: self.
     ].

   This for example would guard errors during unwind as initially
   intended while ensuring that process relative state is retrieved
   correctly.


Yes, but what if I do something like

         [processMemo add: Processor activeProcess.
          self doStuff]
                ensure: [processMemo remove: Processor activeProcess]
?

It would work fine, why? Did you not notice that I used
#evaluate:onBehalfOf: in the above for precisely this reason?

 OK.  But evaluate:onBehalfOf: only fixes one of the two issues.  It fixes
process identity but it doesn't handle exception context.  [BTW, I wrote
evaluate:onBehalfOf: only to get process identity right when simulating
execution in the debugger; I didn't envisage it being used here, but I agree
it would be an improvement].

It may have facets but that doesn't mean that running unwinds in other
than the current process isn't badly broken, and isn't relatively easy
to fix ;)

I don't buy this, and claiming it repeatedly doesn't make it any more real
unless you put up or shut up. I think you're vastly underestimating the
resulting fallout (does that sound like something I've said before? ;-)

For example, here is an interesting one (and please use VW for reference
here if you'd like; I'd actually be interested in finding a *small* download
myself to try some of this stuff): One of the issues in Levente's example of
terminating a process halfways through an unwind block is that if the unwind
block is ill-behaved you can really be in for trouble. Let's say you have
this:

[
 "just to have something to interrupt"
 [(Delay forMilliseconds: 100) wait. true] whileTrue.
] ensure:[
 "this is the actual killer"
 [(Delay forMilliseconds: 100) wait. true] whileTrue.
].

The issue of well-behavedness is orthogonal to in-process or out-of-process
termination.  We can know whether an unwind block has been started or not
and on repeating an attempt to terminate, e.g. after a timeout,  one can
choose whether to restart an in-progress unwind or skip it and start with
the next one, or skip them altogether and abort unwinding.  But these
choices are not tied to in-process or out-of-process termination except
perhaps in implementation details.

If you wish to support completing a well-behaved ensure block that has
been started, it seems tricky to see how this process would actually be
terminated, and how you'd find out that it didn't. In Squeak it's simple -
the fact that the process sending #terminate is still in this code tells you
it's not finished so one can (for example) associate a timeout with
#terminate and catch that. Or, one can just interrupt that process and close
the debugger. Both are quite reasonable options to deal with the invariant.

In VW an unwind block is always executed from a marked sender method so one
can find out whether an unwind is running by looking at its sender:
BlockClosure methods for unwinding
valueAsUnwindBlockFrom: aContextOrNil
"Unwind blocks are evaluated using this wrapper.
This method is marked as special.  When the
system searches for unwind blocks, it skips over
all contexts between this context and the context
passed in as an argument.  If the argument is
nil, it skips over the sender of this context.
The purpose of this is that errors inside an
unwind block should not evaluate that unwind
block, and they should not circumvent the
running of other unwind blocks lower on the
stack."
<exception: #unwindInAction>
| shouldTerminate |
"long comment elided"
shouldTerminate == nil ifTrue: [shouldTerminate := false].
self value.
shouldTerminate
ifTrue:
[Processor activeProcess terminate].
MarkedMethod methods for testing
isMarkedForUnwindInAction
"Answer true if method is marked for unwinding in action."
^markType == #unwindInAction
and exception handling skips over these when looking for handlers.
Context methods for handler search
skipOverUnwindingBlocks
"If the method is marked for unwindInAction, check the first argument
of the method. If it is nil, we just skip to the next context in the chain
which will be one of #ifCurtailed: or #ensure:. If it is not nil, we skip to
the context's (represented by the argument to the method) sender."
^self method isMarkedForUnwindInAction
ifTrue: [(self localAt: 1) == nil
ifTrue: [self sender sender]
ifFalse: [(self localAt: 1) sender]]
ifFalse: [self]

and termination by another process (which causes it to execute
auto-termination) skips over them also to avoid restarting unwinds already
in progress.  [Note there's a per-process interruptProtect semaphore used to
coordinate termination].

Under the assumption that you'd like to support the well-behaved case,
what do you do when using in-process termination in a non-well-behaved case?
How do you find out that a termination request didn't finish, and what do
you do in practice to kill it?

I would expect to be able to write something like
       [process terminate]
             valueIfLongerThanMilliseconds: 1000
             evaluate: [process terminateWithExtremePrejudice]
i.e. nuke it if it doesn't terminate within a specific time, or
      [[:exit|
          [process terminateAbortingCurrentUnwind]
             valueIfLongerThanMilliseconds: 100
             evaluate: exit] valueWithExit.
        process isTerminated] whileFalse
i.e. abort any unwind if it doesn't execute within a specific time but at
least attempt to run all unwinds.

Also, what are the priority implications of the different schemes? One of
the more interesting issues here is that in cases like on our servers I
could see how one of the lower priority processes times out due to
starvation, is then terminated but this may never complete if it's run from
the priority it's at instead of the terminator process. Does for example VW
bump the priority of the terminated process? How does that interact when the
terminated process isn't well-behaved and runs wild?

Priority inversion is again an orthogonal issue.  It would be great to have
priority boosting semaphores.  But in any case one could write
     process
            makePriorityAtLeast: Processor activePriority;
            terminate

I think there are numerous interesting aspects that we have somewhat
reasonable answers in Squeak today that I don't even know what the
implications would be if you'd go for in-process termination.

Do you agree that the right approach is one of specification?  What is the
correct behaviour and how close to implementing that can one reasonably get?
 I think from a specification POV one wants both process identity and
exception context to be respected when running unwinds during termination
and the more straight-forward implementation is to arrange that a process
terminates itself when requested to terminate by some other process.
Orthogonal issues as to what happens to unwinds in progress when repeated
termination requests are made (from within the process or without) and what
priority termination runs at should be specified and then implemented.  It's
reasonable for the implementation to fall short of an ideal if documented.
 It's at least regrettable to have the system specified by an
implementation.
As far as the implications of in-process termination I don't know of a
better way, beyond being able to inspect all current Squeak applications, of
voicing a specification and getting people who have written applications
affected by this functionality to comment on the specification.
I think I get why the Squeak implementation is a it is (there's a /lot/ of
implementation work involved in adding unwinds to the language and it's
unreasonable to expect that all of this work will get done at the first
attempt).  But the fact that the implementation is not ideal shouldn't
discourage us from improving it over time.

Cheers,
 - Andreas

best
Eliot