Re: [Async-sig] Feedback, loop.load() function

[Pau Freixes]

I had an idea based on the Trios implementation. Trying to move the
load method implementation out of the loop scope but avoiding the
subclassing.

My primary concern with the instrumentation was the performance
impact. I've run some tests in one experimental branch, instrumenting
4 events for the _run_once
function. The idea was to see how the performance degrades, despite
there were not listeners/instruments
attached. My gut feelings said that only the call to emit the event
and the proper check to see if there were listeners will have an
expensive performance cost.

The numbers were pretty clear, the default Asyncio loses a 10% of
performance. Degrading from the 30K coroutines per second to 27K in my
laptop. Therefore, looks like we are facing a stopper that would need
to be addressed first.

How might Python address this situation?

Basically implementing a new object type that will provide the same
interface as a Signal. Many languages have their own implementation C#
[1], Java [2] or they have libraries that support Signals or Events.
Indeed Python has plenty of them implemented as independent modules,
the most known are Django Signals [3] or Blinker [4]

>From my understanding here the problem is the language by itself,
having the interpreter behind the scenes.

What I would propose here is a new Python object called Signal
implemented as a new type in the CPython API that will allow the
developers to implement the following snippet having the chance to
speed up the code or at least don't be penalized when there are no
observers listening from a specific event.

class Foo:
def __init__():
self._cb = []

def add_listener(self, cb):
self._cp.append(cb)

def _signal(self, msg):
for cb in self._cb:
cb(msg)

def run():
self._signal("start")
self._signal("stop")


f = Foo()
f.run()

def print_signal(msg):
print(msg)

f.add_listener(print_signal)
f.run()

This previous snippet might be implemented using a new object type
called signal, having the next snippet:

class Foo:
def __init__():
self.start = signal()
self.stop = signal()

def run():
self.start.send()
self.stop.send()


f = Foo()
f.run()

def start():
print(msg)

def stop():
print(msg)

f.start.connect(start)
f.stop.connect(stop)
f.run()

This kind of pattern is spread in many frameworks [5], and they might
get benefited for free thanks to this new type of object.


I'm wondering if there were other attempts in the past to implement
this kind of object in the CPython API. There is some resistance to
modify or add superfluous code in the code if it can be implemented as
a Python module? The simple fact to become a must for an internal
feature might enough to think on that?

Thoughts ?


[1] http://docs.oracle.com/javase/tutorial/uiswing/events/index.html
[2] https://msdn.microsoft.com/en-us/library/aa645739(v=vs.71).aspx
[3] https://docs.djangoproject.com/en/1.11/topics/signals/
[4] http://pythonhosted.org/blinker/
[5] http://flask.pocoo.org/docs/0.12/signals/

On Tue, Aug 22, 2017 at 2:36 PM, Pau Freixes  wrote:
>> One approach would be to add a generic instrumentation API. Trio has
>> something like this, that I think is expressive enough to let Pau
>> implement their busyness checking as a library:
>>
>> https://trio.readthedocs.io/en/latest/reference-core.html#instrument-api
>>
>> This has several advantages over subclassing: multiple libraries can
>> define their own instrumentation without interfering with each other,
>> you don't have to redefine the instrumentation for every loop
>> implementation, and you don't have to hook up the instrumentation when
>> setting up the loop, e.g. you could just do something like:
>>
>> import toobusy
>> with toobusy.Monitor(loop) as monitor:
>> if monitor.toobusy():
>> ...
>
> It will help also other loops to meet the same contract making them
> compatibles with already implemented instruments. Maybe the major
> concern here is the performance penalty, do you have some numbers
> about how negligible is have all of these signals available to be
> used?
> --
> --pau



-- 
--pau
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/

Re: [Async-sig] Feedback, loop.load() function

[Pau Freixes]

> One approach would be to add a generic instrumentation API. Trio has
> something like this, that I think is expressive enough to let Pau
> implement their busyness checking as a library:
>
> https://trio.readthedocs.io/en/latest/reference-core.html#instrument-api
>
> This has several advantages over subclassing: multiple libraries can
> define their own instrumentation without interfering with each other,
> you don't have to redefine the instrumentation for every loop
> implementation, and you don't have to hook up the instrumentation when
> setting up the loop, e.g. you could just do something like:
>
> import toobusy
> with toobusy.Monitor(loop) as monitor:
> if monitor.toobusy():
> ...

It will help also other loops to meet the same contract making them
compatibles with already implemented instruments. Maybe the major
concern here is the performance penalty, do you have some numbers
about how negligible is have all of these signals available to be
used?
-- 
--pau
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/

Re: [Async-sig] Feedback, loop.load() function

[Pau Freixes]

Hi,


> I personally don't think we need this in asyncio.  While the function has a
> relatively low overhead, it's still an overhead, it's a couple more syscalls
> on each loop iteration, and it's a bit of a technical debt.
>
> The bar for adding new functionality to asyncio is very high, and I don't
> see a good rationale for adding this function.  Is it for debugging
> purposes?  Or for profiling live applications?  If it's the latter, then
> there are so many other things we want to see, and some of them are
> protocol-specific.

Let's try to evolve the rationale and put some extra links.

The load of an Asyncio loop can be at some point easily inferred using
the sleeping time vs the overall time,
this information brings us to understand how to saturate is the loop
with a metric that informs you how many CPU
resources are being used, or most important how many CPU resources left.

How helpful can be this method?

In our organization, e use back-pressure at the application layer of
our REST microservices architecture. It allows us to prevent
overloading the services. Once the back pressures kicks in we can
scale horizontally our services to cope the current load. This is
already implemented for other languages and we are currently working
on how to implement it with the aiohttp(asyncio) stack. For more info
about this technique these articles [1] [2]

We are not the first ones running microservices at scale, and this
pattern has been implemented by other organizations. I would like to
mention the Google case [2]. From that link I would like to bold the
following paragraph:

"""
A better solution is to measure capacity directly in available
resources. For example, you may have a total of 500 CPU cores and 1 TB
of memory reserved for a given service in a given datacenter.
Naturally, it works much better to use those numbers directly to model
a datacenter's capacity. We often speak about the cost of a request to
refer to a normalized measure of how much CPU time it has consumed
(over different CPU architectures, with consideration of performance
differences).

In a majority of cases (although certainly not in all), we've found
that simply using CPU consumption as the signal for provisioning works
well
"""

>From my understanding, the comment is pretty aligned with the
implementation proposal for the Asyncio loop Having, as a result, a
way to measure if there are enough resources to cope the ongoing
metric.


[1] https://dzone.com/articles/applying-back-pressure-when
[2] http://engineering.voxer.com/2013/09/16/backpressure-in-nodejs/
[3] https://landing.google.com/sre/book/chapters/handling-overload.html


> If we want to add some tracing/profiling functions there should be a way to
> disable them, otherwise the performance of event loops like uvloop will
> degrade, and I'm not sure that all of its users want to pay a price for
> something they won't ever be using.  All of this just adds to the
> complexity.

The goal will be, have an implementation without impact performance
for real applications. I'm still not sure if this is reachable with
the uvloop, I would like to start working on this as soon as possible,
having the proper numbers and the possibilities to implement this in
libuv will help to get the proper answer.

If at last, there is no way to make it negligible, then I would agree
that is needed a way to switch off or switch on.
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/

Re: [Async-sig] Feedback, loop.load() function

[Pau Freixes]

> It looks like your "load average" is computing something very different than
> the traditional Unix "load average". If I'm reading right, yours is a
> measure of what percentage of the time the loop spent sleeping waiting for
> I/O, taken over the last 60 ticks of a 1 second timer (so generally slightly
> longer than 60 seconds). The traditional Unix load average is an
> exponentially weighted moving average of the length of the run queue.

The implementation proposed wants to expose the load of the loop.
Having a direct metric that comes from the loop instead of using an
external metric such as CPU, load average u others.

Yes, the load average uses a decay function based on the length of the
run queue for those processes that are using or waiting for a CPU,
this gives us extra information about how overloaded is our system. If
you compare it with the CPU load.

In the case presented, the load of the loop is something equivalent
with the load of the CPU and it does not have the ability to inform
you about how much overloaded is your loop once reached the 100%.

>
> Is one of those definitions better for your goal of detecting when to shed
> load? I don't know. But calling them the same thing is pretty confusing :-).
> The Unix version also has the nice property that it can actually go above 1;
> yours doesn't distinguish between a service whose load is at exactly 100% of
> capacity and barely keeping up, versus one that's at 200% of capacity and
> melting down. But for load shedding maybe you always want your tripwire to
> be below that anyway.

Well, I partially disagree with this. The load definition has its
equivalent in computing with other metrics that have a close range,
such as the CPU one. I've never had the intention to align the load of
the loop with the load average, I've just used the concept as an
example of the metric that might be used to check how loaded is your
system.

>
> More broadly we might ask what's the best possible metric for this purpose –
> how do we judge? A nice thing about the JavaScript library you mention is
> that scheduling delay is a real thing that directly impacts the quality of
> service – it's more of an "end to end" measure in a sense. Of course, if you
> really want an end to end measure you can do things like instrument your
> actual logic, see how fast you're replying to HTTP requests or whatever,
> which is even more valid but creates complications because some requests are
> supposed to take longer than others, etc. I don't know which design goals
> are important for real operations.

Here the key for me, something where I should have based my rationale.
How good is the way presented to measure a load of your asynchronous
system compared with the toobusy one? what can we achieve with this
metric?

I will work on that as the base of my rationale for the change
proposed. Then, once if the rationale is accepted the implementation
is peanuts :)

-- 
--pau
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/

Re: [Async-sig] Feedback, loop.load() function

[Nathaniel Smith]

It looks like your "load average" is computing something very different
than the traditional Unix "load average". If I'm reading right, yours is a
measure of what percentage of the time the loop spent sleeping waiting for
I/O, taken over the last 60 ticks of a 1 second timer (so generally
slightly longer than 60 seconds). The traditional Unix load average is an
exponentially weighted moving average of the length of the run queue.

Is one of those definitions better for your goal of detecting when to shed
load? I don't know. But calling them the same thing is pretty confusing
:-). The Unix version also has the nice property that it can actually go
above 1; yours doesn't distinguish between a service whose load is at
exactly 100% of capacity and barely keeping up, versus one that's at 200%
of capacity and melting down. But for load shedding maybe you always want
your tripwire to be below that anyway.

More broadly we might ask what's the best possible metric for this purpose
– how do we judge? A nice thing about the JavaScript library you mention is
that scheduling delay is a real thing that directly impacts quality of
service – it's more of an "end to end" measure in a sense. Of course, if
you really want an end to end measure you can do things like instrument
your actual logic, see how fast you're replying to http requests or
whatever, which is even more valid but creates complications because some
requests are supposed to take longer than others, etc. I don't know which
design goals are important for real operations.

On Aug 6, 2017 3:57 PM, "Pau Freixes"  wrote:

> Hi guys,
>
> I would appreciate any feedback about the idea of implementing a new
> load function to ask about how saturated is your reactor.
>
> I have a proof of concept [1] of how the load function might be
> implemented in the Asyncio python loop.
>
> The idea is to provide a method that can be used to ask about the load
> of the reactor in a specific time, this implementation returns the
> load taking into account the last 60 seconds but it can easily return
> the 5m and 15minutes ones u others.
>
> This method can help services built on to of Asyncio to implement back
> pressure mechanisms that take into account a metric coming from the
> loop, instead of inferring the load using other metrics provided by
> external agents such as the CPU, load average u others.
>
> Nowadays exists some alternatives for other languages that address
> this situation using the lag of a scheduler callback, produced by
> saturated reactors. The most known implementation is toobusy [2] a
> nodejs implementation.
>
> IMHO the solution provided by tobusy has a strong dependency with the
> hardware needing to tune the maximum lag allowed in terms of
> milliseconds [3]. in the POF presented the user can use an exact value
> meaning the percentage of the load, perhaps 0.9
>
> Any comment would be appreciated.
>
> [1] https://github.com/pfreixes/cpython/commit/
> 5fef3cae043abd62165ce40b181286e18f5fb19c
> [2] https://www.npmjs.com/package/toobusy
> [3] https://www.npmjs.com/package/toobusy#tunable-parameters
> --
> --pau
> ___
> Async-sig mailing list
> Async-sig@python.org
> https://mail.python.org/mailman/listinfo/async-sig
> Code of Conduct: https://www.python.org/psf/codeofconduct/
>
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/

Re: [Async-sig] Feedback, loop.load() function

[Chris Jerdonek]

On Sun, Aug 6, 2017 at 3:57 PM, Pau Freixes  wrote:
> Hi guys,
>
> I would appreciate any feedback about the idea of implementing a new
> load function to ask about how saturated is your reactor.

Hi,

Would it be possible for you to rephrase what you've done in terms of
asyncio terminology? From what I can tell, "reactor" isn't a term used
in the asyncio docs or code base. It might also improve the
readability of your asyncio patch to use asyncio terminology in the
code comments, doc strings, etc.

--Chris


>
> I have a proof of concept [1] of how the load function might be
> implemented in the Asyncio python loop.
>
> The idea is to provide a method that can be used to ask about the load
> of the reactor in a specific time, this implementation returns the
> load taking into account the last 60 seconds but it can easily return
> the 5m and 15minutes ones u others.
>
> This method can help services built on to of Asyncio to implement back
> pressure mechanisms that take into account a metric coming from the
> loop, instead of inferring the load using other metrics provided by
> external agents such as the CPU, load average u others.
>
> Nowadays exists some alternatives for other languages that address
> this situation using the lag of a scheduler callback, produced by
> saturated reactors. The most known implementation is toobusy [2] a
> nodejs implementation.
>
> IMHO the solution provided by tobusy has a strong dependency with the
> hardware needing to tune the maximum lag allowed in terms of
> milliseconds [3]. in the POF presented the user can use an exact value
> meaning the percentage of the load, perhaps 0.9
>
> Any comment would be appreciated.
>
> [1] 
> https://github.com/pfreixes/cpython/commit/5fef3cae043abd62165ce40b181286e18f5fb19c
> [2] https://www.npmjs.com/package/toobusy
> [3] https://www.npmjs.com/package/toobusy#tunable-parameters
> --
> --pau
> ___
> Async-sig mailing list
> Async-sig@python.org
> https://mail.python.org/mailman/listinfo/async-sig
> Code of Conduct: https://www.python.org/psf/codeofconduct/
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/

[Async-sig] Feedback, loop.load() function

[Pau Freixes]

Hi guys,

I would appreciate any feedback about the idea of implementing a new
load function to ask about how saturated is your reactor.

I have a proof of concept [1] of how the load function might be
implemented in the Asyncio python loop.

The idea is to provide a method that can be used to ask about the load
of the reactor in a specific time, this implementation returns the
load taking into account the last 60 seconds but it can easily return
the 5m and 15minutes ones u others.

This method can help services built on to of Asyncio to implement back
pressure mechanisms that take into account a metric coming from the
loop, instead of inferring the load using other metrics provided by
external agents such as the CPU, load average u others.

Nowadays exists some alternatives for other languages that address
this situation using the lag of a scheduler callback, produced by
saturated reactors. The most known implementation is toobusy [2] a
nodejs implementation.

IMHO the solution provided by tobusy has a strong dependency with the
hardware needing to tune the maximum lag allowed in terms of
milliseconds [3]. in the POF presented the user can use an exact value
meaning the percentage of the load, perhaps 0.9

Any comment would be appreciated.

[1] 
https://github.com/pfreixes/cpython/commit/5fef3cae043abd62165ce40b181286e18f5fb19c
[2] https://www.npmjs.com/package/toobusy
[3] https://www.npmjs.com/package/toobusy#tunable-parameters
-- 
--pau
___
Async-sig mailing list
Async-sig@python.org
https://mail.python.org/mailman/listinfo/async-sig
Code of Conduct: https://www.python.org/psf/codeofconduct/