Re: Revisiting lazy middleware initialization

[Ben Liyanage]

I'm not currently suffering from this.  In the past with .NET I have
definitely suffered from some long memory loads of DLLs.

The stuff I'm doing now is a lot of back end batch processing where how
fast the initial response is is pretty negligible compared to a process
that lasts for 10+ minutes.

Just weighing in on the general design principle of long initial lazy loads
in frameworks.

-Ben

On Sun, Mar 27, 2016 at 3:42 PM, Fred Stluka  wrote:

> Ben,
>
> If lazy loading is causing you problems, here's good info on how to
> force Django to load everything up front, by calling select_related()
> and prefetch_related() in cases where you need to.  And also how to
> make that the default via use_for_related_fields and custom managers:
> - https://docs.djangoproject.com/en/1.9/topics/db/optimization/
>
> --Fred
> --
> Fred Stluka -- mailto:f...@bristle.com  --
> http://bristle.com/~fred/
> Bristle Software, Inc -- http://bristle.com -- Glad to be of service!
> Open Source: Without walls and fences, we need no Windows or Gates.
> --
> On 3/27/16 4:15 PM, bliyan...@rentlytics.com wrote:
>
> I'm not too familiar with the code you're referencing, but I'm personally
> really annoyed by lazy loading.  It has a tendency to make selenium tests
> timeout inconsistently in CI, as well as give the impression to my bosses
> that the app is slow rather than just the first load which is usually what
> they see on new features.
>
> -Ben
>
> On Thursday, March 24, 2016 at 9:39:50 AM UTC-7, David Evans wrote:
>>
>> Hi all,
>>
>> Currently, middleware is initialized lazily on serving the first request,
>> rather than on application start. There may well have been good reasons for
>> this historically, but I don't think they apply any longer. Lazy
>> initialization is unhelpful if a middleware class throws an error (e.g to
>> report a misconfiguration) because the application will appear to start
>> successfully and only later report the error when a request is made.
>>
>> I'd like to propose initializing middleware when `get_wsgi_application`
>> is called. This solves the problem described above and, as far as I can
>> tell, raises no backwards compatibility issues.
>>
>> More details on all this below.
>>
>>
>> ### 1. Specific example of the problem
>>
>> I recently wrote an adapter for the WhiteNoise static file server so it
>> could function as Django middleware as well as WSGI middleware (
>> https://github.com/evansd/whitenoise). WhiteNoise may be unusual in
>> doing a non-trivial amount of work on initialization, but it doesn't seem
>> unreasonable. When used as WSGI middleware any errors are triggered
>> immediately on start up, but not so when used as Django middleware. This
>> makes for a worse developer experience and an increased chance of
>> deployment errors.
>>
>>
>> ### 2. Reasons previously given for lazy initialization
>>
>> There was some brief discussion in this ticket 4 years ago:
>> https://code.djangoproject.com/ticket/18577
>>
>> The reason given there is that "resolving on first request makes most
>> sense, especially for the case where you might not be serving requests at
>> all". Presumably this refers to running non-http-related management
>> commands. But in those cases we never instantiate a WSGI application anyway
>> (wsgi.py is just never imported) so this is no reason not to initialize
>> eagerly when constructing the WSGI application. (Of course, things may have
>> been different 4 years ago.)
>>
>> Another reason is given in the comments in django.core.handles.wsgi:
>>
>> https://github.com/django/django/blob/3c1b572f1815c295878795b183b1957d0df2ca39/django/core/handlers/wsgi.py#L154
>>
>> This says "Set up middleware if needed. We couldn't do this earlier,
>> because settings weren't available". However `get_wsgi_application` (the
>> only public WSGI API) now calls `django.setup()` before constructing the
>> handler so settings are in fact available.
>>
>>
>> ### 3. Proposed solution
>>
>> My proposal is simply to amend `get_wsgi_application` as follows:
>>
>> def get_wsgi_application():
>> django.setup(set_prefix=False)
>> handler = WSGIHandler()
>> handler.load_middleware()
>> return handler
>>
>> It's possible that this logic could be moved into the handler's __init__
>> method. This caused no problems with existing application when I tried it,
>> however it did cause problems with the test suite which seems to rely on
>> the old behaviour in places. The above proposal passes all existing tests
>> as is.
>>
>>
>> ### 4. Backwards compatibility issues
>>
>> Middleware constructors have no means of accessing the request object or
>> anything that depends on it. They are called right at the start of the
>> handler's `__call__` method before the `request_started` signal is sent and
>> before the `script_prefix` thread-local is set. Therefore it cannot matter,
>> 

Re: Revisiting lazy middleware initialization

[Shai Berger]

Fred,

The thread (and Ben) is talking about "lazy load" in the sense of "wait until 
the first request comes in, only then load things". This used to be the way all 
of Django behaved until 1.7 and the app refactor. These days, it only applies 
to some parts -- middleware is the topic of this thread, and templates is 
another example that comes to mind (at least when cached -- otherwise they may 
be loaded on every request).

HTH,
Shai.

On Monday 28 March 2016 01:42:23 Fred Stluka wrote:
> Ben,
> 
> If lazy loading is causing you problems, here's good info on how to
> force Django to load everything up front, by calling select_related()
> and prefetch_related() in cases where you need to.  And also how to
> make that the default via use_for_related_fields and custom managers:
> - https://docs.djangoproject.com/en/1.9/topics/db/optimization/
> 
> --Fred
> 
> Fred Stluka -- mailto:f...@bristle.com -- http://bristle.com/~fred/
> Bristle Software, Inc -- http://bristle.com -- Glad to be of service!
> Open Source: Without walls and fences, we need no Windows or Gates.
> 
> 
> On 3/27/16 4:15 PM, bliyan...@rentlytics.com wrote:
> > I'm not too familiar with the code you're referencing, but I'm
> > personally really annoyed by lazy loading. It has a tendency to make
> > selenium tests timeout inconsistently in CI, as well as give the
> > impression to my bosses that the app is slow rather than just the
> > first load which is usually what they see on new features.
> > 
> > -Ben
> > 
> > On Thursday, March 24, 2016 at 9:39:50 AM UTC-7, David Evans wrote:
> > Hi all,
> > 
> > Currently, middleware is initialized lazily on serving the first
> > request, rather than on application start. There may well have
> > been good reasons for this historically, but I don't think they
> > apply any longer. Lazy initialization is unhelpful if a middleware
> > class throws an error (e.g to report a misconfiguration) because
> > the application will appear to start successfully and only later
> > report the error when a request is made.
> > 
> > I'd like to propose initializing middleware when
> > `get_wsgi_application` is called. This solves the problem
> > described above and, as far as I can tell, raises no backwards
> > compatibility issues.
> > 
> > More details on all this below.
> > 
> > 
> > ### 1. Specific example of the problem
> > 
> > I recently wrote an adapter for the WhiteNoise static file server
> > so it could function as Django middleware as well as WSGI
> > middleware (https://github.com/evansd/whitenoise
> > ). WhiteNoise may be unusual
> > in doing a non-trivial amount of work on initialization, but it
> > doesn't seem unreasonable. When used as WSGI middleware any errors
> > are triggered immediately on start up, but not so when used as
> > Django middleware. This makes for a worse developer experience and
> > an increased chance of deployment errors.
> > 
> > 
> > ### 2. Reasons previously given for lazy initialization
> > 
> > There was some brief discussion in this ticket 4 years ago:
> > https://code.djangoproject.com/ticket/18577
> > 
> > 
> > The reason given there is that "resolving on first request makes
> > most sense, especially for the case where you might not be serving
> > requests at all". Presumably this refers to running
> > non-http-related management commands. But in those cases we never
> > instantiate a WSGI application anyway (wsgi.py is just never
> > imported) so this is no reason not to initialize eagerly when
> > constructing the WSGI application. (Of course, things may have
> > been different 4 years ago.)
> > 
> > Another reason is given in the comments in django.core.handles.wsgi:
> > https://github.com/django/django/blob/3c1b572f1815c295878795b183b1957
> > d0df2ca39/django/core/handlers/wsgi.py#L154
> >  > 57d0df2ca39/django/core/handlers/wsgi.py#L154>
> > 
> > This says "Set up middleware if needed. We couldn't do this
> > earlier, because settings weren't available". However
> > `get_wsgi_application` (the only public WSGI API) now calls
> > `django.setup()` before constructing the handler so settings are
> > in fact available.
> > 
> > 
> > ### 3. Proposed solution
> > 
> > My proposal is simply to amend `get_wsgi_application` as follows:
> >   def get_wsgi_application():
> >   django.setup(set_prefix=False)
> >   handler = WSGIHandler()
> >   handler.load_middleware()
> >   return handler
> > 

Re: Revisiting lazy middleware initialization

[Fred Stluka]

Ben,

If lazy loading is causing you problems, here's good info on how to
force Django to load everything up front, by calling select_related()
and prefetch_related() in cases where you need to.  And also how to
make that the default via use_for_related_fields and custom managers:
- https://docs.djangoproject.com/en/1.9/topics/db/optimization/

--Fred

Fred Stluka -- mailto:f...@bristle.com -- http://bristle.com/~fred/
Bristle Software, Inc -- http://bristle.com -- Glad to be of service!
Open Source: Without walls and fences, we need no Windows or Gates.

On 3/27/16 4:15 PM, bliyan...@rentlytics.com wrote:
I'm not too familiar with the code you're referencing, but I'm 
personally really annoyed by lazy loading. It has a tendency to make 
selenium tests timeout inconsistently in CI, as well as give the 
impression to my bosses that the app is slow rather than just the 
first load which is usually what they see on new features.


-Ben

On Thursday, March 24, 2016 at 9:39:50 AM UTC-7, David Evans wrote:

Hi all,

Currently, middleware is initialized lazily on serving the first
request, rather than on application start. There may well have
been good reasons for this historically, but I don't think they
apply any longer. Lazy initialization is unhelpful if a middleware
class throws an error (e.g to report a misconfiguration) because
the application will appear to start successfully and only later
report the error when a request is made.

I'd like to propose initializing middleware when
`get_wsgi_application` is called. This solves the problem
described above and, as far as I can tell, raises no backwards
compatibility issues.

More details on all this below.


### 1. Specific example of the problem

I recently wrote an adapter for the WhiteNoise static file server
so it could function as Django middleware as well as WSGI
middleware (https://github.com/evansd/whitenoise
). WhiteNoise may be unusual
in doing a non-trivial amount of work on initialization, but it
doesn't seem unreasonable. When used as WSGI middleware any errors
are triggered immediately on start up, but not so when used as
Django middleware. This makes for a worse developer experience and
an increased chance of deployment errors.


### 2. Reasons previously given for lazy initialization

There was some brief discussion in this ticket 4 years ago:
https://code.djangoproject.com/ticket/18577


The reason given there is that "resolving on first request makes
most sense, especially for the case where you might not be serving
requests at all". Presumably this refers to running
non-http-related management commands. But in those cases we never
instantiate a WSGI application anyway (wsgi.py is just never
imported) so this is no reason not to initialize eagerly when
constructing the WSGI application. (Of course, things may have
been different 4 years ago.)

Another reason is given in the comments in django.core.handles.wsgi:

https://github.com/django/django/blob/3c1b572f1815c295878795b183b1957d0df2ca39/django/core/handlers/wsgi.py#L154



This says "Set up middleware if needed. We couldn't do this
earlier, because settings weren't available". However
`get_wsgi_application` (the only public WSGI API) now calls
`django.setup()` before constructing the handler so settings are
in fact available.


### 3. Proposed solution

My proposal is simply to amend `get_wsgi_application` as follows:

  def get_wsgi_application():
  django.setup(set_prefix=False)
  handler = WSGIHandler()
  handler.load_middleware()
  return handler

It's possible that this logic could be moved into the handler's
__init__ method. This caused no problems with existing application
when I tried it, however it did cause problems with the test suite
which seems to rely on the old behaviour in places. The above
proposal passes all existing tests as is.


### 4. Backwards compatibility issues

Middleware constructors have no means of accessing the request
object or anything that depends on it. They are called right at
the start of the handler's `__call__` method before the
`request_started` signal is sent and before the `script_prefix`
thread-local is set. Therefore it cannot matter, from the
middleware class's perspective, whether it is instantiated before
or after the first request comes in.


I'm aware this issue probably isn't high on anyone else's priority
list, but I think it would 

Re: Revisiting lazy middleware initialization

[bliyanage]

I'm not too familiar with the code you're referencing, but I'm personally 
really annoyed by lazy loading.  It has a tendency to make selenium tests 
timeout inconsistently in CI, as well as give the impression to my bosses 
that the app is slow rather than just the first load which is usually what 
they see on new features.

-Ben

On Thursday, March 24, 2016 at 9:39:50 AM UTC-7, David Evans wrote:
>
> Hi all,
>
> Currently, middleware is initialized lazily on serving the first request, 
> rather than on application start. There may well have been good reasons for 
> this historically, but I don't think they apply any longer. Lazy 
> initialization is unhelpful if a middleware class throws an error (e.g to 
> report a misconfiguration) because the application will appear to start 
> successfully and only later report the error when a request is made.
>
> I'd like to propose initializing middleware when `get_wsgi_application` is 
> called. This solves the problem described above and, as far as I can tell, 
> raises no backwards compatibility issues.
>
> More details on all this below.
>
>
> ### 1. Specific example of the problem
>
> I recently wrote an adapter for the WhiteNoise static file server so it 
> could function as Django middleware as well as WSGI middleware (
> https://github.com/evansd/whitenoise). WhiteNoise may be unusual in doing 
> a non-trivial amount of work on initialization, but it doesn't seem 
> unreasonable. When used as WSGI middleware any errors are triggered 
> immediately on start up, but not so when used as Django middleware. This 
> makes for a worse developer experience and an increased chance of 
> deployment errors.
>
>
> ### 2. Reasons previously given for lazy initialization
>
> There was some brief discussion in this ticket 4 years ago:
> https://code.djangoproject.com/ticket/18577
>
> The reason given there is that "resolving on first request makes most 
> sense, especially for the case where you might not be serving requests at 
> all". Presumably this refers to running non-http-related management 
> commands. But in those cases we never instantiate a WSGI application anyway 
> (wsgi.py is just never imported) so this is no reason not to initialize 
> eagerly when constructing the WSGI application. (Of course, things may have 
> been different 4 years ago.)
>
> Another reason is given in the comments in django.core.handles.wsgi:
>
> https://github.com/django/django/blob/3c1b572f1815c295878795b183b1957d0df2ca39/django/core/handlers/wsgi.py#L154
>
> This says "Set up middleware if needed. We couldn't do this earlier, 
> because settings weren't available". However `get_wsgi_application` (the 
> only public WSGI API) now calls `django.setup()` before constructing the 
> handler so settings are in fact available.
>
>
> ### 3. Proposed solution
>
> My proposal is simply to amend `get_wsgi_application` as follows:
>
> def get_wsgi_application():
> django.setup(set_prefix=False)
> handler = WSGIHandler()
> handler.load_middleware()
> return handler
>
> It's possible that this logic could be moved into the handler's __init__ 
> method. This caused no problems with existing application when I tried it, 
> however it did cause problems with the test suite which seems to rely on 
> the old behaviour in places. The above proposal passes all existing tests 
> as is.
>
>
> ### 4. Backwards compatibility issues
>
> Middleware constructors have no means of accessing the request object or 
> anything that depends on it. They are called right at the start of the 
> handler's `__call__` method before the `request_started` signal is sent and 
> before the `script_prefix` thread-local is set. Therefore it cannot matter, 
> from the middleware class's perspective, whether it is instantiated before 
> or after the first request comes in.
>
>
> I'm aware this issue probably isn't high on anyone else's priority list, 
> but I think it would count as a genuine -- if small -- improvement to 
> Django.
>
> Thanks,
>
> Dave
>

-- 
You received this message because you are subscribed to the Google Groups 
"Django developers  (Contributions to Django itself)" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to django-developers+unsubscr...@googlegroups.com.
To post to this group, send email to django-developers@googlegroups.com.
Visit this group at https://groups.google.com/group/django-developers.
To view this discussion on the web visit 
https://groups.google.com/d/msgid/django-developers/f61a2b50-01e3-48ef-809f-82edc23d9da4%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.