Skip to content

Commit

Permalink
Merge pull request #473 from kytos-ng/blueprint/pacing
Browse files Browse the repository at this point in the history
Add Pacing Blueprint
  • Loading branch information
viniarck authored Jun 3, 2024
2 parents f807667 + 76c3a50 commit 409540b
Showing 1 changed file with 213 additions and 0 deletions.
213 changes: 213 additions & 0 deletions docs/blueprints/EP038.rst
Original file line number Diff line number Diff line change
@@ -0,0 +1,213 @@
:EP: 38
:Title: Action Pacing
:Authors:
- David Ramirez <[email protected]>;
:Issued Date: 2024-05-16
:Status: Pending
:Type: Standards Track

*********************
EP038 - Action Pacing
*********************

########
Abstract
########

This document is intended on a guide for both NApp Developers
and Network Operators for how to implement and use Pacing within Kytos.
Pacing is intended as a mechanism to control
how often certain actions will be executed.
This differs from rate limiting, as the action is not cancelled,
and instead the caller is temporarily paused until the action
is allowed to be executed again.
For an action to be paced, it needs to be instrumented
by developers to support pacing.

##########
Motivation
##########

The purpose of adding in pacing to Kytos is
help maintain system stability by preventing
excessive throughput.
It could occur that sending too many flow mods
or too many DB requests may result in
connection failures, and therefore system
instability.
To prevent such scenarios,
pacing of actions which are sensitive to high
throuhput should be implemented.

#############
Specification
#############

This section describes how to set the pace of actions,
intended for network operators,
and how to add pacing to an action,
intended for developers.

Setting A Pace
==============

For network operators,
setting the pace of an action can be done through
a NApp's ``action_paces`` setting.
The ``action_paces`` setting is a dictionary,
with keys representing the action to be paced,
and values being a dict with the rate to pace them to, ``pace``,
and which ``strategy`` to use when pacing them.
Here's an example:

.. code-block:: python
action_paces = {
"send_flow_mod": {
"pace": "50/2seconds",
"strategy": "fixed_window",
},
}
The ``pace`` is a string formatted as
``{max_concurrent}/{time_period}{time_unit}``,
where ``max_concurrent`` is the max amount of times
the action is allowed to be executed
within the ``time_period`` using the given ``time_unit``.
Supported time units include: ``second``, ``minute``, ``hour``,
and ``day``.

How ``time_period`` is interpreted is controlled by the pacing ``strategy``.
Available strategies include:

- ``fixed_window`` - Actions are paced in a fixed window,
where actions are paced in fixed blocks of time
that reset after the amount of time in the refresh period has passed.
- ``elastic_window`` - Actions are paced in an elastic window,
which acts like a fixed window, except if the pace is exceeded
then the window refresh period is doubled.


Pacing an Action
================

In order to add pacing functionality to an action,
the pacing controller must be integrate into the NApp.
To do so, add in a ``PacerNappWrapper`` to wrap
around the pacer controller provided by the Kytos Controller.
Then with the wrapped pacer, inject the configuration from the
NApp's settings. Here's an example of this:

.. code-block:: python
from kytos.core.pacing import PacerNappWrapper
from napps.kytos.my_napp import settings
class Main(KytosNApp):
...
def setup(self):
...
self.pacer = PacerNappWrapper(
"mynapp",
self.controller.pacer
)
self.pacer.inject_config(
settings.action_paces
)
...
With the pacer added to the NApp at startup,
it can then be used at runtime.
To use the pacer, use the ``hit`` or ``ahit`` methods.
The ``hit`` and ``ahit`` methods take in the name of an action,
and a set of optional keys, which will then
pause execution of the current thread or task until
the action is allowed to be executed.

If we where to want to pace an action basedon a the given switch,
we would use the switch's ID, as one of the keys.
Here's an example of such an implementation:

.. code-block:: python
class Main(KytosNApp):
...
def send_switch_event(self, switch: Switch):
"""Send an event about the switch."""
self.pacer.hit("mynapp.switch_event", switch.dpid)
self.controller.buffers.app.put(
...
)
async def asend_switch_event(self, switch: Switch):
"""Asynchronously send an event about the switch."""
await self.pacer.ahit("mynapp.switch_event", switch.dpid)
await self.controller.buffers.app.aput(
...
)
As for where to implement pacing,
there is on key detail to keep in mind.
Pacing is intended for when an action is
initiated on a resource, therefore
pacing should occur at where that resource is managed.
If we wanted to pace flow mods,
based on what resources we are considering the flow
mods relative to would determine where
pacing would be implemented for it.
For example:

- Flow mod pacing could be implemented per switch.
In that case, the pacing should be implemented
in ``flow_manager``, as it manages the flows
on switches.
- Flow mod pacing could be implemented per EVC.
In that case we would implement pacing in
``mef_eline`` as it manages the EVCs.
- Pacing could be implemented per connection.
In that case we would implement pacing in
the kytos controller, as it manages all connections
to switches.


######################
Implementation Details
######################

This section is to provide developers with
a few details about the implementation,
intended for aiding in further development
and maintenance of the feature.
The way pacing is to be implemented is using the ``limits`` package.
``limits`` has a few additional features, which aren't used,
such as:

- A ``moving_window`` strategy. This isn't exposed to operators
because the behaviour as described in the ``limits``
documentation seems inconsistent with its implementation.
- External stores for limit data. The current implementation
only uses the memory storage, as of this point we don't
have enough unique actions to warrant using an external data store.


##############
Rejected Ideas
##############

This pacing specification has been through
several iterations.
Originally it was intended that pacing would
be done on the event bus.
However, the event bus approach required
a user defined config for how to extract
keys from events, which severely bloated
the implementation, and required developer
knowledge to use.
It was ultimately too unwieldy to use
correctly, and too difficult to maintain.

0 comments on commit 409540b

Please sign in to comment.