This page will contain the activity log of the pyFF+ experiments and endeavours.

Memory profiling

This is the bare import and code usage of using heapy to print heap information while running python code.

...

Another way of profiling pyFF's memory usage is just following RES in top or htop for a long-running pyFF/gunicorn process, that has a 60s refresh interval. I normally use this pipeline

- when update: 
 - load: 
   - edugain.xml 
- when request: 
 - select: 
 - pipe: 
   - when accept application/samlmetadata+xml application/xml: 
     - first 
     - finalize: 
         cacheDuration: PT12H 
         validUntil: P10D 
     - sign: 
         key: cert/sign.key 
         cert: cert/sign.crt 
     - emit application/samlmetadata+xml 
     - break 
   - when accept application/json: 
     - discojson 
     - emit application/json 
     - break


to feed the edugain feed that has been dowloaded using

...

from lxml import etree, objectify
import pickle
# Create pickled datafile
source = open("edugain.xml", "r", encoding="utf-8")
sink = open("edugain.pkl", "w")

t = objectify.parse(source)
p = pickle.dumps(t).decode('latin1')
sink.write(p)

# Read pickled object back in pyFF
def parse_xml
	return pickle.loads(io.encode('latin1'))

In metadata parser:
t = parse_xml(content) #Instead of parse_xml(unicode_stream(content))

Using un/pickling, pyFF's gunicorn starts out using ~800Mb of RES that slowly extends to a steady 1.2-1.5G.

xml.sax etree.ElementTree parser

...

import xml.sax
class XML(xml.sax.handler.ContentHandler):
  def __init__(self):
    self.current = etree.Element("root")
    self.nsmap = { 'xml': 'http://www.w3.org/XML/1998/namespace'}
    self.buffer = ''

  def startElement(self, name, attrs):
    attributes = {}
    for key, value in attrs.items():
        key = key.split(':')
        if len(key) == 2 and:
            if key[0] == 'xmlns':
                self.nsmap[key[-1]] = value
            else:
                attributes[f"{{{ self.nsmap.get(key[0], key[0]) }}}{ key[-1] }"] = value
        elif value:
            attributes[key[-1]] = value

    name = name.split(':')
    if len(name) == 2:
        name = f"{{{ self.nsmap.get(name[0], name[0]) }}}{ name[-1] }"
    else:
        name = name[-1]
    self.current = etree.SubElement(self.current, name, attributes, nsmap=self.nsmap)

  def endElement(self, name):
    self.current.text = self.buffer
    self.current.tail = "\n"
    self.current = self.current.getparent()
    self.buffer = ''

  def characters(self, data):
    d = data.strip()
    if d:
        self.current.textbuffer += d


def parse_xml(io, base_url=None):
    parser = xml.sax.make_parser()
    handler = XML()
    parser.setContentHandler(handler)
    parser.parse(io)
    return etree.ElementTree(handler.current[0])

Using xml.sax parser pyFF's gunicorn starts out using ~800Mb of RES that slowly extends to a steady 1.2-1.5G.

Run pyFF in a uwsgi server

#!/bin/sh

bin/uwsgi \
    --http 127.0.0.1:8080 \
    --module pyff.wsgi \
    --callable app \
    --enable-threads \
    --env PYFF_PIPELINE=edugain.yaml \
    --env PYFF_WORKER_POOL_SIZE=10 \
    --env PYFF_UPDATE_FREQUENCY=60 \
    --env PYFF_LOGGING=pyFFplus/examples/debug.ini

Long-run test reveals comparable memory usage as gunicorn, but there seem to be more knobs to play with.

One of the things we can do against boundless growth of uwsgi is the use of --reload-on-rss <limit>, this kills any worker that exceeds the RSS limit, but results in an empty metadata reply, which is unwanted behaviour. If however, we also supply --lazy, the app is loaded in the worker(s) and the (re)start of each worker then also triggers the reload of metadata. This could be a compromise if the VM is less cpu bound than memory?

Empty Metadata set while refreshing

It turns out pyFF returns an empty metadata set while refreshing, which is unwanted behaviour. The following code, inserted just before the final return in .api#process_handler inspects the validity of the Resource metadata. Having a loadbalancer inspect pyFF and temporarily evicting the server from pool if it receives a 500 could create a stable service.

			def process_handler():
			...

            # Only return request if md is valid?
            valid = True
            log.debug(f"Resource walk")
            for child in request.registry.md.rm.walk():
                log.debug(f"Resource {child.url}")
                valid = valid and child.is_valid()

            if len(request.registry.md.rm) == 0 or not valid:
                log.debug(f"Resource not valid")
				# 500: The server has either erred or is incapable of performing the requested operation.
                raise exc.exception_response(500)
            else:
                log.debug(f"Resource valid")

            return response

Performance-test branch

Incorporated the "store.py" changes in this branch https://github.com/IdentityPython/pyFF/compare/preformance-tests to see how that would change the memory consumption of pyFF, but it didn't change much. It ends up using ~1.8G of RES after several hours of continuously (60s) refreshing the edugain metadata feed.

The changes try to store entities as their serialized (tostring) version of the metadata, and re-parse it on demand. The idea being that we don't need to keep track of the whole parsed tree, but just the serialized entities.

Parked

https://tech.buzzfeed.com/finding-and-fixing-memory-leaks-in-python-413ce4266e7d

Size limitations

We plan to create a controlled mock metadata set containing multitudes of edugain metadata (e.g. 5k, 10k, 20k and 100k entities) to see how pyFF would cope with that amount of entities and metadata.