First, Heroku doesn't actually support websockets, so you must use something like socket.io which can fallback to various long polling mechanisms.

Step 1, disable websocket support in socket.io

Without this, socket.io tries to connect first using websockets and it takes a while to timeout before switching to long polling.

// remove websocket for heroku
var options = {transports:["flashsocket", "htmlfile", "xhr-polling", "jsonp-polling"]};
var socket = io.connect('http://.../", options);

Step 2, configure tornadio to use xheaders

If you don't tell tornadio to use xheaders it will think heroku is trying to hijack sessions and nothing will work. You will get 401 unauthorized messages back from tornado and the error from this statement in your logs:

# If IP address don't match - refuse connection
if handler.request.remote_ip != self.remote_ip:
    logging.error('Attempted to attach to session %s (%s) from different IP (%s)'   % (
                  self.session_id,
                  self.remote_ip,
                  handler.request.remote_ip
                  ))

Enabling xheaders is a good idea when deploying to heroku in general and is not tornadio specific.

Add the xheaders option to the main SocketServer initialization, and everything is happy.

SocketServer(application,xheaders=True)

Whatever you do, make sure you are using versioned python packages, even for simple tasks. And use pip+virtualenv.

So you want to program in python..

It seems like only yesterday, and not 7 years ago, that I decided to learn python. I may not be the best python programmer, but I have made probably every mistake you can, so here are a bunch of things not to do, and a few things you should be doing.

Don't: write python 'scripts'

Don't write programs like this:

temp = input("C: ")
print temp*9/5+32

The way you fix that is not by writing the following:

if __name__ == "__main__":
    temp = input("C: ")
    print temp*9/5+32

And don't write this either:

def main():
    temp = input("C: ")
    print temp*9/5+32
if __name__ == "__main__":
    main()

No matter how good your logic is, if you couple the logic with your input and output you are painting yourself into a corner. I've seen people write scripts like this, and then have other scripts call them using os.system. In a loop. Then they wonder why python is so slow.

Do: Write python modules and packages

Minimally this could look something like:

def ctof(temp):
    return temp*9/5+32
def main():
    temp = input("C: ")
    print ctof(temp)
if __name__ == "__main__":
    main()

Even better would be to have main parse sys.argv rather than working interactively. For simple interactive tools it is hard to beat the cmd module

Now you have a (albeit poorly named) python module that can properly be imported from a larger program:

>>> import temp
>>> print temp.ctof(100)
212

Don't: mess with PYTHONPATH

Now that you have a module you can import, what do you do with it? For years my development/production environment consisted of the following: a lib directory containing modules and packages and a util directory containing scripts that used those modules. This worked fine for a long time, especially when I only had one machine. When I got more systems, I used the high tech method of rsync'ing the entire directory tree to /srv/python or ~/python/ and mucking with the python path. This system worked, but had a number of problems:

• If I wanted to run a program on a new system, I had to rsync the entire directory tree.
• Since there was no dependency information, the first time I wanted to share a program I wrote, I had to figure out the dependencies manually.
• I had no idea what modules were being used, and which were obsolete.
• When I started writing test code and documentation, I did not have a good place to store them. I used a single directory for all my tiny modules because one directory per module seemed like overkill at the time.
• When the version of python on the system was upgraded, bad things happened.

It's very tempting to simply throw all of your python code into a single directory tree, but that method only causes problems later on.

Do: Create python modules

For the example above, we can write a simple setup.py file:

from distutils.core import setup

setup(name="temp",
    version="1.0",
    py_modules = ["temp"], 
    entry_points = {
        'console_scripts': [
            'ctof   = temp:main',
        ]
    },
)

If you have a full package instead of a single file module, you should use packages and not py_modules. The the official documentation should be read if you are doing anything more complicated. There are fields for your name, short and long descriptions, licensing information, etc. This example was kept purposely short to make it clear that there is not much you actually have to do to get started. Even a barebones setup.py is better than no setup.py.

Don't: use 'scripts' in setup.py (Do: Use entry points)

console_scripts entry_points should be preferred over the 'scripts' that setup.py can install. The last time I tried, scripts did not get correctly installed on Windows systems, but console_scripts did. Additionally, the more code you have in scripts, the less testable code you have in your modules. When you use scripts, eventually you will get to the point where they all contain something similar to:

from mypackage.commands import frob
frob()

and at that point, you are just re-implementing what console_scripts does for you.

Do: Version your packages and depend on specific versions.

So, after years of doing-the-wrong-thing, I finally created proper packages for each of my libraries and tools. Shortly after that I started having problems again. While I had been versioning all of my packages, any package that required another package simply depended on the package name and not any specific version or it. This created problems any time I would add new features. I would install the latest version of a utility package on a server, and it would crash since I had forgotten to upgrade the library it depended on. Since I wasn't syncing the entire directory tree anymore, libraries were becoming out of date.

Don't install packages system wide. (Do: Use virtualenv and pip)

Once you get to the point where you are using versioned packages, you'll want to be able install different versions of modules under different python versions. When I was simply sticking everything under /srv/python it was next to impossible to have multiple versions of python. I could change PYTHONPATH to point somewhere else, but there was no easy way to maintain two complete different trees of modules.

It is extremely simple to get started using pip and virtual environments. You can use the -E option to create a virtual environment and install a package in one command. The -E option to pip creates a virtual environment if one doesn't already exist:

justin@eee:~/tmp$ pip  -E python_env install bottle
Creating new virtualenv environment in python_env
  New python executable in python_env/bin/python
  Installing distribute...done........................
Downloading/unpacking bottle
  Downloading bottle-0.9.5.tar.gz (45Kb): 45Kb downloaded
  Running setup.py egg_info for package bottle

Installing collected packages: bottle
  Running setup.py install for bottle

Successfully installed bottle
Cleaning up...
justin@eee:~/tmp$ ./python_env/bin/python 
>>> import bottle
>>> bottle.\_\_file\_\_
'/home/justin/tmp/python_env/lib/python2.7/site-packages/bottle.pyc'
>>> 

I can use that same method to install the toy module I wrote for this post as well:

justin@eee:~/tmp$ pip  -E python_env install ~/tmp/post/temp_mod/
Unpacking ./post/temp_mod
  Running setup.py egg_info for package from file:///home/justin/tmp/post/temp_mod

Installing collected packages: temp
  Running setup.py install for temp

    Installing ctof script to /home/justin/tmp/python_env/bin

Successfully installed temp
Cleaning up...

pip was also nice enough to install my console_script:

justin@eee:~/tmp$ ./python_env/bin/ctof 
C: 34
93

Too long; Did read

The barrier to entry for python is a lot lower compared to a language like java or c++. It's true that helloworld is simply:

print("Hello, World")

However, if you plan on using python for anything more complicated, you will want to learn how to take advantage of modules and packages. Python doesn't force you to do this, but not doing so can quickly turn into a maintenance nightmare.

Consider the following example function:

def logged_in_users():
    users = set()
    for line in os.popen("who"):
        users.add(line.split()[0])
    return users

This runs just fine when everything is working:

In [4]: logged_in_users()
Out[4]: set(['justin'])

But if there is a problem running the command(for the example lets change the 'who' to 'whom':

In [6]: logged_in_users()
sh: whom: not found
Out[6]: set()

What happened was os.popen ran

"sh -c whom"

While sh started fine, the actually command could not be ran. Since os.popen also does not pass the exit code back to the parent process there is no easy method to use to see if anything went wrong.

If we switch over to subprocess.Popen, everything works fine:

    for line in subprocess.Popen(["whom"], stdout=subprocess.PIPE).stdout:

This call will instead immediately raise an exception:

OSError: [Errno 2] No such file or directory

So using subprocess.Popen and not using os.popen has the following benefits:

• Is more secure against potential command injection
• Does not waste a process
• Returns better error message to the parent process

The 'dotted quad' notation for ipv4 is fairly simple, and other than possible zero padding issues, they all look the same. ipv6 addresses are a bit different. Rather than a dotted quad they are 8 hex groups, and there are a lot of rules for displaying the addresses. For working with ipv6 addresses there are two options:

• Convert them to a 16 byte string
• Normalize them

There are some very nice libraries for working with ip addreses, but the low level socket functions can be used to convert and normalize:

>>> import socket
>>> bytes=socket.inet_pton(socket.AF_INET6, "2001:4860:800f:0:0:0:0:0063")
>>> bytes
' \x01H`\x80\x0f\x00\x00\x00\x00\x00\x00\x00\x00\x00c'
>>> 'we can see that the data is the same:'
>>> binascii.hexlify(bytes)
'20014860800f00000000000000000063'
>>> print socket.inet_ntop(socket.AF_INET6, bytes)
2001:4860:800f::63

We can make a simple fuction to do that:

def normalize(ip):
    bytes=socket.inet_pton(socket.AF_INET6, ip)
    return socket.inet_ntop(socket.AF_INET6, bytes)

You can see some of the weird normalization rules in action:

>>> print normalize("2001:4860:800f:0:0:0:0:0063")
2001:4860:800f::63
>>> print normalize("::ffff:c000:280")
::ffff:192.0.2.128
>>> print normalize("ff02:0:0:0:0:0:0:1")
ff02::1

The other day I had some free time so I gave it a go.

I downloaded the mini.iso and dd'd it to a spare usb stick. The kFreeBSD ISOs support both cd and hard disk booting like the linux images. The install took about 40 minutes(including the time taken to download everything).

After that I expected to have a few problems.. but everything worked. I was able to install zfsutils and import the zfs pools. Debian/kFreeBSD doesn't currently support nfs, but it was easy enough to install samba.

I'm left with a speedy, lightweight system, with thousands of packages and full security support:

root@pip:~# df -h /
Filesystem            Size  Used Avail Use% Mounted on
/dev/ad0s1             35G  596M   32G   2% /

root@pip:~# free -m
             total       used       free     shared    buffers     cached
Mem:          2026        222       1804         17          0          0
-/+ buffers/cache:        222       1804
Swap:            0          0          0

root@pip:~# apt-cache search ""|wc -l
26258

Other than a few utilities working a little differently (the main one I noticed was netstat not taking the same flags) it feels exactly like a debian/linux system. But with ZFS.

Nice page titles

The first thing I wanted to do was fix the page titles. Blog posts should automatically have their page title set. This was a trivial change to head.mako:

-<title>${bf.config.blog.name}</title>
+<title>
+    BB.Net
+%if post and post.title:
+- ${post.title}
+%endif
+</title>

Easy blogging

The second thing I needed to do was write a script for easily adding a new post. newblog.py was the result:

justin@eee:~/projects/bbdotnet$ ./newblog.py
Title: Playing with blogofile
cats: tech,python,blogofile

This drops me into a vim session with the following contents

---
categories: tech,python,blogofile
date: 2011/04/17 20:34:08
title: Playing with blogofile
---

all I have to do when I'm done is 'git commit'

Makefile

Finally, I wrote a stupid simple Makefile, that way I can just kick off a :make inside of vim.

all: build

build:
    blogofile build

There has been a push lately to let Google host common JS libraries for you. The main reason for this is increased performance, there are two cases where this helps:

• The user has never loaded jQuery before - They get to download it from fast servers
• The user has visited another site that also hosted jQuery on google - They don't have to download it at all.

However, there are issues with this:

• This will not work on a restricted intranet
• If the copy of jQuery on google was somehow compromised, a large number of sites would be effected.
• If google is unreachable(it happens!), the site will fail to function properly

There should be a way to include a checksum like so:

<script type="text/javascript"
    src="/js/jquery-1.3.2.min.js"
    sha1="3dc9f7c2642efff4482e68c9d9df874bf98f5bcb">
</script>

(sha1 usage here is just an example, a more secure method could easily be used instead)

This would have two benefits:

• If the copy of jQuery was maliciously modified, or simply corrupted, the browser would refuse to load it.
• The browser may be able to use a cached copy of jQuery from another site with the same checksum.

This sort of fits in with one of the ideas in the A New Way to look at Networking talk by Van Jacobson.

I finally got around to figuring out how to remap Caps lock to 'z', the magic lines to add to ~/.Xmodmap are

remove Lock = Caps_Lock
keycode 66 = z

Most of the examples I found are for swapping capslock with control or escape(which are mostly obsolete now that you can use the Keyboard prefs thing in Gnome and swap keys around with a single click). Remapping caps lock to z is still too obscure to be in the nice GUI.

Now, if only two lines in a config file could fix the battery :-)

This post is about my duplicate finding program available under programs. The program is a little bare, and needs a nicer API, but the method it uses is the most efficient one that I am aware of.

There are a couple of different ways you can find duplicate files:

Compute the hash of all the files, and look for duplicates

This method works well if the files on disk are mostly static, and files are added infrequently. In this case you can compute the hashes once, and keep it around for later scans. However, if you are only running the scan once, this method is not ideal since it requires you to read the full contents of every file

Compute the hash of files with the same size

This is the method that I think fdupes still uses. It first builds a candidate list of files that are the same size, and computes the checksum of each. This method works well if most of the files that are the same size are really duplicates, but otherwise triggers too much unneeded IO.

Compare all files with the same size in parallel

This is the method that my program uses. Like fdupes, I first built up a candidate list of files with the same size. Instead of hashing the files, it simply reads each file at the same time, comparing block by block. This is just like what the cmp(1) program does, but for multiple files at the same time. The benefit of this over calculating the files hash, is that as soon as the files differ, you can stop reading.

Implementation

There are a couple of things you need to keep in mind to implement this method.

Don't open too many files.

You have to be careful not to try and open too many files at once. If the user has 5,000 files that all have the same size, the program shouldn't try and open all 5,000 at once. My program uses a simple helper class to handle opening and closing files. The default blocksize in my program would probably waste a bit of memory in this case, but that is easily changed.

Correctly handle diverging sets.

Imagine the filesystem contains 4 files of the same size, 'a', 'b','c', and 'd', where a==c, and b==d. While reading through the files, it will become clear that a!=b, a==c, and a!=d. It is important that at this step the program continues searching using (a,c) and (b,d) as possible duplicates. This is implemented using recursion, the sets (a,c) and (b,d) are fed back into the duplicate finding function.

Example run, compared to fdupes.

Here is dupes.py running against fdupes on a modestly sized directory. Notice how dupes.py only needs to read 600K(not counting metadata).

According to iofileb.d from the dtrace toolkit, dupes.py reads 10M of data (which I think includes python), and fdupes reads 517M. This alone explains the 20x speedup seen in dupes.py

justin@pip:~$ du -hs $DIR
15G   $DIR

justin@pip:~$ time python code/dupes.py $DIR
2896 total files
35 size collisions, max of length 5
bytes read 647168

real    0m1.224s
user    0m0.234s
sys     0m0.494s

justin@pip:~$ time fdupes -r $DIR
real    0m41.694s
user    0m13.612s
sys     0m7.491s

justin@pip:~$ time python code/dupes.py $DIR
2896 total files
35 size collisions, max of length 5
bytes read 647168

real    0m3.662s
user    0m0.256s
sys     0m0.568s

justin@pip:~$ time fdupes -r $DIR
real    0m55.473s
user    0m11.383s
sys     0m6.433s

The Xen documentation on live migration states:

Currently, there is no support for providing automatic remote access to filesystems stored on local disk when a domain is migrated. Administrators should choose an appropriate storage solution (i.e. SAN, NAS, etc.) to ensure that domain filesystems are also available on their destination node. GNBD is a good method for exporting a volume from one machine to another. iSCSI can do a similar job, but is more complex to set up.

This does not mean that it is impossible though. Live migration is a more efficient migration, and migration can be seen as a save on one node, and a restore on another. Normally, if you save a VM on one machine, and try to restore it on another machine, it will fail when it is unable to read its filesystems. But what would happen if you coppied the filesystem to the other node between the save and restore? If done right, it works pretty well.

The solution?

The solution is simple:

• Save running image
• Sync disks
• copy image to other node, restore

This can be somewhat sped up by syncing the disks twice:

• Sync disks
• Save running image
• Sync disks - only having to save any changes in the last few seconds
• copy image to other node, restore

Syncronizing block devices

File backed

If you are using plain files as vbds, you can sync the disks using rsync.

Raw devices

If you are using raw devices, rsync can not be used. I wrote a small utility called [[blocksync|/programs/blocksync.py]] which can syncronize 2 block devices over the network. In my testing it was easily able to max out the network on an initial sync, and max out the disk read speed on a resync.

$ blocksync.py /dev/xen/vm-root 1.2.3.4

Will sync /dev/xen/vm-root onto 1.2.3.4. The device should already exist on the destination and be the same size.

Solaris ZFS

If you are using ZFS, it should be possible to use zfs send to sync the block devices before migration. This would give an almost instantaneous sync time.

Automation

A simple script [[/programs/xen_migrate.sh]] and its helper [[/programs/xen_vbds.py]] will migrate a domain to another host. File and raw vbds are supported. ZFS send support is not yet implemented.

Example migration

#migrating a 1G / + 128M swap over the network
#physical machines are 350mhz with 64M of ram,
#total downtime is about 3 minutes

xen1:~# time ./migrate.sh test 192.168.1.2
+ '[' 2 -ne 2 ']'
+ DOMID=test
+ DSTHOST=192.168.1.2
++ xen_vbds.py test
+ FILES=/dev/xen/test-root
/dev/xen/test-swap
+ main
+ check_running
+ xm list test
Name              Id  Mem(MB)  CPU  State  Time(s)  Console
test              87       15    0  -b---      0.0    9687
+ sync_disk
+ blocksync.py /dev/xen/test-root 192.168.1.2
ssh -c blowfish 192.168.1.2 blocksync.py server /dev/xen/test-root -b 1048576
same: 942, diff: 82, 1024/1024
+ blocksync.py /dev/xen/test-swap 192.168.1.2
ssh -c blowfish 192.168.1.2 blocksync.py server /dev/xen/test-swap -b 1048576
same: 128, diff: 0, 128/128
+ save_image
+ xm save test test.dump
+ sync_disk
+ blocksync.py /dev/xen/test-root 192.168.1.2
ssh -c blowfish 192.168.1.2 blocksync.py server /dev/xen/test-root -b 1048576
same: 1019, diff: 5, 1024/1024
+ blocksync.py /dev/xen/test-swap 192.168.1.2
ssh -c blowfish 192.168.1.2 blocksync.py server /dev/xen/test-swap -b 1048576
same: 128, diff: 0, 128/128
+ copy_image
+ scp test.dump 192.168.1.2:
test.dump                                       100%   16MB   3.2MB/s   00:05
+ restore_image
+ ssh 192.168.1.2 'xm restore test.dump && rm test.dump'
(domain
    (id 89)
    [domain info stuff cut out]
)
+ rm test.dump

real    6m6.272s
user    1m29.610s
sys     0m30.930s