Blog entries december 2012 [3]

Building Debian images for an OpenStack (private) cloud

2012/12/23 by David Douard

Now I have a working OpenStack cloud at Logilab, I want to provide my fellow collegues a bunch of ready-made images to create instances.

Strangely, there are no really usable ready-made UEC Debian images available out there. There have been recent efforts made to provide Debian images on Amazon Market Place, and the toolsuite used to build these is available as a collection of bash shell scripts from a github repository. There are also some images for Eucalyptus, but I have not been able to make them boot properly on my kvm-based OpenStack install.

So I have tried to build my own set of Debian images to upload in my glance shop.

Vocabulary

A bit of vocabulary may be useful for the one not very accustomed with OpenStack nor AWS jargons.

When you want to create an instance of an image, ie. boot a virtual machine in a cloud, you generally choose from a set of ready made system images, then you choose a virtual machine flavor (ie. a combination of a number of virtual CPUs, an amount of RAM, and a harddrive size used as root device). Generally, you have to choose between tiny (1 CPU, 512MB, no disk), small (1 CPU, 2G of RAM, 20G of disk), etc.

In the cloud world, an instance is not meant to be sustainable. What is sustainable is a volume that can be attached to a running instance.

If you want your instance to be sustainable, there are 2 choices:

  • you can snapshot a running instance and upload it as a new image ; so it is not really a sustainable instance, instead, it's the ability to configure an instance that is then the base for booting other instances,
  • or you can boot an instance from a volume (which is the sustainable part of a virtual machine in a cloud).

In the Amazon world, a "standard" image (the one that is instanciated when creating a new instance) is called an instance store-backed AMI images, also called an UEC image, and a volume image is called an EBS-backed AMI image (EBS stands for Elastic Block Storage). So an AMI images stored in a volume cannot be instanciated, it can be booted once and only once at a time. But it is sustainable. Different usage.

An UEC or AMI image consist in a triplet: a kernel, an init ramdisk and a root file system image. An EBS-backed image is just the raw image disk to be booted on a virtulization host (a kvm raw or qcow2 image, etc.)

Images in OpenStack

In OpenStack, when you create an instance from a given image, what happens depends on the kind of image.

In fact, in OpenStack, one can upload traditional UEC AMI images (need to upload the 3 files, the kernel, the initial ramdisk and the root filesystem as a raw image). But one can also upload bare images. These kind of images are booted directly by the virtualization host. So it is some kind of hybrid between a boot from volume (an EBS-backed boot in the Amazon world) and the traditional instanciation from an UEC image.

Instanciating an AMI image

When one creates an instance from an AMI image in an OpenStack cloud:

  • the kernel is copied to the virtualization host,
  • the initial ramdisk is copied to the virtualization host,
  • the root FS image is copied to the virtualization host,
  • then, the root FS image is :
    • duplicated (instanciated),
    • resized (the file is increased if needed) to the size of the asked instance flavor,
    • the file system is resized to the new size of the file,
    • the contained filesystem is mounted (using qemu-nbd) and the configured SSH acces key is added to /root/.ssh/authorized_keys
    • the nbd volume is then unmounted
  • a libvirt domain is created, configured to boot from the given kernel and init ramdisk, using the resized and modified image disk as root filesystem,
  • the libvirt domain is then booted.

Instantiating a BARE image

When one creates an instance from a BARE image in an OpenStack cloud:

  • the VM image file is copied on the virtualization host,
  • the VM image file is duplicated (instantiated),
  • a libvirt domain is created, configured to boot from this copied image disk as root filesystem,
  • the libvirt domain is then booted.

Differences between the 2 instantiation methods

Instantiating a BARE image:
  • Involves a much simpler process.
  • Allows to boot a non-linux system (depends on the virtualization system, especially true when using kvm vitualization).
  • Is slower to boot and consumes more resources, since the virtual machine image must be the size of the required/wanted virtual machine (but can remain minimal if using a qcow2 image format). If you use a 10G raw image, then 10G of data will be copied from the image provider to the virtualization host, and this big file will be duplicated each time you instantiate this image.
  • The root filesystem size corresponding to the flavor of the instance is not honored; the filesystem size is the one of the BARE images.
Instantiating an AMI image:
  • Honours the flavor.
  • Generally allows quicker instance creation process.
  • Less resource consumption.
  • Can only boot Linux guests.

If one wants to boot a Windows guest in OpenStack, the only solution (as far as I know) is to use a BARE image of an installed Windows system. It works (I have succeeded in doing so), but a minimal Windows 7 install is several GB, so instantiating such a BARE image is very slow, because the image needs to be uploaded on the virtualization host.

Building a Debian AMI image

So I wanted to provide a minimal Debian image in my cloud, and to provide it as an AMI image so the flavor is honoured, and so the standard cloud injection mechanisms (like setting up the ssh key to access the VM) work without having to tweak the rc.local script or use cloud-init in my guest.

Here is what I did.

1. Install a Debian system in a standard libvirt/kvm guest.

david@host:~$ virt-install  --connect qemu+tcp://virthost/system   \
                 -n openstack-squeeze-amd64 -r 512 \
                 -l http://ftp2.fr.debian.org/pub/debian/dists/stable/main/installer-amd64/ \
                 --disk pool=default,bus=virtio,type=qcow2,size=5 \
                 --network bridge=vm7,model=virtio  --nographics  \
                 --extra-args='console=tty0 console=ttyS0,115200'

This creates a new virtual machine, launch the Debian installer directly downloaded from a Debian mirror, and start the usual Debian installer in a virtual serial console (I don't like VNC very much).

I then followed the installation procedure. When asked for the partitioning and so, I chose to create only one primary partition (ie. with no swap partition; it wont be necessary here). I also chose only "Default system" and "SSH server" to be installed.

2. Configure the system

After the installation process, the VM is rebooted, I log into it (by SSH or via the console), so I can configure a bit the system.

david@host:~$ ssh root@openstack-squeeze-amd64.vm.logilab.fr
Linux openstack-squeeze-amd64 2.6.32-5-amd64 #1 SMP Sun Sep 23 10:07:46 UTC 2012 x86_64

The programs included with the Debian GNU/Linux system are free software;
the exact distribution terms for each program are described in the
individual files in /usr/share/doc/*/copyright.

Debian GNU/Linux comes with ABSOLUTELY NO WARRANTY, to the extent
permitted by applicable law.
Last login: Sun Dec 23 20:14:24 2012 from 192.168.1.34
root@openstack-squeeze-amd64:~# apt-get update
root@openstack-squeeze-amd64:~# apt-get install vim curl parted # install some must have packages
[...]
root@openstack-squeeze-amd64:~# dpkg-reconfigure locales # I like to have fr_FR and en_US in my locales
[...]
root@openstack-squeeze-amd64:~# echo virtio_baloon >> /etc/modules
root@openstack-squeeze-amd64:~# echo acpiphp >> /etc/modules
root@openstack-squeeze-amd64:~# update-initramfs -u
root@openstack-squeeze-amd64:~# apt-get clean
root@openstack-squeeze-amd64:~# rm /etc/udev/rules.d/70-persistent-net.rules
root@openstack-squeeze-amd64:~# rm .bash_history
root@openstack-squeeze-amd64:~# poweroff

What we do here is to install some packages, do some configurations. The important part is adding the acpiphp module so the volume attachment will work in our instances. We also clean some stuffs up before shutting the VM down.

3. Convert the image into an AMI image

Since I created the VM image as a qcow2 image, I needed to convert it back to a raw image:

david@host:~$ scp root@virthost:/var/lib/libvirt/images/openstack-squeeze-amd64.img .
david@host:~$ qemu-img convert -O raw openstack-squeeze-amd64.img openstack-squeeze-amd64.raw

Then, as I want a minimal-sized disk image, the filesystem must be resized to minimal. I did this like described below, but I think there are simpler methods to do so.

david@host:~$ fdisk -l openstack-squeeze-amd64.raw  # display the partition location in the disk

Disk openstack-squeeze-amd64.raw: 5368 MB, 5368709120 bytes
149 heads, 8 sectors/track, 8796 cylinders, total 10485760 sectors
Units = sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 512 bytes
I/O size (minimum/optimal): 512 bytes / 512 bytes
Disk identifier: 0x0001fab7

                   Device Boot      Start         End      Blocks   Id  System
debian-squeeze-amd64.raw1            2048    10483711     5240832   83  Linux
david@host:~$ # extract the filesystem from the image
david@host:~$ dd if=openstack-squeeze-amd64.raw of=openstack-squeeze-amd64.ami bs=1024 skip=1024 count=5240832
david@host:~$ losetup /dev/loop1 openstack-squeeze-amd64.ami
david@host:~$ mkdir /tmp/img
david@host:~$ mount /dev/loop1 /tmp/img
david@host:~$ cp /tmp/img/boot/vmlinuz-2.6.32-5-amd64 .
david@host:~$ cp /tmp/img/boot/initrd.img-2.6.32-5-amd64 .
david@host:~$ umount /tmp/img
david@host:~$ e2fsck -f /dev/loop1 # required before a resize

e2fsck 1.42.5 (29-Jul-2012)
Pass 1: Checking inodes, blocks, and sizes
Pass 2: Checking directory structure
Pass 3: Checking directory connectivity
Pass 4: Checking reference counts
Pass 5: Checking group summary information
/dev/loop1: 26218/327680 files (0.2% non-contiguous), 201812/1310208 blocks
david@host:~$ resize2fs -M /dev/loop1 # minimize the filesystem

resize2fs 1.42.5 (29-Jul-2012)
Resizing the filesystem on /dev/loop1 to 191461 (4k) blocks.
The filesystem on /dev/loop1 is now 191461 blocks long.
david@host:~$ # note the new size ^^^^ and the block size above (4k)
david@host:~$ losetup -d /dev/loop1 # detach the lo device
david@host:~$ dd if=debian-squeeze-amd64.ami of=debian-squeeze-amd64-reduced.ami bs=4096 count=191461

4. Upload in OpenStack

After all this, you have a kernel image, a init ramdisk file and a minimized root filesystem image file. So you just have to upload them to your OpenStack image provider (glance):

david@host:~$ glance add disk_format=aki container_format=aki name="debian-squeeze-uec-x86_64-kernel" \
                 < vmlinuz-2.6.32-5-amd64
Uploading image 'debian-squeeze-uec-x86_64-kernel'
==================================================================================[100%] 24.1M/s, ETA  0h  0m  0s
Added new image with ID: 644e59b8-1503-403f-a4fe-746d4dac2ff8
david@host:~$ glance add disk_format=ari container_format=ari name="debian-squeeze-uec-x86_64-initrd" \
                 < initrd.img-2.6.32-5-amd64
Uploading image 'debian-squeeze-uec-x86_64-initrd'
==================================================================================[100%] 26.7M/s, ETA  0h  0m  0s
Added new image with ID: 6f75f1c9-1e27-4cb0-bbe0-d30defa8285c
david@host:~$ glance add disk_format=ami container_format=ami name="debian-squeeze-uec-x86_64" \
                 kernel_id=644e59b8-1503-403f-a4fe-746d4dac2ff8 ramdisk_id=6f75f1c9-1e27-4cb0-bbe0-d30defa8285c \
                 < debian-squeeze-amd64-reduced.ami
Uploading image 'debian-squeeze-uec-x86_64'
==================================================================================[100%] 42.1M/s, ETA  0h  0m  0s
Added new image with ID: 4abc09ae-ea34-44c5-8d54-504948e8d1f7
http://www.logilab.org/file/115220?vid=download

And that's it (!). I now have a working Debian squeeze image in my cloud that works fine:

http://www.logilab.org/file/115221?vid=download

Nazca is out !

2012/12/21 by Simon Chabot

What is it for ?

Nazca is a python library aiming to help you to align data. But, what does “align data” mean? For instance, you have a list of cities, described by their name and their country and you would like to find their URI on dbpedia to have more information about them, as the longitude and the latitude. If you have two or three cities, it can be done with bare hands, but it could not if there are hundreds or thousands cities. Nazca provides you all the stuff we need to do it.

This blog post aims to introduce you how this library works and can be used. Once you have understood the main concepts behind this library, don't hesitate to try Nazca online !

Introduction

The alignment process is divided into three main steps:

  1. Gather and format the data we want to align. In this step, we define two sets called the alignset and the targetset. The alignset contains our data, and the targetset contains the data on which we would like to make the links.
  2. Compute the similarity between the items gathered. We compute a distance matrix between the two sets according to a given distance.
  3. Find the items having a high similarity thanks to the distance matrix.

Simple case

  1. Let's define alignset and targetset as simple python lists.
alignset = ['Victor Hugo', 'Albert Camus']
targetset = ['Albert Camus', 'Guillaume Apollinaire', 'Victor Hugo']
  1. Now, we have to compute the similarity between each items. For that purpose, the Levenshtein distance [1], which is well accurate to compute the distance between few words, is used. Such a function is provided in the nazca.distance module.

    The next step is to compute the distance matrix according to the Levenshtein distance. The result is given in the following table.

     

    Albert Camus

    Guillaume Apollinaire

    Victor Hugo

    Victor Hugo

    6

    9

    0

    Albert Camus

    0

    8

    6

  2. The alignment process is ended by reading the matrix and saying items having a value inferior to a given threshold are identical.

[1]Also called the edit distance, because the distance between two words is equal to the number of single-character edits required to change one word into the other.

A more complex one

The previous case was simple, because we had only one attribute to align (the name), but it is frequent to have a lot of attributes to align, such as the name and the birth date and the birth city. The steps remain the same, except that three distance matrices will be computed, and items will be represented as nested lists. See the following example:

alignset = [['Paul Dupont', '14-08-1991', 'Paris'],
            ['Jacques Dupuis', '06-01-1999', 'Bressuire'],
            ['Michel Edouard', '18-04-1881', 'Nantes']]
targetset = [['Dupond Paul', '14/08/1991', 'Paris'],
             ['Edouard Michel', '18/04/1881', 'Nantes'],
             ['Dupuis Jacques ', '06/01/1999', 'Bressuire'],
             ['Dupont Paul', '01-12-2012', 'Paris']]

In such a case, two distance functions are used, the Levenshtein one for the name and the city and a temporal one for the birth date [2].

The cdist function of nazca.distances enables us to compute those matrices :

  • For the names:
>>> nazca.matrix.cdist([a[0] for a in alignset], [t[0] for t in targetset],
>>>                    'levenshtein', matrix_normalized=False)
array([[ 1.,  6.,  5.,  0.],
       [ 5.,  6.,  0.,  5.],
       [ 6.,  0.,  6.,  6.]], dtype=float32)
  Dupond Paul Edouard Michel Dupuis Jacques Dupont Paul
Paul Dupont 1 6 5 0
Jacques Dupuis 5 6 0 5
Edouard Michel 6 0 6 6
  • For the birthdates:
>>> nazca.matrix.cdist([a[1] for a in alignset], [t[1] for t in targetset],
>>>                    'temporal', matrix_normalized=False)
array([[     0.,  40294.,   2702.,   7780.],
       [  2702.,  42996.,      0.,   5078.],
       [ 40294.,      0.,  42996.,  48074.]], dtype=float32)
  14/08/1991 18/04/1881 06/01/1999 01-12-2012
14-08-1991 0 40294 2702 7780
06-01-1999 2702 42996 0 5078
18-04-1881 40294 0 42996 48074
  • For the birthplaces:
>>> nazca.matrix.cdist([a[2] for a in alignset], [t[2] for t in targetset],
>>>                    'levenshtein', matrix_normalized=False)
array([[ 0.,  4.,  8.,  0.],
       [ 8.,  9.,  0.,  8.],
       [ 4.,  0.,  9.,  4.]], dtype=float32)
  Paris Nantes Bressuire Paris
Paris 0 4 8 0
Bressuire 8 9 0 8
Nantes 4 0 9 4

The next step is gathering those three matrices into a global one, called the global alignment matrix. Thus we have :

  0 1 2 3
0 1 40304 2715 7780
1 2715 43011 0 5091
2 40304 0 43011 48084

Allowing some misspelling mistakes (for example Dupont and Dupond are very closed), the matching threshold can be set to 1 or 2. Thus we can see that the item 0 in our alignset is the same that the item 0 in the targetset, the 1 in the alignset and the 2 of the targetset too : the links can be done !

It's important to notice that even if the item 0 of the alignset and the 3 of the targetset have the same name and the same birthplace they are unlikely identical because of their very different birth date.

You may have noticed that working with matrices as I did for the example is a little bit boring. The good news is that Nazca makes all this job for you. You just have to give the sets and distance functions and that's all. An other good news is the project comes with the needed functions to build the sets !

[2]Provided in the nazca.distances module.

Real applications

Just before we start, we will assume the following imports have been done:

from nazca import dataio as aldio   #Functions for input and output data
from nazca import distances as ald  #Functions to compute the distances
from nazca import normalize as aln  #Functions to normalize data
from nazca import aligner as ala    #Functions to align data

The Goncourt prize

On wikipedia, we can find the Goncourt prize winners, and we would like to establish a link between the winners and their URI on dbpedia (Let's imagine the Goncourt prize winners category does not exist in dbpedia)

We simply copy/paste the winners list of wikipedia into a file and replace all the separators (- and ,) by #. So, the beginning of our file is :

1903#John-Antoine Nau#Force ennemie (Plume)
1904#Léon Frapié#La Maternelle (Albin Michel)
1905#Claude Farrère#Les Civilisés (Paul Ollendorff)
1906#Jérôme et Jean Tharaud#Dingley, l'illustre écrivain (Cahiers de la Quinzaine)

When using the high-level functions of this library, each item must have at least two elements: an identifier (the name, or the URI) and the attribute to compare. With the previous file, we will use the name (so the column number 1) as identifier (we don't have an URI here as identifier) and attribute to align. This is told to python thanks to the following code:

alignset = aldio.parsefile('prixgoncourt', indexes=[1, 1], delimiter='#')

So, the beginning of our alignset is:

>>> alignset[:3]
[[u'John-Antoine Nau', u'John-Antoine Nau'],
 [u'Léon Frapié', u'Léon, Frapié'],
 [u'Claude Farrère', u'Claude Farrère']]

Now, let's build the targetset thanks to a sparql query and the dbpedia end-point. We ask for the list of the French novelists, described by their URI and their name in French:

query = """
     SELECT ?writer, ?name WHERE {
       ?writer  <http://purl.org/dc/terms/subject> <http://dbpedia.org/resource/Category:French_novelists>.
       ?writer rdfs:label ?name.
       FILTER(lang(?name) = 'fr')
    }
 """
 targetset = aldio.sparqlquery('http://dbpedia.org/sparql', query)

Both functions return nested lists as presented before. Now, we have to define the distance function to be used for the alignment. This is done thanks to a python dictionary where the keys are the columns to work on, and the values are the treatments to apply.

treatments = {1: {'metric': ald.levenshtein}} # Use a levenshtein on the name
                                              # (column 1)

Finally, the last thing we have to do, is to call the alignall function:

alignments = ala.alignall(alignset, targetset,
                       0.4, #This is the matching threshold
                       treatments,
                       mode=None,#We'll discuss about that later
                       uniq=True #Get the best results only
                      )

This function returns an iterator over the different alignments done. You can see the results thanks to the following code :

for a, t in alignments:
    print '%s has been aligned onto %s' % (a, t)

It may be important to apply some pre-treatment on the data to align. For instance, names can be written with lower or upper characters, with extra characters as punctuation or unwanted information in parenthesis and so on. That is why we provide some functions to normalize your data. The most useful may be the simplify() function (see the docstring for more information). So the treatments list can be given as follow:

def remove_after(string, sub):
    """ Remove the text after ``sub`` in ``string``
        >>> remove_after('I like cats and dogs', 'and')
        'I like cats'
        >>> remove_after('I like cats and dogs', '(')
        'I like cats and dogs'
    """
    try:
        return string[:string.lower().index(sub.lower())].strip()
    except ValueError:
        return string


treatments = {1: {'normalization': [lambda x:remove_after(x, '('),
                                    aln.simply],
                  'metric': ald.levenshtein
                 }
             }

Cities alignment

The previous case with the Goncourt prize winners was pretty simply because the number of items was small, and the computation fast. But in a more real use case, the number of items to align may be huge (some thousands or millions…). In such a case it's unthinkable to build the global alignment matrix because it would be too big and it would take (at least...) fews days to achieve the computation. So the idea is to make small groups of possible similar data to compute smaller matrices (i.e. a divide and conquer approach). For this purpose, we provide some functions to group/cluster data. We have functions to group text and numerical data.

This is the code used, we will explain it:

targetset = aldio.rqlquery('http://demo.cubicweb.org/geonames',
                           """Any U, N, LONG, LAT WHERE X is Location, X name
                              N, X country C, C name "France", X longitude
                              LONG, X latitude LAT, X population > 1000, X
                              feature_class "P", X cwuri U""",
                           indexes=[0, 1, (2, 3)])
alignset = aldio.sparqlquery('http://dbpedia.inria.fr/sparql',
                             """prefix db-owl: <http://dbpedia.org/ontology/>
                             prefix db-prop: <http://fr.dbpedia.org/property/>
                             select ?ville, ?name, ?long, ?lat where {
                              ?ville db-owl:country <http://fr.dbpedia.org/resource/France> .
                              ?ville rdf:type db-owl:PopulatedPlace .
                              ?ville db-owl:populationTotal ?population .
                              ?ville foaf:name ?name .
                              ?ville db-prop:longitude ?long .
                              ?ville db-prop:latitude ?lat .
                              FILTER (?population > 1000)
                             }""",
                             indexes=[0, 1, (2, 3)])


treatments = {1: {'normalization': [aln.simply],
                  'metric': ald.levenshtein,
                  'matrix_normalized': False
                 }
             }
results = ala.alignall(alignset, targetset, 3, treatments=treatments, #As before
                       indexes=(2, 2), #On which data build the kdtree
                       mode='kdtree',  #The mode to use
                       uniq=True) #Return only the best results

Let's explain the code. We have two files, containing a list of cities we want to align, the first column is the identifier, and the second is the name of the city and the last one is location of the city (longitude and latitude), gathered into a single tuple.

In this example, we want to build a kdtree on the couple (longitude, latitude) to divide our data in few candidates. This clustering is coarse, and is only used to reduce the potential candidats without loosing any more refined possible matchs.

So, in the next step, we define the treatments to apply. It is the same as before, but we ask for a non-normalized matrix (ie: the real output of the levenshtein distance). Thus, we call the alignall function. indexes is a tuple saying the position of the point on which the kdtree must be built, mode is the mode used to find neighbours [3].

Finally, uniq ask to the function to return the best candidate (ie: the one having the shortest distance below the given threshold)

The function outputs a generator yielding tuples where the first element is the identifier of the alignset item and the second is the targetset one (It may take some time before yielding the first tuples, because all the computation must be done…)

[3]The available modes are kdtree, kmeans and minibatch for numerical data and minhashing for text one.

Try it online !

We have also made this little application of Nazca, using Cubicweb. This application provides a user interface for Nazca, helping you to choose what you want to align. You can use sparql or rql queries, as in the previous example, or import your own cvs file [4]. Once you have choosen what you want to align, you can click the Next step button to customize the treatments you want to apply, just as you did before in python ! Once done, by clicking the Next step, you start the alignment process. Wait a little bit, and you can either download the results in a csv or rdf file, or directly see the results online choosing the html output.

[4]Your csv file must be tab-separated for the moment…

Openstack, Wheezy and ZFS on Linux

2012/12/19 by David Douard

Openstack, Wheezy and ZFS on Linux

A while ago, I started the install of an OpenStack cluster at Logilab, so our developers can play easily with any kind of environment. We are planning to improve our Apycot automatic testing platform so it can use "elastic power". And so on.

http://www.openstack.org/themes/openstack/images/open-stack-cloud-computing-logo-2.png

I first tried a Ubuntu Precise based setup, since at that time, Debian packages were not really usable. The setup never reached a point where it could be relased as production ready, due to the fact I tried a too complex and bleeding edge configuration (involving Quantum, openvswitch, sheepdog)...

Meanwhile, we went really short of storage capacity. For now, it mainly consists in hard drives distributed in our 19" Dell racks (generally with hardware RAID controllers). So I recently purchased a low-cost storage bay (SuperMicro SC937 with a 6Gb/s JBOD-only HBA) with 18 spinning hard drives and 4 SSDs. This storage bay being driven by ZFS on Linux (tip: the SSD-stored ZIL is a requirement to get decent performances). This storage setup is still under test for now.

http://zfsonlinux.org/images/zfs-linux.png

I also went to the last Mini-DebConf in Paris, where Loic Dachary presented the status of the OpenStack packaging effort in Debian. This gave me the will to give a new try to OpenStack using Wheezy and a bit simpler setup. But I could not consider not to use my new ZFS-based storage as a nova volume provider. It is not available for now in OpenStack (there is a backend for Solaris, but not for ZFS on Linux). However, this is Python and in fact, the current ISCSIDriver backend needs very little to make it work with zfs instead of lvm as "elastics" block-volume provider and manager.

So, I wrote a custom nova volume driver to handle this. As I don't want the nova-volume daemon to run on my ZFS SAN, I wrote this backend mixing the SanISCSIDriver (which manages the storage system via SSH) and the standard ISCSIDriver (which uses standard Linux isci target tools). I'm not very fond of the API of the VolumeDriver (especially the fact that the ISCSIDriver is responsible for 2 roles: managing block-level volumes and exporting block-level volumes). This small design flaw (IMHO) is the reason I had to duplicate some code (not much but...) to implement my ZFSonLinuxISCSIDriver...

So here is the setup I made:

Infrastructure

My OpenStack Essex "cluster" consists for now in:

  • one control node, running in a "normal" libvirt-controlled virtual machine; it is a Wheezy that runs:
    • nova-api
    • nova-cert
    • nova-network
    • nova-scheduler
    • nova-volume
    • glance
    • postgresql
    • OpenStack dashboard
  • one computing node (Dell R310, Xeon X3480, 32G, Wheezy), which runs:
    • nova-api
    • nova-network
    • nova-compute
  • ZFS-on-Linux SAN (3x raidz1 poools made of 6 1T drives, 2x (mirrored) 32G SLC SDDs, 2x 120G MLC SSDs for cache); for now, the storage is exported to the SAN via one 1G ethernet link.

OpensStack Essex setup

I mainly followed the Debian HOWTO to setup my private cloud. I mainly tuned the network settings to match my environement (and the fact my control node lives in a VM, with VLAN stuff handled by the host).

I easily got a working setup (I must admit that I think my previous experiment with OpenStack helped a lot when dealing with custom configurations... and vocabulary; I'm not sure I would have succeded "easily" following the HOWTO, but hey, it is a functionnal HOWTO, meaning if you do not follow the instructions because you want special tunings, don't blame the HOWTO).

Compared to the HOWTO, my nova.conf looks like (as of today):

[DEFAULT]
logdir=/var/log/nova
state_path=/var/lib/nova
lock_path=/var/lock/nova
root_helper=sudo nova-rootwrap
auth_strategy=keystone
dhcpbridge_flagfile=/etc/nova/nova.conf
dhcpbridge=/usr/bin/nova-dhcpbridge
sql_connection=postgresql://novacommon:XXX@control.openstack.logilab.fr/nova

##  Network config
# A nova-network on each compute node
multi_host=true
# VLan manger
network_manager=nova.network.manager.VlanManager
vlan_interface=eth1
# My ip
my-ip=172.17.10.2
public_interface=eth0
# Dmz & metadata things
dmz_cidr=169.254.169.254/32
ec2_dmz_host=169.254.169.254
metadata_host=169.254.169.254

## More general things
# The RabbitMQ host
rabbit_host=control.openstack.logilab.fr

## Glance
image_service=nova.image.glance.GlanceImageService
glance_api_servers=control.openstack.logilab.fr:9292
use-syslog=true
ec2_host=control.openstack.logilab.fr

novncproxy_base_url=http://control.openstack.logilab.fr:6080/vnc_auto.html
vncserver_listen=0.0.0.0
vncserver_proxyclient_address=127.0.0.1

Volume

I had a bit more work to do to make nova-volume work. First, I got hit by this nasty bug #695791 which is trivial to fix... when you know how to fix it (I noticed the bug report after I fixed it by myself).

Then, as I wanted the volumes to be stored and exported by my shiny new ZFS-on-Linux setup, I had to write my own volume driver, which was quite easy, since it is Python, and the logic to implement was already provided by the ISCSIDriver class on the one hand, and by the SanISCSIDrvier on the other hand. So I ended with this firt implementation. This file should be copied to nova volumes package directory (nova/volume/zol.py):

# vim: tabstop=4 shiftwidth=4 softtabstop=4

# Copyright 2010 United States Government as represented by the
# Administrator of the National Aeronautics and Space Administration.
# Copyright 2011 Justin Santa Barbara
# Copyright 2012 David DOUARD, LOGILAB S.A.
# All Rights Reserved.
#
#    Licensed under the Apache License, Version 2.0 (the "License"); you may
#    not use this file except in compliance with the License. You may obtain
#    a copy of the License at
#
#         http://www.apache.org/licenses/LICENSE-2.0
#
#    Unless required by applicable law or agreed to in writing, software
#    distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
#    WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
#    License for the specific language governing permissions and limitations
#    under the License.
"""
Driver for ZFS-on-Linux-stored volumes.

This is mainly a custom version of the ISCSIDriver that uses ZFS as
volume provider, generally accessed over SSH.
"""

import os

from nova import exception
from nova import flags
from nova import utils
from nova import log as logging
from nova.openstack.common import cfg
from nova.volume.driver import _iscsi_location
from nova.volume import iscsi
from nova.volume.san import SanISCSIDriver


LOG = logging.getLogger(__name__)

san_opts = [
    cfg.StrOpt('san_zfs_command',
               default='/sbin/zfs',
               help='The ZFS command.'),
    ]

FLAGS = flags.FLAGS
FLAGS.register_opts(san_opts)


class ZFSonLinuxISCSIDriver(SanISCSIDriver):
    """Executes commands relating to ZFS-on-Linux-hosted ISCSI volumes.

    Basic setup for a ZoL iSCSI server:

    XXX

    Note that current implementation of ZFS on Linux does not handle:

      zfs allow/unallow

    For now, needs to have root access to the ZFS host. The best is to
    use a ssh key with ssh authorized_keys restriction mechanisms to
    limit root access.

    Make sure you can login using san_login & san_password/san_private_key
    """
    ZFSCMD = FLAGS.san_zfs_command

    _local_execute = utils.execute

    def _getrl(self):
        return self._runlocal
    def _setrl(self, v):
        if isinstance(v, basestring):
            v = v.lower() in ('true', 't', '1', 'y', 'yes')
        self._runlocal = v
    run_local = property(_getrl, _setrl)

    def __init__(self):
        super(ZFSonLinuxISCSIDriver, self).__init__()
        self.tgtadm.set_execute(self._execute)
        LOG.info("run local = %s (%s)" % (self.run_local, FLAGS.san_is_local))

    def set_execute(self, execute):
        LOG.debug("override local execute cmd with %s (%s)" %
                  (repr(execute), execute.__module__))
        self._local_execute = execute

    def _execute(self, *cmd, **kwargs):
        if self.run_local:
            LOG.debug("LOCAL execute cmd %s (%s)" % (cmd, kwargs))
            return self._local_execute(*cmd, **kwargs)
        else:
            LOG.debug("SSH execute cmd %s (%s)" % (cmd, kwargs))
            check_exit_code = kwargs.pop('check_exit_code', None)
            command = ' '.join(cmd)
            return self._run_ssh(command, check_exit_code)

    def _create_volume(self, volume_name, sizestr):
        zfs_poolname = self._build_zfs_poolname(volume_name)

        # Create a zfs volume
        cmd = [self.ZFSCMD, 'create']
        if FLAGS.san_thin_provision:
            cmd.append('-s')
        cmd.extend(['-V', sizestr])
        cmd.append(zfs_poolname)
        self._execute(*cmd)

    def _volume_not_present(self, volume_name):
        zfs_poolname = self._build_zfs_poolname(volume_name)
        try:
            out, err = self._execute(self.ZFSCMD, 'list', '-H', zfs_poolname)
            if out.startswith(zfs_poolname):
                return False
        except Exception as e:
            # If the volume isn't present
            return True
        return False

    def create_volume_from_snapshot(self, volume, snapshot):
        """Creates a volume from a snapshot."""
        zfs_snap = self._build_zfs_poolname(snapshot['name'])
        zfs_vol = self._build_zfs_poolname(snapshot['name'])
        self._execute(self.ZFSCMD, 'clone', zfs_snap, zfs_vol)
        self._execute(self.ZFSCMD, 'promote', zfs_vol)

    def delete_volume(self, volume):
        """Deletes a volume."""
        if self._volume_not_present(volume['name']):
            # If the volume isn't present, then don't attempt to delete
            return True
        zfs_poolname = self._build_zfs_poolname(volume['name'])
        self._execute(self.ZFSCMD, 'destroy', zfs_poolname)

    def create_export(self, context, volume):
        """Creates an export for a logical volume."""
        self._ensure_iscsi_targets(context, volume['host'])
        iscsi_target = self.db.volume_allocate_iscsi_target(context,
                                                            volume['id'],
                                                      volume['host'])
        iscsi_name = "%s%s" % (FLAGS.iscsi_target_prefix, volume['name'])
        volume_path = self.local_path(volume)

        # XXX (ddouard) this code is not robust: does not check for
        # existing iscsi targets on the host (ie. not created by
        # nova), but fixing it require a deep refactoring of the iscsi
        # handling code (which is what have been done in cinder)
        self.tgtadm.new_target(iscsi_name, iscsi_target)
        self.tgtadm.new_logicalunit(iscsi_target, 0, volume_path)

        if FLAGS.iscsi_helper == 'tgtadm':
            lun = 1
        else:
            lun = 0
        if self.run_local:
            iscsi_ip_address = FLAGS.iscsi_ip_address
        else:
            iscsi_ip_address = FLAGS.san_ip
        return {'provider_location': _iscsi_location(
                iscsi_ip_address, iscsi_target, iscsi_name, lun)}

    def remove_export(self, context, volume):
        """Removes an export for a logical volume."""
        try:
            iscsi_target = self.db.volume_get_iscsi_target_num(context,
                                                           volume['id'])
        except exception.NotFound:
            LOG.info(_("Skipping remove_export. No iscsi_target " +
                       "provisioned for volume: %d"), volume['id'])
            return

        try:
            # ietadm show will exit with an error
            # this export has already been removed
            self.tgtadm.show_target(iscsi_target)
        except Exception as e:
            LOG.info(_("Skipping remove_export. No iscsi_target " +
                       "is presently exported for volume: %d"), volume['id'])
            return

        self.tgtadm.delete_logicalunit(iscsi_target, 0)
        self.tgtadm.delete_target(iscsi_target)

    def check_for_export(self, context, volume_id):
        """Make sure volume is exported."""
        tid = self.db.volume_get_iscsi_target_num(context, volume_id)
        try:
            self.tgtadm.show_target(tid)
        except exception.ProcessExecutionError, e:
            # Instances remount read-only in this case.
            # /etc/init.d/iscsitarget restart and rebooting nova-volume
            # is better since ensure_export() works at boot time.
            LOG.error(_("Cannot confirm exported volume "
                        "id:%(volume_id)s.") % locals())
            raise

    def local_path(self, volume):
        zfs_poolname = self._build_zfs_poolname(volume['name'])
        zvoldev = '/dev/zvol/%s' % zfs_poolname
        return zvoldev

    def _build_zfs_poolname(self, volume_name):
        zfs_poolname = '%s%s' % (FLAGS.san_zfs_volume_base, volume_name)
        return zfs_poolname

To configure my nova-volume instance (which runs on the control node, since it's only a manager), I added these to my nova.conf file:

# nove-volume config
volume_driver=nova.volume.zol.ZFSonLinuxISCSIDriver
iscsi_ip_address=172.17.1.7
iscsi_helper=tgtadm
san_thin_provision=false
san_ip=172.17.1.7
san_private_key=/etc/nova/sankey
san_login=root
san_zfs_volume_base=data/openstack/volume/
san_is_local=false
verbose=true

Note that the private key (/etc/nova/sankey here) is stored in clear and that it must be readable by the nova user.

This key being stored in clear and giving root acces to my ZFS host, I have limited a bit this root access by using a custom command wrapper in the .ssh/authorized_keys file.

Something like (naive implementation):

[root@zfshost ~]$ cat /root/zfswrapper
#!/bin/sh
CMD=`echo $SSH_ORIGINAL_COMMAND | awk '{print $1}'`
if [ "$CMD" != "/sbin/zfs" && "$CMD" != "tgtadm" ]; then
  echo "Can do only zfs/tgtadm stuff here"
  exit 1
fi

echo "[`date`] $SSH_ORIGINAL_COMMAND" >> .zfsopenstack.log
exec $SSH_ORIGINAL_COMMAND

Using this in root's .ssh/authorized_keys file:

[root@zfshost ~]$ cat /root/.ssh/authorized_keys | grep control
from="control.openstack.logilab.fr",no-pty,no-port-forwarding,no-X11-forwarding, \
      no-agent-forwarding,command="/root/zfswrapper" ssh-rsa AAAA[...] root@control

I had to set the iscsi_ip_address (the ip address of the ZFS host), but I think this is a result of something mistakenly implemented in my ZFSonLinux driver.

Using this config, I can boot an image, create a volume on my ZFS storage, and attach it to the running image.

I have to test things like snapshot, (live?) migration and so. This is a very first draft implementation which needs to be refined, improved and tested.

What's next

Besides the fact that it needs more tests, I plan to use salt for my OpenStack deployment (first to add more compute nodes in my cluster), and on the other side, I'd like to try the salt-cloud so I have a bunch of Debian images that "just work" (without the need of porting the cloud-init Ubuntu package).

On the side of my zol driver, I need to port it to Cinder, but I do not have a Folsom install to test it...


Announcing pylint.org

2012/12/04 by Arthur Lutz

Pylint - the world renowned Python code static checker - now has a landing page : http://www.pylint.org

http://www.python.org/images/python-logo.gif

We've tried to summarize all the things a newcomer should know about pylint. We hope it reflects the diversity of uses and support canals for pylint.

Open and decentralized Web

Note that pylint is not hosted on github or another well-known forge, since we firmly believe in a decentralized architecture for the web.

This applies especially to open source software development. Pylint's development is self-hosted on a forge and its code is version-controlled with mercurial, a distributed version control system (DVCS). Both tools are free software written in python.

http://www.zjulian.com/wp-content/uploads/2012/05/Centralized-Decentralized-And-Distributed-System.jpg

We know centralized (and closed source) platforms for managing software projects can make things easier for contributors. We have enabled a mirror on bitbucket (and pylint-brain) so as to ease forks and pull requests. Pull requests can be made there and even from a self-hosted mercurial (with a quick email on the mailing-list).

Feel free to add your comments or feedback below.