We are pleased to announce the availability of this new release, which comes with some exciting new features and improvements. We would like to thank everyone who participated in this release in a way or another.
This version contains a few bug fixes and improvements over 1.9, mostly affecting packaging:
clustershellto be installed as user in a
pip installor using
pip install --userwith man pages. Root installation using pip is now discouraged. If done,
/usr/localis likely to be used as the install prefix. See Installing ClusterShell the Python way for more information.
$CFGDIRwas broken if
/etc/clustershelldid not exist
- Add support for negative ranges in
For more details, please have a look at GitHub Issues for 1.9.1 milestone.
Main changes in 1.9¶
Support for Python 2.6 has been dropped in this version. Upgrading to Python 3 is highly recommended as Python 2 reached end of life in 2020. See Requirements.
- clush has now support for Run modes to support more
authentication use cases. A run mode has pre-defined
clush.conf settings with a given name, and can then
be activated with
--mode=MODE. We also added the new options
password_prompt(see clush.conf). Two examples of run modes are included and can be easily enabled:
clush.conf comes with a new variable
confdir to specify where to look for run mode
configuration files. If you upgrade from 1.8.4 and want to use run modes,
confdir is present in your
- clush: add arguments
--errdir=ERRDIR; similar to pssh(1), it allows to save the standard output (stdout) and/or error (stderr) of all remote commands to local files. See Saving output in files.
Node sets and node groups¶
To support mixed-length 0-padding ranges, version 1.9 introduces
RangeSet's API that might break existing code. If you
RangeSet directly, see below for more information.
RangeSetNDobjects now support sets with mixed length zero padding, meaning you can safely mix ranges like
005-123. The following example with nodeset shows that not only
001are now seen as separate indexes, but it is also possible to mix non-padded indexes like
1with zero-padded indexes:
$ nodeset --fold node001 node1 node01 node[1,01,001]
nodeset's zero padding for more examples.
RangeSetnow internally manages indexes as strings with the zero padding included. Prior to v1.9, indexes were stored as integers and zero padding was a simple display feature of fixed length per
RangeSetobject. If you are using this class directly in your code, please see the RangeSet class in the Programming Guide section for portability recommendations (especially the new method
NodeSet class API has NOT changed so as long as you do
RangeSet directly, you may safely upgrade to 1.9.
Listing group names in expressions: the @@ operator may be used in any node set expression to manipulate group names as a node set:
$ nodeset -l -s rack @rack:J1 @rack:J2 @rack:J3 $ nodeset -f @@rack J[1-3]
RangeSet: multidimensional folding performance optimization, useful for "xnames" on HPE Cray EX supercomputers that encode up to 5 dimensions.
Slurm group bindings: filter out more Slurm node state flags
$CLUSTERSHELL_CFGDIRas an alternate location for configuration files; useful on a cluster where ClusterShell is provided as a user-facing tool installed on a shared file system (see clush.conf, groups.conf and Library Defaults).
- Fix start by implementing a proper asynchronous start for
TreeWorker, which is now only triggered when the engine actually starts.
- Fix error with intermediate gateways
For more details, please have a look at GitHub Issues for 1.9 milestone.
This adaptive major release is now compatible with both Python 2 and Python 3.
We hope this release will help you manage your clusters, server farms or cloud farms! Special thanks to the many of you that have sent us feedback on GitHub!
Support for Python 2.5 and below has been dropped in this version.
This version contains a few bug fixes and improvements:
- allow out-of-tree worker modules
- use default local_worker and allow overriding Library Defaults (tree mode)
- return maxrc properly in the case of the Rsh Worker
- clush: improve stdin support with Python 3
- clush: add maxrc option to clush.conf
- clush: add support for NO_COLOR and CLICOLOR
For more details, please have a look at GitHub Issues for 1.8.4 milestone.
This version contains a few bug fixes and improvements, mostly affecting the tree mode:
CLUSTERSHELL_GW_PYTHON_EXECUTABLEenvironment variable to remote gateways (see Remote Python executable)
- fix defect to properly close gateway channel when worker has aborted
- improve error reporting from gateways
- clush: now properly handles
--worker=sshwhen topology.conf is present to explicitly disable tree mode
- use safe yaml load variant to avoid warning from
For more details, please have a look at GitHub Issues for 1.8.3 milestone.
We also added a Python support matrix for the main Linux distributions.
This version contains a few minor fixes:
- clush: support UTF-8 string encoding with --diff
- in some cases, timers were too fast due to an
- fix issue in the Slurm group bindings where job ids were used instead of user names
- performance update for xCAT group bindings
For more details, please have a look at GitHub Issues for 1.8.2 milestone.
Version 1.8.2 adds support for Python 3.7.
This version still supports Python 2.6 and thus also RHEL/CentOS 6, but please note that ClusterShell 1.9 is expected to require at least Python 2.7.
Version 1.8.2 adds support for RHEL 8/CentOS 8 and Fedora 31+, where only the
Python 3 package is provided. The
clustershell packages will be made
available in EPEL-8 as soon as possible.
No packaging changes were made to
clustershell in RHEL/CentOS 6 or 7.
This update contains a few bug fixes and some performance improvements of the
The tree mode has been fixed to properly support offline gateways.
We added the following command line options:
--confto specify alternative clush.conf (clush only)
--groupsconfto specify alternative groups.conf (all CLIs)
EventHandler, we reinstated
EventHandler.ev_error(): (as deprecated) for compatibility purposes.
Please see below for more details about important
changes in 1.8.
Finally, cluset/nodeset have been improved by adding support for:
- literal new line in
- multiline shell variables in options
For more details, please have a look at GitHub Issues for 1.8.1 milestone.
Main changes in 1.8¶
For more details, please have a look at GitHub Issues for 1.8 milestone.
CLI (command line interface)¶
If you use the clush or cluset/nodeset tools, there are no major changes since 1.7, though a few bug fixes and improvements have been done:
It is now possible to work with numeric node names with cluset/nodeset:
$ nodeset --fold 6704 6705 r931 r930 [6704-6705],r[930-931] $ squeue -h -o '%i' -u $USER | cluset -f [680240-680245,680310]
As a reminder, cluset/nodeset has always had an option to switch to numerical cluster ranges (only), using
$ squeue -h -o '%i' -u $USER | cluset -f -R 680240-680245,680310
Node group configuration is now loaded and processed only when required. This is actually an improvement of the
NodeSetclass that the tools readily benefit. This should improve both usability and performance.
YAML group files are now ignored for users that don't have the permission to read them (see File-based group sources for more info about group files).
clush now use slightly different colors that are legible on dark backgrounds.
- Better detection of the Python executable, and, if needed, we added a new environment variable to override it, see Remote Python executable.
- You must use the same major version of Python on the gateways and the root node.
If you're a developer and use the ClusterShell Python library, please read below.
Python 3 support¶
Starting in 1.8, the library can also be used with Python 3. The code is compatible with both Python 2 and 3 at the same time. To make it possible, we performed a full code refactoring (without changing the behavior).
When using Python 3, we recommend Python 3.4 or any more recent version.
Improved Event API¶
We've made some changes to
EventHandler, a class that defines a
simple interface to handle events generated by
Please note that all programs already based on
work with this new version of ClusterShell without any code change (backward
API compatibility across 1.x versions is enforced). We use object
introspection, the ability to determine the type of an object at runtime,
to make the Event API evolve smoothly. We do still recommend to change your
code as soon as possible as we'll break backward compatibility in the future
major release 2.0.
The signatures of the following
EventHandler methods changed in
Both old and new signatures are supported in 1.8. The old signatures will be deprecated in a future 1.x release and removed in version 2.0.
The new methods aims to be more convenient to use by avoiding the need of
Worker attributes like
worker.current_node (replaced with the
node argument in that case).
Also, please note that the following
EventHandler methods will be
removed in 2.0:
EventHandler.ev_error(): its use should be replaced with
EventHandler.ev_read()by comparing the stream name
Worker.SNAME_STDERR, like in the example below:
class MyEventHandler(EventHandler): def ev_read(self, worker, node, sname, msg): if sname == worker.SNAME_STDERR: print('error from %s: %s' % (node, msg))
EventHandler.ev_timeout(): its use should be replaced with
EventHandler.ev_close()by checking for the new
timedoutargument, which is set to
Truewhen a timeout occurred.
We recommend developers to start using the improved
Event API now.
Please don't forget to update your packaging requirements to use ClusterShell
1.8 or later.
Task and standard input (stdin)¶
Task.run() have a new
argument which if set to
False prevents the use of stdin by sending
EOF at first read, like if it is connected to /dev/null.
If not specified, its value is managed by the Library Defaults.
Its default value in
Defaults is set to
True for backward
compatibility, but could change in a future major release.
If your program doesn't plan to listen to stdin, it is recommended to set
stdin=False when calling these two methods.
We recommend that package maintainers use separate subpackages for Python 2 and Python 3, to install ClusterShell modules and related command line tools. The Python 2 and Python 3 stacks should be fully installable in parallel.
For the RPM packaging, there is now two subpackages
python34-clustershell in EPEL), each providing
the library and tools for the corresponding version of Python.
clustershell package includes the common configuration files and
documentation and requires
python2-clustershell, mainly because
Python 2 is still the default interpreter on most operating systems.
vim-clustershell was confusing so we removed it and added the vim
extensions to the main
Version 1.8 should be readily available as RPMs in the following distributions or RPM repositories:
- EPEL 6 and 7
- Fedora 26 and 27
- openSUSE Factory and Leap
On a supported environment, you can expect a smooth upgrade from version 1.6+.
We also expect the packaging to be updated for Debian.
It's just a small version bump from the well-known 1.6 version, but ClusterShell 1.7 comes with some nice new features that we hope you'll enjoy! Most of these features have already been tested on some very large Linux production systems.
Version 1.7 and possible future minor versions 1.7.x are compatible with Python 2.4 up to Python 2.7 (for example: from RedHat EL5 to EL7). Upgrade from version 1.6 to 1.7 should be painless and is fully supported.
This update contains a few bug fixes and some interesting performance improvements. This is also the first release published under the GNU Lesser General Public License, version 2.1 or later (LGPL v2.1+). Previous releases were published under the CeCILL-C V1.
Quite a bit of work has been done on the fanout of processes that the library uses to execute commands. We implemenented a basic per-worker fanout to fix the broken behaviour in tree mode. Thanks to this, it is now possible to use fanout=1 with gateways. The documentation has also been clarified.
An issue that led to broken pipe errors but also affected performance has been fixed in tree mode when copying files.
An issue with clush -L where nodes weren't always properly sorted has been fixed.
The performance of
MsgTree, the class used by the library to
aggregate identical command outputs, has been improved. We have seen up to 75%
speed improvement in some cases.
Finally, a cluset command has been added to avoid a conflict with xCAT nodeset command. It is the same command as nodeset.
For more details, please have a look at GitHub Issues for 1.7.3 milestone.
ClusterShell 1.7.3 is compatible with Python 2.4 up to Python 2.7 (for example: from RedHat EL5 to EL7). Upgrades from versions 1.6 or 1.7 are supported.
This minor version fixes a defect in tree mode that led to broken pipe errors or unwanted backtraces.
NodeSet class now supports the empty string as input. In
practice, you may now safely reuse the output of a
nodeset command as input argument for another
nodeset command, even if the result is an empty string.
A new option
--pick is available for clush and
nodeset to pick N node(s) at random from the resulting
For more details, please have a look at GitHub Issues for 1.7.2 milestone.
ClusterShell 1.7.2 is compatible with Python 2.4 up to Python 2.7 (for example: from RedHat EL5 to EL7). Upgrades from versions 1.6 or 1.7 are supported.
This minor version contains a few bug fixes, mostly related to Node sets handling.
This version also contains bug fixes and performance improvements in tree propagation mode.
For more details, please have a look at GitHub Issues for 1.7.1 milestone.
ClusterShell 1.7.1 is compatible with Python 2.4 up to Python 2.7 (for example: from RedHat EL5 to EL7). Upgrades from versions 1.6 or 1.7 are supported.
Main changes in 1.7¶
This new version comes with a refreshed documentation, based on the Sphinx documentation generator, available on http://clustershell.readthedocs.org.
The main new features of version 1.7 are described below.
NodeSet class and nodeset command-line
have been improved to support multidimentional node sets with folding
capability. The use of nD naming scheme is sometimes used to map node names to
physical location like
name-<rack>-<position> or node position within the
cluster interconnect network topology.
A first example of 3D nodeset expansion is a good way to start:
$ nodeset -e gpu-[1,3]-[4-5]-[0-6/2] gpu-1-4-0 gpu-1-4-2 gpu-1-4-4 gpu-1-4-6 gpu-1-5-0 gpu-1-5-2 gpu-1-5-4 gpu-1-5-6 gpu-3-4-0 gpu-3-4-2 gpu-3-4-4 gpu-3-4-6 gpu-3-5-0 gpu-3-5-2 gpu-3-5-4 gpu-3-5-6
You've probably noticed the
/2 notation of the last dimension. It's called
a step and behaves as one would expect, and is fully supported with nD
All other nodeset commands and options are supported with nD nodesets. For example, it's always useful to have a quick way to count the number of nodes in a nodeset:
$ nodeset -c gpu-[1,3]-[4-5]-[0-6/2] 16
Then to show the most interesting new capability of the underlying
NodeSet class in version 1.7, a folding example is probably
$ nodeset -f compute-1-[1-34] compute-2-[1-34] compute-[1-2]-[1-34]
In the above example, nodeset will try to find a very compact nodesets representation whenever possible. ClusterShell is probably the first and only cluster tool capable of doing such complex nodeset folding.
Attention, as not all cluster tools are supporting this kind of complex
nodesets, even for nodeset expansion, we added an
--axis option to select
to fold along some desired dimension:
$ nodeset --axis 2 -f compute-[1-2]-[1-34] compute-1-[1-34],compute-2-[1-34]
The last dimension can also be selected using
$ nodeset --axis -1 -f compute-[1-2]-[1-34] compute-1-[1-34],compute-2-[1-34]
All set-like operations are also supported with several dimensions, for
example difference (
$ nodeset -f c-[1-10]-[1-44] -x c-[5-10]-[1-34] c-[1-4]-[1-44],c-[5-10]-[35-44]
Hard to follow? Don't worry, ClusterShell does it for you!
File-based node groups¶
Cluster node groups have been a great success of previous version of ClusterShell and are now widely adopted. So we worked on improving it even more for version 1.7.
For those of you who use the file
/etc/clustershell/group to describe
node groups, that is still supported in 1.7 and upgrade from your 1.6 setup
should work just fine. However, for new 1.7 installations, we have put this
file in a different location by default:
$ vim /etc/clustershell/groups.d/local.cfg
Especially if you're starting a new setup, you have also the choice to switch to a more advanced groups YAML configuration file that can define multiple sources in a single file (equivalent to separate namespaces for node groups). The YAML format possibly allows you to edit the file content with YAML tools but it's also a file format convenient to edit just using the vim editor. To enable the example file, you need to rename it first as it needs to have the .yaml extension:
$ cd /etc/clustershell/groups.d $ mv cluster.yaml.example cluster.yaml
You can make the first dictionary found on this file (named roles) to be the
default source by changing
default: local to
default: roles in
/etc/clustershell/groups.conf (main config file for groups).
For more info about the YAML group files, please see File-based group sources.
Please also see node groups configuration for node groups configuration in general.
nodeset -L/--list-all option¶
Additionally, the nodeset command also has a new option
--list-all to list groups from all sources (
-l only lists
groups from the default source). This can be useful when configuring
ClusterShell and/or troubleshooting node group sources:
$ nodeset -LL @adm example0 @all example[2,4-5,32-159] @compute example[32-159] @gpu example[156-159] @io example[2,4-5] @racks:new example[4-5,156-159] @racks:old example[0,2,32-159] @racks:rack1 example[0,2] @racks:rack2 example[4-5] @racks:rack3 example[32-159] @racks:rack4 example[156-159] @cpu:hsw example[64-159] @cpu:ivy example[32-63]
Special group @*¶
The special group syntax
@source:* if using explicit source
selection) has been added and can be used in configuration files or with
command line tools. This special group is always available for file-based node
groups (return the content of the all group, or all groups from the source
otherwise). For external sources, it is available when either the all
upcall is defined or both map and list upcalls are defined. The all
special group is also used by
clush -a and
nodeset -a. For example,
the two following commands are equivalent:
$ nodeset -a -f example[2,4-5,32-159] $ nodeset -f @* example[2,4-5,32-159]
Version 1.7 introduces a new generic execution worker named
ExecWorker as the new base class for most exec()-based worker
classes. In practice with clush, you can now specify the worker in
command line using
-R and use exec. It also supports
special placeholders for the node (%h) or rank (%n). For example, the
following command will execute ping commands in parallel, each with a
different host from hosts cs01, etc. to cs05 as argument and then
aggregate the results:
$ clush -R exec -w cs[01-05] -bL 'ping -c1 %h >/dev/null && echo ok' cs[01-04]: ok clush: cs05: exited with exit code 1
This feature allows the system administrator to use non cluster-aware tools in
a more efficient way. You may also want to explicitly set the fanout (using
-f) to limit the number of parallel local commands launched.
Please see also clush worker selection.
Version 1.7 adds support for
rsh or any of its variants like
ssh also share a lot of common mechanisms. Worker Rsh was
added moving a lot of Worker Ssh code into it.
clush, please see clush worker selection to
rsh by default instead of
ssh at the library level, install the
provided example file named
Tree Propagation Mode¶
The ClusterShell Tree Mode allows you to send commands to target nodes through a set of predefined gateways (using ssh by default). It can be useful to access servers that are behind some other servers like bastion hosts, or to scale on very large clusters when the flat mode (eg. sliding window of ssh commands) is not enough anymore.
The tree mode is now documented, it has been improved and
is enabled by default when a
topology.conf file is found. While it is still
a work in progress, the tree mode is known to work pretty well when all gateways
are online. We'll continue to improve it and make it more robust in the next
$XDG_CONFIG_HOME are defined,
ClusterShell will use them to search for additional configuration files.
$CLUSTERSHELL_CFGDIR is not defined, the global configuration files will
be searched for in /etc/clustershell
PIP user installation support¶
ClusterShell 1.7 is now fully compatible with PIP and supports user configuration files:
$ pip install --user clustershell
Please see Installing ClusterShell as user using pip.