Changes in IPython Parallel
A tiny release fixing issues seen building notebooks with jupyter-book.
Fix KeyError on parent_header when streaming output with %%px
Allow disabling streaming/progress defaults with IPP_NONINTERACTIVE=1 environment variable (e.g. when building notebooks in documentation)
A tiny bugfix release
Fix error preventing creation of new profiles via the lab extension
A small bugfix and compatibility release.
Updates dependencies in jupyterlab extension to jupyterlab 3.6
fix ResourceWarnings about closed clusters
Avoid some deprecated APIs in jupyter-client and pyzmq
8.4.1 is a tiny release, adding support for Python 3.11
8.4.0 is a small release, with some bugfixes and improvements to the release process.
%px) only skip redisplay of streamed errors if outputs are complete
Avoid use of recently deprecated asyncio/tornado APIs around ‘current’ event loops that are not running.
Switch to hatch backend for packaging
8.3.0 is a small release, with some bugfixes and improvements to the release process.
Workaround SSL issues with recent builds of nodejs + webpack
Build with flit, removing setup.py
Remove remaining references to deprecated
distutilspackage (has surprising impact on process memory)
Improve logging when engine registration times out
Maintenance changes that shouldn’t affect users:
Releases are now built with pip instead of
Updates to autoformatting configuration
Contributors to this release#
8.2.1 Fixes some compatibility issues with latest dask, ipykernel, and setuptools, as well as some typos and improved documentation.
8.2.0 is a small release, mostly of small bugfixes and improvements.
len(AsyncMapResult) and progress ports now use the number of items in the map,
not the number of messages.
Show output prior to errors in
Fix cases where engine id could be
pbsto engine launcher entrypoints
8.1.0 is a small release, adding a few new features and bugfixes.
relay KeyboardInterrupt to engines in blocking
Cluster.start_and_connect(activate=True)to include activation of
%pxmagics in one-liner startup.
initial support for Clusters tab in RetroLab
ensure profile config is always loaded for
build lab extension in production mode, apply trove classifiers
pass through keyword arguments to constructor in
This is marked as a major revision because of the change to pass connection information via environment variables.
BatchSystem launchers with a custom template will need to make sure to set flags that inherit environment variables,
#PBS -V or
Cluster(engines="mpi")signature for setting the engine (or controller) launcher class.
The first (and usually only) engine set can be accessed as
Cluster.engine_set, rather than digging through the
environmentconfiguration to all Launchers.
Support more configuration via environment variables, including passing connection info to engines via
$IPP_CONNECTION_INFO, which is used by default, avoiding the need to send connection files to engines in cases of non-shared filesystems.
Launchers send connection info to engines via
$IPP_CONNECTION_INFOby default. This is governed by
Cluster.send_engines_connection_env, which is True by default.
EngineLauncher.get_output()via output files in batch system launchers
Capture output in Batch launchers by setting output file options in the default templates.
LazyMapIteratorwhich has a
.cancel()method, for stopping consumption of the map input.
LoadBalancedView.imap(max_outstanding=n)()limits the number of tasks submitted to the cluster, instead of limiting the number not-yet-consumed. Prior to this, the cluster could be idle if several results were waiting to be consumed.
output streamed by
%%pxincludes errors and results, for immediate feedback when only one engine fails.
Various bugs preventing use of non-default Controller launchers
Fixed crash in jupyterlab extension when IPython directory does not exist
imapresults, like Executors in the standard library
Removed spurious jupyterlab plugin options that had no effect.
%autopxstreams output just like
Add BroadcastView benchmark code
Tag releases with tbump
Client.start_and_connect()method for starting a cluster and returning a connected client in one call.
view.executorfor better consistency with standard library Executors.
Client.wait_for_engines()will raise an informative error if the parent Cluster object notices that its engines have halted while waiting, or any engine unregisters, rather than continuing to wait for engines that will never come
Show progress if
%pxis taking significant time
Improved support for streaming output, e.g. with
%px, including support for updating output in-place with standard terminal carriage-return progress bars.
Fix dropped IOPub messages when using large numbers of engines, causing
Fix error waiting for connection files when controller is started over ssh.
Fix missing setupbase.py in tarball
Require Python 3.6
Fix compatibility issues with ipykernel 6 and jupyter-client 7
Remove dependency on deprecated ipython-genutils
New dependencies on psutil, entrypoints, tqdm
ipcluster cleancommands derived from the Cluster API.
Client.send_signal()for sending signals to single engines.
New KernelNanny process for signaling and monitoring engines for improved responsiveness of handing engine crashes.
BroadcastSchedulerwith vastly improved scaling in ‘do-on-all’ operations on large numbers of engines, c/o Tom-Olav Bøyum’s Master’s thesis at University of Oslo. Broadcast view documentation.
Client.wait_for_engines()method to wait for engines to be available.
Nicer progress bars for interactive waits, such as
AsyncResult.stream_output()context manager for streaming output. Stream output by default in parallel magics.
Launchers registered via entrypoints for better support of third-party Launchers.
New JupyterLab extension (enabled by default) based on dask-labextension for managing clusters.
LoadBalancedView.imap()consumes inputs as-needed, producing a generator of results instead of an AsyncMapResult, allowing for consumption of very large or infinite mapping inputs.
Improvements and other fixes:
Greatly improved performance of heartbeat and registration with large numbers of engines, tested with 5000 engines and default configuration.
IPController.portsconfiguration to specify the pool of ports for the controller to use, e.g.
ipcontroller --ports 10101-10120.
fas keyword-argument to
joblib backend will start and stop a cluster by default if the default cluster is not running.
The repo has been updated to use pre-commit, black, myst, and friends and GitHub Actions for CI, but this should not affect users, only making it a bit nicer for contributors.
Require Python 3.5
Fix compatibility with joblib 0.14
Fix crash recovery test for Python 3.8
Fix repeated name when cluster-id is set
Fix CSS for notebook extension
Fix KeyError handling heartbeat failures
Fix compatibility with Python 3.8
Fix compatibility with recent dask
Improve compatibility with ipykernel 5
%autopxwith IPython 7
Fix non-local ip warning when using current hostname
Fix compatibility for execute requests with ipykernel 5 (now require ipykernel >= 4.4)
Fix compatibility with tornado 4, broken in 6.2.0
Fix encoding of engine and controller logs in
ipcluster --debugon Python 3
Fix compatiblity with joblib 0.12
Include LICENSE file in wheels
Workaround a setuptools issue preventing installation from sdist on Windows
Drop support for Python 3.3. IPython parallel now requires Python 2.7 or >= 3.4.
Further fixes for compatibility with tornado 5 when run with asyncio (Python 3)
Fix for enabling clusters tab via nbextension
Multiple fixes for handling when engines stop unexpectedly
Installing IPython Parallel enables the Clusters tab extension by default, without any additional commands.
Fix regression in 6.1.0 preventing BatchSpawners (PBS, etc.) from launching with ipcluster.
Compatibility fixes with related packages:
Fix compatibility with pyzmq 17 and tornado 5.
Fix compatibility with IPython ≥ 6.
Improve compatibility with dask.distributed ≥ 1.18.
namespaceto BatchSpawners for easier extensibility.
Support serializing partial functions.
Support hostnames for machine location, not just ip addresses.
--locationargument to ipcluster for setting the controller location. It can be a hostname or ip.
Engine rank matches MPI rank if engines are started with
Avoid duplicate pickling of the same object in maps, etc.
Documentation has been improved significantly.
Upload fixed sdist for 6.0.1.
Small encoding fix for Python 2.
Due to a compatibility change and semver, this is a major release. However, it is not a big release. The main compatibility change is that all timestamps are now timezone-aware UTC timestamps. This means you may see comparison errors if you have code that uses datetime objects without timezone info (so-called naïve datetime objects).
become_distributed()remains as an alias.
import joblib from a public API instead of a private one when using IPython Parallel as a joblib backend.
Compatibility fix in extensions for security changes in notebook 4.3
Fix compatibility with changes in ipykernel 4.3, 4.4
Improve inspection of
Client.wait()accepts any Future.
--userflag to ipcluster nbextension
Default to one core per worker in
Client.become_distributed(). Override by specifying
Subprocess logs are no longer sent to files by default in ipcluster.
To turn an IPython cluster into a dask.distributed cluster,
executor = client.become_distributed(ncores=1)
which returns a distributed
To register IPython Parallel as the backend for joblib:
import ipyparallel as ipp ipp.register_joblib_backend()
IPython parallel now supports the notebook-4.2 API for enabling server extensions, to provide the IPython clusters tab:
jupyter serverextension enable --py ipyparallel jupyter nbextension install --py ipyparallel jupyter nbextension enable --py ipyparallel
though you can still use the more convenient single-call:
ipcluster nbextension enable
which does all three steps above.
Slurm support is added to ipcluster.
Fix imports in
Various typos and documentation updates to catch up with 5.0.
The highlight of ipyparallel 5.0 is that the Client has been reorganized a bit to use Futures.
AsyncResults are now a Future subclass, so they can be
yield ed in coroutines, etc.
Views have also received an Executor interface.
This rewrite better connects results to their handles,
so the Client.results cache should no longer grow unbounded.
Part of the Future refactor is that Client IO is now handled in a background thread,
which means that
Client.spin_thread() is obsolete and deprecated.
Add ipcluster nbextension enable|disable to toggle the clusters tab in Jupyter notebook
Less interesting development changes for users:
Some IPython-parallel extensions to the IPython kernel have been moved to the ipyparallel package:
ipykernel Python serialization is now in
apply_request message handling is implememented in a Kernel subclass, rather than the base ipykernel Kernel.
Improvements for specifying engines with SSH launcher.
First release of
ipyparallel as a standalone package.