The Dask distributed scheduler provides live feedback in two forms:
- An interactive dashboard containing many plots and tables with live information
- A progress bar suitable for interactive use in consoles or notebooks
If Bokeh is installed then the dashboard will start up automatically whenever the scheduler is created. For local use this happens when you create a client with no arguments:
from dask.distributed import Client client = Client() # start distributed scheduler locally. Launch dashboard
It is typically served at http://localhost:8787/status ,
but may be served elsewhere if this port is taken.
The address of the dashboard will be displayed if you are in a Jupyter Notebook,
or can be queried from
There are numerous pages with information about task runtimes, communication, statistical profiling, load balancing, memory use, and much more. For more information we recommend the video guide above.
||Connect to and submit computation to a Dask cluster|
||Collect task stream within a context block|
||Collect statistical profiling information about recent work|
||Gather performance report|
You can capture some of the same information that the dashboard presents for
offline processing using the
functions. These capture the start and stop time of every task and transfer,
as well as the results of a statistical profiler.
with get_task_stream(plot='save', filename="task-stream.html") as ts: x.compute() client.profile(filename="dask-profile.html") history = ts.data
Additionally, Dask can save many diagnostics dashboards at once including the
task stream, worker profiles, bandwidths, etc. with the
from dask.distributed import performance_report with performance_report(filename="dask-report.html"): ## some dask computation
The following video demonstrates the
performance_report context manager in greater
||Track progress of futures|
dask.distributed progress bar differs from the
ProgressBar used for
progress function takes a Dask object that is executing in the background:
# Single machine progress bar from dask.diagnostics import ProgressBar with ProgressBar(): x.compute() # Distributed scheduler ProgressBar from dask.distributed import Client, progress client = Client() # use dask.distributed by default x = x.persist() # start computation in the background progress(x) # watch progress x.compute() # convert to final result when done if desired
Connecting to the Dashboard¶
Some computer networks may restrict access to certain ports or only allow access from certain machines. If you are unable to access the dashboard then you may want to contact your IT administrator.
Some common problems and solutions follow:
Specify an accessible port¶
Some clusters restrict the ports that are visible to the outside world. These
ports may include the default port for the web interface,
8787. There are
a few ways to handle this:
- Open port
8787to the outside world. Often this involves asking your cluster administrator.
- Use a different port that is publicly accessible using the
--dashboard-address :8787option on the
- Use fancier techniques, like Port Forwarding
If you have SSH access then one way to gain access to a blocked port is through SSH port forwarding. A typical use case looks like the following:
local$ ssh -L 8000:localhost:8787 user@remote remote$ dask-scheduler # now, the web UI is visible at localhost:8000 remote$ # continue to set up dask if needed -- add workers, etc
It is then possible to go to
localhost:8000 and see Dask Web UI. This same approach is
not specific to dask.distributed, but can be used by any service that operates over a
network, such as Jupyter notebooks. For example, if we chose to do this we could
forward port 8888 (the default Jupyter port) to port 8001 with
ssh -L 8001:localhost:8888 user@remote.
progress(*futures, notebook=None, multi=True, complete=True, **kwargs)¶
Track progress of futures
This operates differently in the notebook and the console
- Notebook: This returns immediately, leaving an IPython widget on screen
- Console: This blocks until the computation completes
- futures: Futures
A list of futures or keys to track
- notebook: bool (optional)
Running in the notebook or not (defaults to guess)
- multi: bool (optional)
Track different functions independently (defaults to True)
- complete: bool (optional)
Track all keys (True) or only keys that have not yet run (False) (defaults to True)
In the notebook, the output of progress must be the last statement in the cell. Typically, this means calling progress at the end of a cell.
>>> progress(futures) # doctest: +SKIP [########################################] | 100% Completed | 1.7s
get_task_stream(client=None, plot=False, filename='task-stream.html')¶
Collect task stream within a context block
This provides diagnostic information about every task that was run during the time when this block was active.
This must be used as a context manager.
- plot: boolean, str
If true then also return a Bokeh figure If plot == ‘save’ then save the figure to a file
- filename: str (optional)
The filename to save to if you set
- Function version of this context manager
>>> with get_task_stream() as ts: ... x.compute() >>> ts.data [...]
Get back a Bokeh figure and optionally save to a file
>>> with get_task_stream(plot='save', filename='task-stream.html') as ts: ... x.compute() >>> ts.figure <Bokeh Figure>
To share this file with others you may wish to upload and serve it online. A common way to do this is to upload the file as a gist, and then serve it on https://raw.githack.com
$ python -m pip install gist $ gist task-stream.html https://gist.github.com/8a5b3c74b10b413f612bb5e250856ceb
You can then navigate to that site, click the “Raw” button to the right of the
task-stream.htmlfile, and then provide that URL to https://raw.githack.com . This process should provide a sharable link that others can use to see your task stream plot.