You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Dask dashboards don't work from Dask Gateway Helm release 2022.11.0 possibly due to bokeh version #7512
This has been reported at dask/dask-gateway#676, however it seems from consideRatio's reply this may need reporting here (I haven't yet managed to resolve).
Describe the issue:
Since upgrading our Dask Gateway version (deployed with Helm on K8s) to >2022.10.0, the dashboards don't seem to be working.
The error must be related to the Python version (and libraries) running in the Dask scheduler pod, since I've experimented with different client environments and different Dask Gateway deployments; holding all things constant client-side, and upgrading DG from Helm release 2022.10.0 to 2022.11.0, the issue manifests.
When trying to view the dashboard URL from 2022.11.0 onwards, an error message in the browser states: "Dask needs bokeh >= 2.1.1, < 3 for the dashboard."
If I manually jump into the spawned dask-scheduler pod, I can verify that this is likely the cause; using pip3 freeze | grep bokeh in that pod, the versions corresponding to each DG Helm release are:
DG Helm version
scheduler bokeh version
dashboards working?
2022.10.0
bokeh==2.4.3
true
2022.11.0
bokeh==3.0.1
false
2023.01.0
bokeh==3.0.3
false
i.e. as soon as the scheduler version exceeds v3, the dashboards seemingly break.
I tried to manually reinstall an older version of bokeh in the scheduler pod, to no avail.
Minimal Complete Verifiable Example:
Client side code I've been using to get the dashboard URL:
# List Gateway clusters:fromdask_gatewayimportGatewayClustercluster=GatewayCluster('http://<ip_address>/dask-gateway/', auth="jupyterhub")
cluster.scale(3)
client=cluster.get_client()
client
client widget yields a dashboard URL like: http://<ip_address>/dask-gateway/clusters/my-dask-gateway.ea1...246/status
which is what I've been using to test (i.e. not the Dask Jupyter extension)
I've worked around this issue for now (described at: dask/dask-gateway#676 (comment)) - I suspect that as the bokeh version is specified in the requirements files under the Dask Gateway repo it is a Dask Gateway issue rather than a Distributed one, but I could be wrong.
Activity
JColl88 commentedon Jan 30, 2023
I've worked around this issue for now (described at: dask/dask-gateway#676 (comment)) - I suspect that as the bokeh version is specified in the
requirementsfiles under the Dask Gateway repo it is a Dask Gateway issue rather than a Distributed one, but I could be wrong.