Repository navigation
Dashboard is wonky with bokeh 3 #7327
Description
With bokeh 3.0.2, the task stream doesn't show up, nor do the 'tasks processing' bars. (I've more commonly seen that a few tasks intermittently show up on the task stream, but most are missing. The 'tasks processing' bars always seem to be broken.)
After pip install 'bokeh<3':
In [1]: import distributed
In [2]: import dask.array as da
In [3]: client = distributed.Client()
In [4]: da.random.random(200_000, chunks=100).sum().compute()I see this with both macOS and linux running the scheduler, both Chrome and Safari as the browser.
Environment:
- Distributed version: 176a84c, or 2022.11.0
- Python version: 3.9.1
- Operating System: macOS
- Install method (conda, pip, source): source
Activity
jrbourbeau commentedon Nov 17, 2022
Thanks @gjoseph92 -- closing as a duplicate of #7173
gjoseph92 commentedon Nov 17, 2022
Thanks @jrbourbeau. Until this is fixed though, shouldn't we set an upper bound on Bokeh versions? distributed isn't compatible with bokeh 3, so the appropriate thing to do from a packaging standpoint would be to set a limit. That is, revert dask/dask#9659 and the yaml changes in #5648. Getting a broken dashboard when installing dask in a new environment isn't a great user experience.
jrbourbeau commentedon Nov 17, 2022
As you saw, previously we were setting an upper bound and then we added
bokeh=3compatibility code and removed the upper bound. You bring up a fair point that, while the compatibility code was needed, given we know there are some plot rendering issues, removing thebokeh<3was probably too premature.FWIW as
bokehis an optional dependency, we can't 100% prevent users from installingbokeh=3in all cases. For example, I thinkpip install distributed bokehwill still pull in latest releases, but IIRC there are some things we can do withcondain this situation (mayberun_constrained?). That said, we do have full control over certain installation options (i.e.pip install dask[complete],conda install dask) and a best-effort approach there would still add value.@charlesbluca do you have bandwidth / interest in reinstating those
bokeh<3constraints? No worries if notEDIT: Also, @gjoseph92 @ncclementi if either of you have bandwidth to make these changes, that would also be welcome
gjoseph92 commentedon Nov 17, 2022
Agreed, I don't think we can do anything about that. I tried experimenting with setting bounds on optional dependencies, but unsurprisingly pip doesn't care about them unless you're installing the option.
Mostly having
dask[complete]andconda install daskgive you the right thing seems like the primary concern right now. I'm happy to open some PRs adding the pin back.gjoseph92 commentedon Nov 17, 2022
bokeh=3support dask#9673bokeh=3support #7329