[jupyterlab] [esrf] [jupyter-slurm] very slow reading and broken `.h5`?

[jpcbertoldo]

Hi, I was using jupyterlab-h5web on jupyter-slurm.esrf.fr, and some files are veeeery slow, and I have one that doesn't even open. I am not sure if this is somehow expected/normal, but I recorded a few of this issues so you can check it.

See: https://youtu.be/qSM2JsVVEHY

I open 4 files in the video, respectively:

1) 0:08: an .h5 from the acquisition system at id11 path: /data/id11/3dxrd/blc12852/id11/bmg_l1/bmg_l1_bmg_dct2/bmg_l1_bmg_dct2.h5 issue: 0:38: displaying one frame is quite slow

2) 1:15: a "small" .h5 written by my code is quite slow path: /data/id11/3dxrd/blc12852/id11/analysis/bmg_l1_bmg_dct2/difspots/000000.h5 issue: isn't it slow relative to the size of the things inside?

3) 1:33: an .h5 regrouping (external links) about 3300 .h5s like the one above path: /data/id11/3dxrd/blc12852/id11/analysis/bmg_l1_bmg_dct2/difspots.h5 issue: slow to open, then very slow to expand the group with the external links

4) 2:30: an .h5 with only 2 datasets, about the same size of the first one path: /data/id11/3dxrd/blc12852/id11/bmg_l1/bmg_l1_bmg_dct2/bmg_l1_bmg_dct2.preprocessed.h5 issue: veeeeeeeery slow; I kept waiting after the end of the video for several minutes, the entire jupyterlab becomes very slow, and it ends up breaking the page (screenshot below). When I reload the page, the h5web tab (in jupyterlab) is not there anymore.

A few notes (Idk if this might have some influence):

• I am using symbolic links (ln -s) to be able to see those files in jupyter-slurm because they are in /data.
• I am writting my .h5 files with silx.io.dictdump.dicttoh5. I can show you specifically where in my code.
• I saw similar effects on a jupyterlab instances on gpid11-nice and rnice8 nodes. In these cases I launch jupyterlab on those machines, connect to them with ssh -L 8888:localhost:8888 to forward the application's port, then access it using my local browser (my personal computer, not an ESRF one).

[t20100]

There is an issue for the integration of h5web in JLab file browser which needs to be solved in jupyterlab (see #24): There is no way to avoid loading the hdf5 file in the browser!.... so next version of JLab v3.1.x will be needed (jupyter-slurm runs JLab v2).

There is no such issue when opening a file from within a notebook (within JLab):

from jupyterlab_h5web import H5Web
H5Web('<path to the HDF5 file>')

https://github.com/silx-kit/jupyterlab-h5web#in-jupyter-notebooks This should avoid at least the time for the first loading of a file.

[jpcbertoldo]

Ok, thanks!

Just to be sure: to use it in my own jupyterlab installation, so I need to install jupyterlab-h5web in the jupyterlab's environment or on the kernel's environment (I use different kernels)? Or both?

Last but not least, is JLab v2 --> JLab v3.1.X upgrade supposed to happen anytime soon? I'm just wondering so I can organize my own priorities.

[t20100]

to use it in my own jupyterlab installation, so I need to install jupyterlab-h5web in the jupyterlab's environment or on the kernel's environment (I use different kernels)? Or both?

Both: kernel side for the jupyterlab_h5web import in the notebook, jupyterlab side for enabling the feature in JLab. On JLab side, you might also need to rebuild JLab (see https://github.com/silx-kit/jupyterlab-h5web#install)

is JLab v2 --> JLab v3.1.X upgrade supposed to happen anytime soon?

Not soon, to upgrade it requires quite a bit of change in the jupyter-slurm installation first (to workaround incompatibilty of matplotlib version installed on the system and JLab 3 compatible extension), and it's not my priority for now.

[loichuder]

Hello @joaopcbertoldo and thanks for the feedback ! To add on @t20100 comments:

• Even though jupyterlab-h5web only officially supports JupyterLab 2.x, CI tests should ensure that the extension should work for JupyterLab 3.0.x. Note that so far (v0.0.9 and below), the extension does not work with JupyterLab 3.1.x (see #49).
• Even if we release a version of jupyterlab-h5web compatible with JupyterLab 3.1.x (which we will do!), this will not fix #24 straight away. There are still roadblocks on the way that I need to overcome by changing things in the extension and in core JupyterLab. That is why we put forward the opening of files in the notebook rather than in the browser.
• Now getting back to the issues you reported, I expect 4. to be due to #24 so there is not much to be done.
• The other operations seem indeed slower than what I would expect so I will investigate. I will have a go with your files to see if I get the same.

[loichuder]

As already mentioned, the root issue is that opening an h5web tab by double-clicking on a file fetched the whole file as a base64 blob. Not only this is quite inefficient but also pending requests of this type will slow down the whole "regular" fetching process.

I thought we needed to wait for a newer version of JLab to remove this blob fetching but I was able to bypass the problematic machinery in 0.0.10. Therefore, 0.0.10 removed this blob fetching.

Getting back to the issue at hand: 0.0.10 is now deployed on jupyter-slurm and I tried all the operations above. Except for the external links one (3.), I found that operations were way faster than what you presented so I would invite you to retry and tell us if you get the same impression.

[loichuder]

Latest versions (1.1.0 and 0.0.14) fix 3. and I already mentioned that other points are fixed for versions above 0.0.10.

Therefore I will now close this issue. We can reopen if needed.