ZIM for ab_all_maxi has different sizes between 1.13 and 1.14

[audiodude]

The ZIM scraped in July 2024 by 1.14 has a different size than the one scraped in June 2024 by 1.13:

wikipedia_ab_all_maxi_2024-06.zim             2024-06-17 02:21   26M   
wikipedia_ab_all_maxi_2024-07.zim             2024-07-22 20:56   36M

Oddly, this is the opposite problem as #2070. We don't yet know what the issue might be.

[audiodude]

ZIM file links:

wikipedia_ab_all_maxi_2024-06.zim

wikipedia_ab_all_maxi_2024-07.zim

[audiodude]

Here is a .tsv for every entry in the ZIMs. It has the format:

path,june size, july size

comparison.zip

Doing some analysis in pandas, we see that there are 681 webps that are larger in July, out of 2969 total webps:

The mean size difference is +10,661 bytes for those webps that are larger.

However, the total difference, including webps that are smaller in July, is only 6.77 MB:

[audiodude]

Clearly I'm doing something wrong, because the total sums of July sizes and June sizes are only 107 MB and 110 MB:

I tried to iterate over all entries in the ZIM using the _get_entry_by_id hack:

import csv

from libzim.reader import Archive

june = Archive("zims/wikipedia_ab_all_maxi_2024-06.zim")
july = Archive("zims/wikipedia_ab_all_maxi_2024-07.zim")

path_to_sizes = {}

for i in range(0, june.all_entry_count):
  entry = june._get_entry_by_id(i)
  path_to_sizes[entry.path] = [entry.get_item().size]

for i in range(0, july.all_entry_count):
  entry = july._get_entry_by_id(i)
  if entry.path in path_to_sizes:
    path_to_sizes[entry.path].append(entry.get_item().size)
  else:
    path_to_sizes[entry.path] = [None, entry.get_item().size]

for entry, sizes in path_to_sizes.items():
  if len(sizes) == 1:
    sizes.append(None)

with open('comparison.tsv', 'w', newline='') as csvfile:
  csvwriter = csv.writer(csvfile, delimiter='\t')
  sorted_keys = sorted(path_to_sizes.keys())
  csvwriter.writerow(('path', 'june', 'july'))
  for key in sorted_keys:
    csvwriter.writerow((key, *path_to_sizes[key]))

[rgaudin]

Your tsv is not filtered on WEBP files ; it contains all entries, including compressed ones (text) and indexes.

[rgaudin]

sum(june._get_entry_by_id(i).get_item().size for i in range(0, june.all_entry_count) if june._get_entry_by_id(i).get_item().mimetype == "image/webp")
> 13594704  # 12.96 MiB

[rgaudin]

July WEBP are 24129838 / 23.01 MiB

[audiodude]

Your tsv is not filtered on WEBP files ; it contains all entries, including compressed ones (text) and indexes.

The pandas code limits it to webp:

webps = df[df['path'].str.endswith('.webp')]

[audiodude]

The larger question is why is my total size 118 MB?

Edit: I misread the original ZIM sizes as 26/36 GB instead of MB. So actually, uncompressed, 110/118 MB makes sense.

[audiodude]

Here is my Jupyter notebook with analysis: https://github.com/audiodude/zim-investigation/blob/main/compare.ipynb

[audiodude]

Doing a more "apples to apples" comparison of the wiki scraped right now with 1.13 versus 1.14, the discrepancy is much less:

30244	zims/wikipedia_ab_all_maxi_2024-08.113.zim
34728	zims/wikipedia_ab_all_maxi_2024-08.114.zim

[audiodude]

Dumping some of the webps from the respective ZIMs, we see that the 1.14 ones are much bigger:

$ du zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/*
16	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/1924WOlympicPoster.jpg.webp
16	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/200412_-_Plaqueminier_et_ses_kakis.jpg.webp
12	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Ambara_church_ruins_in_Abkhazia%2C_1899.jpg.webp
20	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Carmen_habanera_original.jpg.webp
20	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Carmen_-_illustration_by_Luc_for_Journal_Amusant_1911.jpg.webp
8	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Christos_Acheiropoietos.jpg.webp
28	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Hovenia_dulcis.jpg.webp
12	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Lashkendar_temple_ruins.JPG.webp
28	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Paliurus_fg01.jpg.webp
20	zim_file_dump/wikipedia_ab_all_maxi_2024-08.113.zim/Tsebelda_iconostasis.jpg.webp

$ du zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/*
112	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/1924WOlympicPoster.jpg.webp
16	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/200412_-_Plaqueminier_et_ses_kakis.jpg.webp
120	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Ambara_church_ruins_in_Abkhazia%2C_1899.jpg.webp
112	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Carmen_habanera_original.jpg.webp
144	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Carmen_-_illustration_by_Luc_for_Journal_Amusant_1911.jpg.webp
112	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Christos_Acheiropoietos.jpg.webp
28	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Hovenia_dulcis.jpg.webp
168	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Lashkendar_temple_ruins.JPG.webp
28	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Paliurus_fg01.jpg.webp
140	zim_file_dump/wikipedia_ab_all_maxi_2024-08.114.zim/Tsebelda_iconostasis.jpg.webp

Confirmed manually that the 1.14 images have much bigger dimensions.

[audiodude]

Confirmed this is due to larger images.

[kelson42]

@audiodude What does that mean concretly it term of resolution and quality? Are they all impacted in the same manner?

[audiodude]

@kelson42 The resolutions are much bigger, they have larger widths and heights in terms of pixels. I didn't do any systematic analysis of to what degree that is the case, but it's most likely due to #1925

[kelson42]

@audiodude I'm not against to downscale images but:

• To do that properly we need to understand properly hownthis should be done. We should probably find the piece of code of the MCS responsible of that and copy it
• We shoukd clarify if - on the top - there is not a problem with the quality of the webp pictures we do (what we suspect)

[Jaifroid]

Doing a more "apples to apples" comparison of the wiki scraped right now with 1.13 versus 1.14, the discrepancy is much less:

30244	zims/wikipedia_ab_all_maxi_2024-08.113.zim
34728	zims/wikipedia_ab_all_maxi_2024-08.114.zim

I noticed this with https://download.kiwix.org/zim/wikivoyage/wikivoyage_en_all_maxi_2024-08.zim, which is scraped with 1.13 from new endpoint, has larger images at least in terms of display dimensions, but which hardly increases the ZIM size compared to ZIMs scraped from the old endpoint.

I actually rather like the larger display size for images at least in that Wikivoyage version (which I've just released as a packaged app). If we could hit that sweet-spot in terms of display-size vs compression, it would be a good solution IMHO. What is 1.13 doing right here?