The slowest part of large images scaling in production is their retrieval from Swift, which could be much faster for bucketed images.
On Thu, May 1, 2014 at 7:57 AM, Gilles Dubuc gilles@wikimedia.org wrote:
An extremely crude benchmark on our multimedia labs instance, still using the same test image:
original -> 3002 (original size) 0m0.268s original -> 2048 0m1.344s original -> 1024 0m0.856s original -> 512 0m0.740s original -> 256 0m0.660s 2048 -> 1024 0m0.444s 2048 -> 512 0m0.332s 2048 -> 256 0m0.284s 1024 -> 512 0m0.112s 512 -> 256 0m0.040s
Which confirms that chaining instead of generating all thumbnails based on the biggest bucket saves a significant amount of processing time. It's definitely in the same order or magnitude as the savings achieved by going from original as the source to the biggest bucket as the source.
It's also worth noting that generating the thumbnail of the same size as the original is relatively cheap. Using it as the source for the 2048 image doesn't save that much time, though: 0m1.252s (for 3002 -> 2048).
And here's a side-by-side comparison of these images generated with chaining and images that come from our regular image scalers: https://dl.dropboxusercontent.com/u/109867/imagickchaining/index.html Try to guess which is which before inspecting the page for the answer :)
On Thu, May 1, 2014 at 4:02 PM, Gilles Dubuc gilles@wikimedia.org wrote:
Another point about picking the "one true bucket list": currently Media Viewer's buckets have been picked based on the most common screen resolutions, because Media Viewer tries to always use the entire width of the screen to display the image, so trying to achieve a 1-to-1 pixel correspondence makes sense, because it should give the sharpest result possible to the average user.
However, sticking to that approach will likely introduce a cost. As I've just mentioned, we will probably need to generate more than one of the
high
buckets based on the original, in order to avoid resizing artifacts.
On the other hand, we could decide that the unified bucket list shouldn't be based on screen resolutions (after all the full width display scenario experienced in Media Viewer might be the exception, and the buckets will
be
for everything mediawiki) and instead would progress by powers of 2. Then creating a given bucket could always be done without resizing artifacts, based on the bucket above the current one. This should provide the
biggest
savings possible in image scaling time to generate thumbnail buckets.
To illustrate with an example, the bucket list could be: 256, 512, 1024, 2048, 4096. The 4096 bucket would be generated first, based on the original, then 2048 would be generated based on 4096, then 1024 based on 2048, etc.
The big downside is that there's less progression in the 1000-3000 range (4 buckets in the Media Viewer resolution-oriented strategy, 2 buckets here) where the majority of devices currently are. If I take a test image as an example ( https://commons.wikimedia.org/wiki/File:Swallow_flying_drinking.jpg),
the
file size progression is quite different between the screen resolution buckets and the geometric (powers of 2) buckets:
- screen resolution buckets
320 11.7kb 640 17kb 800 37.9kb 1024 58kb 1280 89.5kb 1920 218.9kb 2560 324.6kb 2880 421.5kb
- geometric buckets
256 9.4kb 512 20kb 1024 58kb 2048 253.1kb 4096 test image is smaller than 4096
It seems like it's not ideal that a screen resolution slightly above 1024 would suddenly need to download an image 5 times as heavy, for not that many extra pixels on the actual screen. A similar thing can be said for
the
screen resolution progression, where the file size more than doubles between 1280 and 1920. We could probably use at least an extra step
between
those two if we use screen resolution buckets, like 1366 and/or 1440.
I think that the issue of buckets between 1000 and 3000 is tricky, it's going to be difficult to avoid generating them based on the original
while
not getting visual artifacts.
Maybe we can get away with generating 1280 (and possibly 1366, 1440)
based
on 2048, the distance between the two guaranteeing that the quality
issues
will be negligible. We definitely can't generate a 1920 based on a 2048 thumbnail, though, otherwise artifacts on thin lines will look awful.
A mixed progression like this might be the best of both worlds, if we confirm that between 1024 and 2048 the resizing is artifact-free enough:
256, 512, 1024, 1280, 1366, 1440, 2048, 4096 where 2048 is generated
based
on 4096, 1024, 1280, 1366, 1440 are generated based on 2048, 512 based on 1024, 256 based on 512.
If for example the image width is between 1440 and 2048, then 1024, 1280, 1366, 1440 would be generated based on the original. That's fine performance-wise, since the original is small.
Something that might also be useful to generate is a thumbnail of the
same
size as the original if original < 4096 (or whatever the highest bucket is). Currently we seem to block generating such a thumbnail, but the difference in file size is huge. For the test image mentioned above,
which
is 3002 pixels wide, the original is 3.69MB, while a thumbnail of the
same
size would be 465kb. For the benefit of retina displays that are
2560/2880,
displaying a thumbnail of the same size as a 3002 original would
definitely
be better than the highest available bucket (2048).
All of this is benchmark-worthy anyway, I might be splitting hair looking for powers of two if rendering a bucket chain (each bucket generated
based
on the next one) isn't that much faster than generating all buckets based on the biggest bucket.
On Thu, May 1, 2014 at 12:54 PM, Gilles Dubuc <gilles@wikimedia.org wrote:
_don't consider the upload complete_ until those are done! a web
uploader
or API-using bot should probably wait until it's done before uploading
the
next file, for instance...
You got me a little confused at that point, are you talking about the client generating the intermediary sizes, or the server?
I think client-side thumbnail generation is risky when things start getting corrupt. A client-side bug could result in a user uploading thumbnails that are for a different image. And if you want to run a
visual
signature check on the server-side to avoid that issue, you might be looking at similar processing time checking that the thumbnail is for
the
correct image then if the server was to generate the actual thumbnail.
It
would be worth researching if there's a very fast "is this thumbnail a smaller version of that image" algorithm out there. We don't need 100% confidence either, if we're looking to avoid shuffling bugs in a given upload batch.
Regarding the issue of a single intermediary size versus multiple, there's still a near-future plan to have pregenerated buckets for Media Viewer (which can be reused for a whole host of other things). Those
could
be used like mip-maps like you describe. Since these sizes will be generated at upload time, why not use them?
However quality starts to introduce noticeable visual artifacts when the bucket (source image)'s dimensions are too close to the thumbnail you
want
to render.
Consider the existing Media Viewer width buckets: 320, 640, 800, 1024, 1280, 1920, 2560, 2880
I think that generating the 300px thumbnail based on the 320 bucket is likely to introduce very visible artifacts with thin lines, etc.
compared
to using the biggest bucket (2880px). Maybe there's a smart compromise, like picking the higher bucket (eg. 300px thumbnail would use the 640 bucket as its source, etc.). I think that we need a battery of visual
test
to determine what's the best strategy here.
All of this is dependent on Ops giving the green light for pregenerating the buckets, though. The swift capacity for it is slowing being brought online, but I think Ops' prerequisite wish to saying yes to it is that
we
focus on the post-swift strategy for thumbnails. We also need to figure
out
the performance impact of generating all these thumbnails on upload. On
a
very meta note, we might generate the smaller buckets based on the
biggest
bucket and only the 2-3 biggest buckets based on the original (still to avoid visual artifacts).
Another related angle I'd like to explore is to submit a simplified version of this RFC:
https://www.mediawiki.org/wiki/Requests_for_comment/Standardized_thumbnails_... propose a single bucket list option instead of multiple
(presumably, the Media Viewer ones, if not, we'd update Media Viewer to
use
the new canon list of buckets). And where we would still allow arbitrary thumbnail sizes below a certain limit. For example, people would still
be
allowed to request thumbnails that are smaller than 800px at any size
they
want, because these are likely to be thumbnails in the real sense of the term, and for anything above 800px they would be limited to the
available
buckets (eg. 1024, 1280, 1920, 2560, 2880). This would still allow foundation-hosted wikis to have flexible layout strategies with their thumbnail sizes, while reducing the craziness of this attack vector on
the
image Scalers and gigantic waste of disk and memory space on the
thumbnail
hosting. I think it would be an easier sell for the community, the
current
RFC is too extreme in banning all arbitrary sizes and offers too many bucketing options. I feel like the standardization of true thumbnail
sizes
(small images, <800px) is much more subject to endless debate with no consensus.
On Thu, May 1, 2014 at 12:21 PM, Erwin Dokter erwin@darcoury.nl
wrote:
On 04/30/2014 12:51 PM, Brion Vibber wrote:
- at upload time, perform a series of scales to produce the mipmap
levels
- _don't consider the upload complete_ until those are done! a web
uploader or API-using bot should probably wait until it's done before uploading the next file, for instance...
- once upload is complete, keep on making user-facing thumbnails as
before... but make them from the smaller mipmap levels instead of the full-scale original
Would it not suffice to just produce *one* scaled down version (ie. 2048px) which the real-time scaler can use to produce the thumbs?
Regards,
Erwin Dokter
Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l