16 bit images were not returning the correct dominant color due truncation
The routine expected an 8bit color eg: #FFAA00, but ended up getting a 16bit one eg: #FFFAAA000. This caused a truncation, which leads to wildly off colors.
We've had the UploadReference table for some time now in core,
but it was added after ChatUpload was and chat was just never
moved over to this new system.
This commit changes all chat code dealing with uploads to create/
update/delete/query UploadReference records instead of ChatUpload
records for consistency. At a later date we will drop the ChatUpload
table, but for now keeping it for data backup.
The migration + post migration are the same, we need both in case
any chat uploads are added/removed during deploy.
Previously we would unconditionally fetch all images via HTTP to grab
original sizing from cooked post processor in 2 different spots.
This was wasteful as we already calculate and cache this info in upload records.
This also simplifies some specs and reduces use of mocks.
We previously had a system which would generate a 10x10px preview of images and add their URLs in a data-small-upload attribute. The client would then use that as the background-image of the `<img>` element. This works reasonably well on fast connections, but on slower connections it can take a few seconds for the placeholders to appear. The act of loading the placeholders can also break or delay the loading of the 'real' images.
This commit replaces the placeholder logic with a new approach. Instead of a 10x10px preview, we use imagemagick to calculate the average color of an image and store it in the database. The hex color value then added as a `data-dominant-color` attribute on the `<img>` element, and the client can use this as a `background-color` on the element while the real image is loading. That means no extra HTTP request is required, and so the placeholder color can appear instantly.
Dominant color will be calculated:
1. When a new upload is created
2. During a post rebake, if the dominant color is missing from an upload, it will be calculated and stored
3. Every 15 minutes, 25 old upload records are fetched and their dominant color calculated and stored. (part of the existing PeriodicalUpdates job)
Existing posts will continue to use the old 10x10px placeholder system until they are next rebaked
Discourse shouldn't dynamically calculate the path of uploads and optimized images after a file has been stored on disk or S3. Otherwise it might calculate the wrong path if the SHA1 or extension stored in the database doesn't match the actual file path.
It's possible that the original topic image is broken in some form, so
we shouldn't try and generate a topic thumbnail for it. The fix will
prevent the generate_topic_thumbnails job being enqueued every time the
topic is viewed.
Fixed bugs, added specs, extracted the upload downsizing code to a class, added support for non-S3 setups, changed it so that images aren't downloaded twice.
This code has been tested on production and successfully resized ~180k uploads.
Includes:
* DEV: Extract upload downsizing logic
* DEV: Add support for non-S3 uploads
* DEV: Process only images uploaded by users
* FIX: Incorrect usage of `count` and `exist?` typo
* DEV: Spec S3 image downsizing
* DEV: Avoid downloading images twice
* DEV: Update filesizes earlier in the process
* DEV: Return false on invalid upload
* FIX: Download images that currently above the limit (If the image size limit is decreased, then there was no way to resize those images that now fall outside the allowed size range)
* Update script/downsize_uploads.rb (Co-authored-by: Régis Hanol <regis@hanol.fr>)
* FIX: Perform crop using user-specified image sizes
It used to resize the images to max width and height first and then
perform the crop operation. This is wrong because it ignored the user
specified image sizes from the Markdown.
* DEV: Use real images in test
Basically, say you had already downloaded a certain image from a certain URL
using pull_hotlinked_images and the onebox. The upload would be stored
by its sha as an upload record. Whenever you linked to the same URL again
in a post (e.g. in our case an og:image on review.discourse) we would
would reuse the original upload record because of the sha1.
However when you turned on secure media this could cause problems as
the first post that uses that upload after secure media is enabled
will set the access control post for the upload to the new post.
Then if the post is deleted every single onebox/link to that same image
URL will fail forever with 403 as the secure-media-uploads URL fails
if the access control post has been deleted.
To fix this when cooking posts and pulling hotlinked images, we only
allow using an original upload by URL if its access control post
matches the current post, and if the original_sha1 is filled in,
meaning it was uploaded AFTER secure media was enabled. otherwise
we just redownload the media again to be safe, as the URL will always
be new then.
### General Changes and Duplication
* We now consider a post `with_secure_media?` if it is in a read-restricted category.
* When uploading we now set an upload's secure status straight away.
* When uploading if `SiteSetting.secure_media` is enabled, we do not check to see if the upload already exists using the `sha1` digest of the upload. The `sha1` column of the upload is filled with a `SecureRandom.hex(20)` value which is the same length as `Upload::SHA1_LENGTH`. The `original_sha1` column is filled with the _real_ sha1 digest of the file.
* Whether an upload `should_be_secure?` is now determined by whether the `access_control_post` is `with_secure_media?` (if there is no access control post then we leave the secure status as is).
* When serializing the upload, we now cook the URL if the upload is secure. This is so it shows up correctly in the composer preview, because we set secure status on upload.
### Viewing Secure Media
* The secure-media-upload URL will take the post that the upload is attached to into account via `Guardian.can_see?` for access permissions
* If there is no `access_control_post` then we just deliver the media. This should be a rare occurrance and shouldn't cause issues as the `access_control_post` is set when `link_post_uploads` is called via `CookedPostProcessor`
### Removed
We no longer do any of these because we do not reuse uploads by sha1 if secure media is enabled.
* We no longer have a way to prevent cross-posting of a secure upload from a private context to a public context.
* We no longer have to set `secure: false` for uploads when uploading for a theme component.
This change both speeds up specs (less strings to allocate) and helps catch
cases where methods in Discourse are mutating inputs.
Overall we will be migrating everything to use #frozen_string_literal: true
it will take a while, but this is the first and safest move in this direction
This reduces the number of scans that the db has to do in the query
to fetch orphan uploads. Futheremore, we were not batching our
records which bloats memory.
