Actions like deleting realms may leave unreferenced uploads in the
attachment storage backend.
Fix these by walking the complete contents of the attachment storage
backend, and removing files which are no longer present in the
database. This may take quite some time, as it is necessarily O(n) in
the number of files uploaded to the system.
The Content-Type of user-provided uploads was provided by the browser
at initial upload time, and stored in S3; however, 04cf68b45e
switched to determining the Content-Disposition merely from the
filename. This makes uploads vulnerable to a stored XSS, wherein a
file uploaded with a content-type of `text/html` and an extension of
`.png` would be served to browsers as `Content-Disposition: inline`,
which is unsafe.
The `Content-Security-Policy` headers in the previous commit mitigate
this, but only for browsers which support them.
Revert parts of 04cf68b45e, specifically by allowing S3 to provide
the Content-Disposition header, and using the
`ResponseContentDisposition` argument when necessary to override it to
`attachment`. Because we expect S3 responses to vary based on this
argument, we include it in the cache key; since the query parameter
has dashes in it, we can't use use the helper `$arg_` variables, and
must parse it from the query parameters manually.
Adding the disposition may decrease the cache hit rate somewhat, but
downloads are infrequent enough that it is unlikely to have a
noticeable effect. We take care to not adjust the cache key for
requests which do not specify the disposition.
When file uploads are stored in S3, this means that Zulip serves as a
302 to S3. Because browsers do not cache redirects, this means that
no image contents can be cached -- and upon every page load or reload,
every recently-posted image must be re-fetched. This incurs extra
load on the Zulip server, as well as potentially excessive bandwidth
usage from S3, and on the client's connection.
Switch to fetching the content from S3 in nginx, and serving the
content from nginx. These have `Cache-control: private, immutable`
headers set on the response, allowing browsers to cache them locally.
Because nginx fetching from S3 can be slow, and requests for uploads
will generally be bunched around when a message containing them are
first posted, we instruct nginx to cache the contents locally. This
is safe because uploaded file contents are immutable; access control
is still mediated by Django. The nginx cache key is the URL without
query parameters, as those parameters include a time-limited signed
authentication parameter which lets nginx fetch the non-public file.
This adds a number of nginx-level configuration parameters to control
the caching which nginx performs, including the amount of in-memory
index for he cache, the maximum storage of the cache on disk, and how
long data is retained in the cache. The currently-chosen figures are
reasonable for small to medium deployments.
The most notable effect of this change is in allowing browsers to
cache uploaded image content; however, while there will be many fewer
requests, it also has an improvement on request latency. The
following tests were done with a non-AWS client in SFO, a server and
S3 storage in us-east-1, and with 100 requests after 10 requests of
warm-up (to fill the nginx cache). The mean and standard deviation
are shown.
| | Redirect to S3 | Caching proxy, hot | Caching proxy, cold |
| ----------------- | ------------------- | ------------------- | ------------------- |
| Time in Django | 263.0 ms ± 28.3 ms | 258.0 ms ± 12.3 ms | 258.0 ms ± 12.3 ms |
| Small file (842b) | 586.1 ms ± 21.1 ms | 266.1 ms ± 67.4 ms | 288.6 ms ± 17.7 ms |
| Large file (660k) | 959.6 ms ± 137.9 ms | 609.5 ms ± 13.0 ms | 648.1 ms ± 43.2 ms |
The hot-cache performance is faster for both large and small files,
since it saves the client the time having to make a second request to
a separate host. This performance improvement remains at least 100ms
even if the client is on the same coast as the server.
Cold nginx caches are only slightly slower than hot caches, because
VPC access to S3 endpoints is extremely fast (assuming it is in the
same region as the host), and nginx can pool connections to S3 and
reuse them.
However, all of the 648ms taken to serve a cold-cache large file is
occupied in nginx, as opposed to the only 263ms which was spent in
nginx when using redirects to S3. This means that to overall spend
less time responding to uploaded-file requests in nginx, clients will
need to find files in their local cache, and skip making an
uploaded-file request, at least 60% of the time. Modeling shows a
reduction in the number of client requests by about 70% - 80%.
The `Content-Disposition` header logic can now also be entirely shared
with the local-file codepath, as can the `url_only` path used by
mobile clients. While we could provide the direct-to-S3 temporary
signed URL to mobile clients, we choose to provide the
served-from-Zulip signed URL, to better control caching headers on it,
and greater consistency. In doing so, we adjust the salt used for the
URL; since these URLs are only valid for 60s, the effect of this salt
change is minimal.
Importing `upload_backend` directly means that in testing it must also
be mocked where it is imported, in order to correctly test the right
backend. Since `get_avatar_url` is part of the public
`ZulipUploadBackend` API, add another helper method to call that.
Enforcing a consistent `type` helps us double-check that we're not
playing fast-and-loose with any file paths for local files. As noted
in the comment, this is purely for defense-in-depth.
Passing `write_local_file` a consistent `type` requires removing the
"avatars" out of `realm_avatar_and_logo_path` -- which makes it
consistent across upload backends.
This, in turn, requires a compensatory change to zerver.lib.export, to
be explicit that the realm icons are exported from the avatars
directory. This clarity is likely an improvement.
Anders Kaseorg
Alex Vandiver