Thread resolving is one of the few tasks that isn't retried on failure.
One common cause for failure of this task is a well-connected user replying to
a toot from a little-connected user on a small instance: the small instance
will get many requests at once, and will often fail to answer requests within
the 10 seconds timeout used by Mastodon.

This changes makes the ThreadResolveWorker retry a few times, with a
rapidly-increasing time before retries and large random contribution in order
to spread the load over time.

must be added to the Sidekiq invokation in your systemd file

The pull queue will handle link crawling, thread resolving, and OStatus
processing. Such tasks are more likely to hang for a longer time (due to
network requests) so it is more sensible to not make the "in-house" tasks
wait for them.

since that is only ever called in the background

This is a big one, so let me enumerate:

Accounts as well as stream entry pages now contain Link headers that
reference the Atom feed and Webfinger URL for the former and Atom entry
for the latter. So you only need to HEAD those resources to get that
information, no need to download and parse HTML <link>s.

ProcessFeedService will now queue ThreadResolveWorker for each remote
status that it cannot find otherwise. Furthermore, entries are now
processed in reverse order (from bottom to top) in case a newer entry
references a chronologically previous one.

ThreadResolveWorker uses FetchRemoteStatusService to obtain a status
and attach the child status it was queued for to it.

FetchRemoteStatusService looks up the URL, first with a HEAD, tests
if it's an Atom feed, in which case it processes it directly. Next
for Link headers to the Atom feed, in which case that is fetched
and processed. Lastly if it's HTML, it is checked for <link>s to the Atom
feed, and if such is found, that is fetched and processed. The account for
the status is derived from author/name attribute in the XML and the hostname
in the URL (domain). FollowRemoteAccountService and ProcessFeedService
are used.

This means that potentially threads are resolved recursively until a dead-end
is encountered, however it is performed asynchronously over background jobs,
so it should be ok.