:discard marks jobs as "discarded", i.e. jobs which permanently failed
due to e.g. exhausting all retries or explicitly being discared due to a
fatal error.
:cancel marks jobs as "cancelled" which does not imply failure.
While neither method counts as a job "exception" in the set of
telemetries we currently export via Prometheus, the different state
is visible in the (not-exported) metadata of oban job telemetry.
We can use handlers of those events to build bespoke statistics.
Ideally we'd like to distinguish in the receiver worker between
"invalid" and "already present or delete of unknown" documents,
but this is cumbersome to get get right with a list of
free-form, human-readable descriptions oof the violated constraints.
For now, just count both as an fatal error.
# but that is cumbersome to get right with a list of string error descriptions
E.g. \*oma federates (most) follower-only posts multiple times
to each personal inbox. This commonly leads to race conditions
with jobs of several copies running at the same time and getting
past the initial "already known" check but then later all but
one will crash with an exception from the unique db index.
Since the only special thing we do with copies anyway is to discard them,
just don't create such duplicate jobs in the first place.
For the same reason and since failed jobs don't count towards
duplicates, this should have virtually no effect on federation.
With the current strategy the individual
and cumulative backoff looks like this
(the + part denotes max extra random delay):
attempt backoff_single cumulative
1 16+30 16+30
2 47+60 63+90
3 243+90 ≈ 4min 321+180
4 1024+120 ≈17min 1360+300 ≈23+5min
5 3125+150 ≈20min 4500+450 ≈75+8min
6 7776+180 ≈ 2.1h 12291+630 ≈3.4h
7 16807+210 ≈ 4.6h 29113+840 ≈8h
8 32768+240 ≈ 9.1h 61896+1080 ≈17h
9 59049+270 ≈16.4h 120960+1350 ≈33h
10 100000+300 ≈27.7h 220975+1650 ≈61h
We default to 5 retries meaning the least backoff runs with attempt=4.
Therefore outgoing activiities might already be permanently dropped by a
downtime of only 23 minutes which doesn't seem too implausible to occur.
Furthermore it seems excessive to retry this quickly this often at the
beginning.
At the same time, we’d like to have at least one quick'ish retry to deal
with transient issues and maintain reasonable federation responsiveness.
If an admin wants to tolerate one -day downtime of remotes,
retries need to be almost doubled.
The new backoff strategy implemented in this commit instead
switches to an exponetial after a few initial attempts:
attempt backoff_single cumulative
1 16+30 16+30
2 143+60 159+90
3 2202+90 ≈37min 2361+180 ≈40min
4 8160+120 ≈ 2.3h 10521+300 ≈ 3h
5 77393+150 ≈21.5h 87914+450 ≈24h
Initial retries are still fast, but the same amount of retries
now allows a remote downtime of at least 40 minutes. Customising
the retry count to 5 allows for whole-day downtimes.
Oban cataches crashes to handle job failure and retry,
thus it never bubbles up all the way and nothing is logged by default.
For better debugging, catch and log any crashes.
To facilitate this ObjectValidator.fetch_actor_and_object is adapted to
return an informative error. Otherwise we’d be unable to make an
informed decision on retrying or not later. There’s no point in
retrying to fetch MRF-blocked stuff or private posts for example.
This is the only user of fetch_actor_and_object which previously just
always preteneded to be successful. For all the activity types handled
here, we absolutely need the referenced object to be able to process it
(other than Announce whether or not processing those activity types for
unknown remote objects is desirable in the first place is up for debate)
All other users of the similar fetch_actor already properly check success.
Note, this currently lumps all reolv failure reasons together,
so even e.g. boosts of MRF rejected posts will still exhaust all
retries. The following commit improves on this.
It makes decisions based on error sources harder since all possible
nesting levels need to be checked for. As shown by the return values
handled in the receiver worker something else still nests those,
but this is a first start.
Ideally we’d like to split this up more and count most invalid documents
as an error, but silently drop e.g. Deletes for unknown objects.
However, this is hard to extract from the changeset and jobs canceled
with :discard don’t count as exceptions and I’m not aware of a idiomatic
way to cancel further retries while retaining the exception status.
Thus at least keep a log, but since superfluous "Delete"s
seem kinda frequent, don't log at error, only info level.
There were two issues leading to needles effort:
Most importnatly, the use of AP IDs as "source_url" meant multiple
simultaneous jobs got scheduled for the same instance even with the
default unique settings.
Also jobs were scheduled uncontionally for each processed AP object
meaning we incured oberhead from managing Oban jobs even if we knew it
wasn't necessary. By comparison the single query to check if an update
is needed should be cheaper overall.
It was used to migrate OStatus connections to ActivityPub if possible,
but support for OStatus was long since dropped, all new actors always AP
and if anything wasn't migrated before, their instance is already marked
as unreachable anyway.
The associated logic was also buggy in several ways and deleted users
got set to ap_enabled=false also causing some issues.
This patch is a pretty direct port of the original Pleroma MR;
follow-up commits will further fix and clean up remaining issues.
Changes made (other than trivial merge conflict resolutions):
- converted CHANGELOG format
- adapted migration id for Akkoma’s timeline
- removed ap_enabled from additional tests
Ported-from: https://git.pleroma.social/pleroma/pleroma/-/merge_requests/3880
Otherwise attachments have a high chance to disappear with akkoma-fe’s
“delete & redraft” feature when cleanup is enabled in the backend. Since
we don't know whether a deletion was intended to be part of a redraft
process or even if whether the redraft was abandoned we still have to
delete attachments eventually.
A thirty minute delay should provide sufficient time for redrafting.
Fixes: https://akkoma.dev/AkkomaGang/akkoma/issues/775
Rich Media parsing was previously handled on-demand with a 2 second HTTP request timeout and retained only in Cachex. Every time a Pleroma instance is restarted it will have to request and parse the data for each status with a URL detected. When fetching a batch of statuses they were processed in parallel to attempt to keep the maximum latency at 2 seconds, but often resulted in a timeline appearing to hang during loading due to a URL that could not be successfully reached. URLs which had images links that expire (Amazon AWS) were parsed and inserted with a TTL to ensure the image link would not break.
Rich Media data is now cached in the database and fetched asynchronously. Cachex is used as a read-through cache. When the data becomes available we stream an update to the clients. If the result is returned quickly the experience is almost seamless. Activities were already processed for their Rich Media data during ingestion to warm the cache, so users should not normally encounter the asynchronous loading of the Rich Media data.
Implementation notes:
- The async worker is a Task with a globally unique process name to prevent duplicate processing of the same URL
- The Task will attempt to fetch the data 3 times with increasing sleep time between attempts
- The HTTP request obeys the default HTTP request timeout value instead of 2 seconds
- URLs that cannot be successfully parsed due to an unexpected error receives a negative cache entry for 15 minutes
- URLs that fail with an expected error will receive a negative cache with no TTL
- Activities that have no detected URLs insert a nil value in the Cachex :scrubber_cache so we do not repeat parsing the object content with Floki every time the activity is rendered
- Expiring image URLs are handled with an Oban job
- There is no automatic cleanup of the Rich Media data in the database, but it is safe to delete at any time
- The post draft/preview feature makes the URL processing synchronous so the rendered post preview will have an accurate rendering
Overall performance of timelines and creating new posts which contain URLs is greatly improved.
by default just prevent job floods with a 1-seconds
uniqueness check, but override in RemoteFetcherWorker
for 5 minute uniqueness check over all states
:infinity is an option we can go for maybe at some point,
but that would prevent any refetches so maybe not idk.
