Streams can have lots of subscribers, meaning that the archiving process
will be moving tons of UserMessages per message. For that reason, using
a smaller batch size for stream messages is justified.
Some personal messages need to be added in test_scrub_realm to have
coverage of do_delete_messages_by_sender after these changes.
Currently, we use -1 as the Realm.message_retention_days value to retain
message forever unless specified at stream level for a particular stream,
that is, no policy set at the realm level. But this is incoherent with what
we use for Stream.message_retention_days where -1 means
> disable retention policy for this stream unconditionally
that can be confusing from an API standpoint.
So instead of trying some hack to reset the value to NULL or using some
other value like -2 for RETAIN_MESSAGE_FOREVER and use that for API. It is
much more intuitive to use a string like 'forever' that can be mapped to
RETAIN_MESSAGE_FOREVER at the backend. And this is similar to what we use
for streams settings as well.
To be more consistent with the meaning in the Stream model, and to make
it easier to have a reasonable settings API, we get rid of the None
value for Realm.message_retention_days in favor of the value -1 to
represent the "don't delete messages" default policy.
Fixes#2665.
Regenerated by tabbott with `lint --fix` after a rebase and change in
parameters.
Note from tabbott: In a few cases, this converts technical debt in the
form of unsorted imports into different technical debt in the form of
our largest files having very long, ugly import sequences at the
start. I expect this change will increase pressure for us to split
those files, which isn't a bad thing.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Automatically generated by the following script, based on the output
of lint with flake8-comma:
import re
import sys
last_filename = None
last_row = None
lines = []
for msg in sys.stdin:
m = re.match(
r"\x1b\[35mflake8 \|\x1b\[0m \x1b\[1;31m(.+):(\d+):(\d+): (\w+)", msg
)
if m:
filename, row_str, col_str, err = m.groups()
row, col = int(row_str), int(col_str)
if filename == last_filename:
assert last_row != row
else:
if last_filename is not None:
with open(last_filename, "w") as f:
f.writelines(lines)
with open(filename) as f:
lines = f.readlines()
last_filename = filename
last_row = row
line = lines[row - 1]
if err in ["C812", "C815"]:
lines[row - 1] = line[: col - 1] + "," + line[col - 1 :]
elif err in ["C819"]:
assert line[col - 2] == ","
lines[row - 1] = line[: col - 2] + line[col - 1 :].lstrip(" ")
if last_filename is not None:
with open(last_filename, "w") as f:
f.writelines(lines)
Signed-off-by: Anders Kaseorg <anders@zulipchat.com>
For unknown reasons, deleting 10,000s of ArchiveTransaction objects
results in rapidly growing memory in the job making the request in the
Django process, eventually leading to an OOM kill.
I don't understand why Django behaves that way; I would have expected
the failure mode to instead be a serious load problem on the database
server, but perhaps the way Django's internal deletion logic handles
cascading the deletes to many millions of ArchiveMessages and other
ForeignKey objects requires tracking a lot of data in memory.
The solution is the same in any case, which is to batch the deletions
to execute a reasonable number of them at once. Doing a single
ArchiveTransaction at a time would likely result in huge numbers of
database queries in a loop, which performs very poorly. So we balance
by batching deletions in groups of 100 ArchiveTransactions; testing
this in production, I saw no spike of memory usage materially beyond
that of a normal Django process, and each bulk-deletion transaction
takes several seconds to process (meaning per-transaction overhead is
negligible).
For realms with no retention policy on themselves or any of their
streams, no archiving happens, but 3 lines of logs would be generated.
That's redundant and we make changes in this commit to avoid logging
those lines if nothing of interest is happening.
In https://github.com/zulip/zulip/pull/12823 some changes to the realms
structure have been made, so now both in production and development
cross-realm bots live in the realm with string_id "zulipinternal".
There was a TODO in retention code to eliminate a conditional in a query
that became redundant with this change, and also the zulipinternal realm
should be omitted from the archiving process in archive_messages().
With the recipient field being denormalized into the UserProfile and
Streams models, all current uses of get_stream_recipients can be done
more efficiently, by simply checking the .recipient_id attribute on the
appropriate objects.
Fixes#1727.
With the server down, apply migrations 0245 and 0246. 0246 will remove
the pub_date column, so it's essential that the previous migrations
ran correctly to copy data before running this.
Previous cleanups (mostly the removals of Python __future__ imports)
were done in a way that introduced leading newlines. Delete leading
newlines from all files, except static/assets/zulip-emoji/NOTICE,
which is a verbatim copy of the Apache 2.0 license.
Signed-off-by: Anders Kaseorg <anders@zulipchat.com>
We can simply archive cross-realm personal messages according to the
retention policy of the recipient's realm. It requires adding another
message-archiving query for this case however.
What remains is to figure out how to treat cross-realm huddle messages.
This reverts commit 8f15884c7d. Using the
WITH ( ) ... DELETE method leads to a small performance drop, while
probably not offering many positives, so it seems appropriate to go to
the simpler case of just letting things get cleaned up by CASCADE.
The way the code changed in this commit was written caused Django to
fetch stream.realm from the database for every stream, leading to
redundant, identical queries. Each stream's realm is already known, so
we use that information.
Duplicate handling when INSERTing is switched from "LEFT JOIN ... id IS
NULL" approach to "ON CONFLICT (id) DO NOTHING", since we now have
postgresql 9.5. The ON CONFLICT approach is more natural as well as also
potentially being faster,
When archiving Messages, we stop relying on LEFT JOIN ... IS NULL to
avoid duplicates when INSERTing. Instead we use ON CONFLICT DO UPDATE
(added in postgresql 9.5) to, in case of archiving a Message that
already has a corresponding archived objects (this happens if a Message
gets archived, restored and then archived again), re-assign the existing
ArchivedMessage to the new transaction.
This also allows us to fix test_archiving_messages_second_time, which
was temporarily disable a few commits before.
We combine run_message_batch_query and run_archiving_in_chunks
functions, which makes the code simpler and more readable - we get rid
of hacky generator usage, for example.
In the process, move_expired_messages_* functions are adjusted, and now
they archive Messages as well as their related objects.
Appropriate adjustments in reaction to this are made in the main
archiving functions which call move_expired_messages_* (they no longer
need to call move_related_objects_to_archive).
Instead of having a bunch of custom code in the function, we make it use
run_message_batch_query and run_archiving_in_chunks to do the necessary
operations in a consistent way, using the same codepaths as the rest of
the archiving system.
This breaks test_archiving_messages_second_time temporarily, but we will
fix it and re-enable the test in the next commits, where we'll address
various other issues with re-archiving of messages.
We also remove the @transaction.atomic wrapper, because atomicity is
handled by the logic inside run_archiving_in_chunks.
We add a new model, ArchiveTransaction, to tie archived objects together
in a coherent way, according to the batches in which they are archived.
This enables making a better system for restoring from archive, and it
seems just more sensible to tie the archived objects in this way, rather
the somewhat vague setting of archive_timestamp to each object using
timezone_now().
Rather than relying on the CASCADING property of the ForeignKey to the
Message table to clean up these objects, we delete them in the same
query as we archive them - since it's guaranteed that any of these
objects that we archive will be deleted due to their Message being
deleted later.
We don't have this guarantee for Attachment objects, which is why we
can't apply this scheme to them.
To ensure the database retains a consistent state if archiving gets
interrupted, we process each Messages chunk together with related
objects in a single atomic transaction.
We had two duplicate functions for archiving zerver_attachment_messages
rows, doing the same thing - archiving by message_id. One of them had a
redundant INNER JOIN, so we get rid of that too.
Since we loop over realms in the functions for archiving stream messages
and then personal+huddle messages, and also want to split cleaning up
attachments by realm - it makes sense to do it all in one single loop.
We batch queries that archive Messages, to limit the maximum amount of
Message objects archived in a single query. This leads to the archiving
of other related objects being batched as well, because we loop over
chunks of archived messages and archive their related objects per-chunk.
We add the following behavior:
If stream has message_retention_days set to -1, archiving for it is
disabled.
If stream has message_retention_days set to null, use the realm's
policy. If the realm has no policy, we don't archive for this stream.
UserMessages no longer need special handling, they can be archived by
move_models_with_message_key_to_archive and automatically cleaned up
like the other models with a message key with CASCADING=True.
We change the archiving scheme to allow having stream based retention
policies. In the first step of the archiving process, we loop over
streams and archive their expired messages and related objects.
Then we separately archive all expired personal and huddle messages and
related objects. As the last step, we scan for redundant attachments
which can now be deleted.
To achieve this, we have to rewrite a significant portion of the
retention code and rework some of the database queries.
For the sake of simplicity, we neither archive nor delete cross-realm
messages, except cross-realm stream messages – in their case they can
be processed in the same manner as ordinary stream messages.
In the query for archiving personal and huddle messages we simply
exclude those sent by cross-realm bots.
We change the tests to adapt to these modifications.
Mateusz Mandera
Pragati Agrawal
Tim Abbott
Anders Kaseorg
Anders Kaseorg
Mateusz Mandera
Tim Abbott