We let supervisor create the socket for us by making humbug-django a
fcig-program. Unfortunately, supevisor doesn't support putting
fcgi-programs in groups (see
https://github.com/Supervisor/supervisor/issues/148), so we have to
restart tornado and django separately.
To deploy, copy the config files over and restart nginx and
supervisor (via stopping and then starting it because restart is
broken). I believe the automated restart as part of
update-deployment will fail because of the way supervisor treats
programs in groups. If so, after restarting supervisor, you will
also need to run restart-server manually to fill the caches and then
delete the lock directory in humbug-deployments.
(imported from commit bfb5db7dd42dcbc4bfefa2944355b3cbb2ef9104)
The amount of process downtime during a supervisord-mediated restart
appears to be linear in the number of processes that are being
restarted. Therefore, restarting just Django and Tornado causes less
downtime than doing them at the same time as the other worker
processes.
(imported from commit 1fa9ef547bcd88caeec49800664e37d5f2fcb7a8)
Currently our test database is backed by sqlite; this commit moves
us to using postgres for our all database needs. This, in conjunction
with the patched django on github, allow us to have fewer hacks and
more true-to-life tests. It also sets the stage for testing the bulk_create
and schema search_path patches made to django.
Developers will need to run:
./tools/postgres-init-test-db
./tools/do-destroy-rebuild-test-database
this is assuming that they have already run:
./tools/postgres-init-db
./tools/do-destroy-rebuild-database
at some point on this pg_cluster. (The ordering is important; it will other-
wise complain about the south_migration table).
(imported from commit c56c6f27e13df7ae10b2e643e65d669dde61af3d)
This change will make it so that processes related to the app.humbughq.com
server are run under supervisord, which uses a state machine model to ensure
that programs are running. It also ensure process startup order.
We will need to manually switch the old way of running server (in screen) into
this new way of doing things, on both staging and prod (app_frontend.pp has been
updated appropriately). This means:
1) cp servers/puppet/modules/humbug/files/supervisord/conf.d/humbug.conf /etc/supervisord/conf.d
2) installing the supervisor package.
3) killing those while loops in that screen session
4) mkdir /var/log/humbug (as root)
5) /etc/init.d/supervisord start
6) check that nothing broke
(imported from commit 055269a70973db89acd69049e01b185fabdc8f90)
Most of the model logic pertaining to unread counts had been in
zephyr.js, along with a couple global variables. Now the code
is encapsulated in unread.js. It was a pretty straightforward
extraction with some minor method name changes. Also, a small
bit of the logic had also been in stream_list.js.
Conflicts:
tools/jslint/check-all.js
(imported from commit f0abdd48f26ab20c5beaef203479eb5a70dacfff)
The only known outstanding bug with this is that it doesn't properly
handle the updating of a message's highlighting/presence in a narrowed
view (e.g. in theory, a message should disappear if it is edited such
that its subject doesn't match your narrow or it no longer matches
your search). I think I'll just open a trac ticket about that once
this is merged, since it's a little hairy to deal with and kinda a
marginal use case.
Also it's not pretty, but that should be easy to tweak once we get the
framework merged.
Conflicts:
tools/jslint/check-all.js
(imported from commit 2d0e3a440bcd885546bd8e28aff97bf379649950)
South doesn't properly deal with removing the Django User model, so
this commit redoes our South history to instead start after that
migration has already been applied. This allows us to get rid of some
annoying hacks.
Note that developers and staging will need to run
./manage.py migrate --delete-ghost-migrations zephyr
in order to clear out the old versions of the migrations.
(imported from commit 7f45ea601b809dde33720f76e7dfb0ab348b0e65)
This is a prefactoring to eventually eliminate the home_unread_messages
global variable. More commits to follow.
In order to set up process_loaded_for_unread() not to modify
global variable to get its job done, we want to pull it out of
add_messages(), so that add_messages() doesn't have to pass back
state to the 9 different places in the codebase where it's called.
There are only 2 places where process_loaded_for_unread() get
called after this commit.
In order to facilitate pulling up process_loaded_for_unread(), I
made it so that the contract for add_messages() was to accept
already-hydrated messages. This way I could hydrate the messages
before calling process_loaded_for_unread() without have to
worry about double-caching them in add_messages. This will
slightly improve performance, but it was mostly done for code
clarity.
(imported from commit ad5aaad5b1f22c31647370f4c9dcb5f89d7d99a7)
The default is 50, which we believe was causing too much CPU
contention in prod during restarts. This commit lowers the max to 20.
The django-command-line script needs to be restarted manually to pick
up this change.
(imported from commit c2d097d2141be50b6222efb4a34142108c241cce)
The String.localeCompare function is really slow, at least partially
because it creates a locale-aware collator object each time. So now,
when we can, we create and cache a locale-aware collator object.
However, this is not supported on most browsers, so we fall back to a
non-locale-aware comparison. This is not ideal, but for now we are
mostly working with English-speaking customers.
(imported from commit 51aa02e3b9fe4a0ef0cb084874fe26e91c57f65e)
The previous version ended up being (at least sometimes) wrong after
the recent deployment system changes.
(imported from commit dec3beb1b1bf8b9c9ad6820b93b0a5d730d020e8)
This requires manual steps on deploy to each of staging and prod:
(1) Run the new update-deployment code to setup the initial deployment directory.
(2) Restart all the programs running in screen sessions.
(3) Deploy the nginx changes and restart nginx.
(imported from commit 1ffe27933ee79274dc0a93d35c9938712de0ef36)
This is a V1 of this feature. For now, the only way to expand is by narrowing
to the stream---future revisions may add a manual toggle if it is found to be
useful.
Additionally, showing per-subject unread counts will be coming in a future revision
as well.
(imported from commit fb5df0d27e928fa3b0f32b9ff2c1c508202cf7e5)
Django's South migrations support for setting up a new database
doesn't properly handle AUTH_USER_MODEL changing over time. Fix this
by having the initial migration be run with AUTH_USER_MODEL set to the
default value.
(imported from commit c373db9edc61f26527c486c741f8e870614600e3)
We accidentally lost this when we did the User/UserProfile merge (this
commit also deletes the old code to add the auth_user index in
do-destroy-rebuild-database).
This below is mostly just notes for future reference, but when
deploying this change to staging, we should consider running the
following instead of using the migration directly:
CREATE UNIQUE INDEX CONCURRENTLY zephyr_userprofile_email_uniq ON zephyr_userprofile(email);
ALTER TABLE zephyr_userprofile ADD CONSTRAINT zephyr_userprofile_email_uniq UNIQUE USING INDEX zephyr_userprofile_email_uniq;
CREATE INDEX CONCURRENTLY zephyr_userprofile_email ON zephyr_userprofile(email);
But I think it might be the case that it's fine to just run it
directly, since the ALTER TABLE part seems to hang if there's an open
transaction working on a UserProfile object anyway.
(imported from commit 1bf34ce242de51e97c91c8bab86b6b273e17fb43)
This allows blueslip to catch exceptions from the event handlers on
these elements in addition to the other benefits that not using
inline handlers provide.
(imported from commit 2bdcb2496c6c08fa7228a20ce6164b527cf64e41)
When the backend is failing, the frontend often fails in much less
clear ways (e.g. timing out), so it's generally more useful to run the
backend tests first.
(imported from commit 36ac862ad1dbb21e32c0f44ba135c3c29bbea2f5)
It doesn't save much time (maybe 0.5s out of 12+s).
I'm leaving the option in because I think it is still useful for
iteratively testing a single test case.
(imported from commit a0ac43f4c48eec101f05d731740394b30a15773b)
...rather than embedding them into index.html.
This is only acceptable for dev, but the next commit adds an alternative
mechanism for prod.
There isn't actually a manual deployment step here. However, this commit won't
work on staging / prod without the next one (since we don't serve
zephyr/static/templates in prod).
(imported from commit dce7ddfe89e07afc3a96699bb972fd124335aa05)
Before this commit, if you try to arrow around when the selected
message is outside the pointer threshold for recentering, you get a
big jump, even if you are arrowing towards the center of the viewport.
(imported from commit 5c15d5ccccdf027a8bfa8b79bf519fccbfa971d8)
The new nginx configuration file needs to be copied to
/etc/nginx/humbug-include and nginx needs to be restarted when this
commit is deployed.
(imported from commit 6c43f3c2c7a6acee6a852c672c96a38bda01dd0d)
This means all GET parameters were chopped off previously in local testing.
Nginx does pattern matching so this is not a problem on staging.
(imported from commit 25a28155b70d168228ca793fc0122b2ebea408e9)
If the user scrolls super fast, our scroll handler might not catch
the user passing by some messages.
(imported from commit 14cebffcd1321f02443971ac5e1c922db19648ab)
restart-server has been relatively slow recently, and it'd be nice to
know what it is spending its time doing when it hangs for a few
seconds.
(imported from commit a411c951f5a3f2a1366b6d5d3a40d0660ebec11b)
Our previous code could in theory end up clearing the caches it had
just filled, if Tornado's cache filling work happened to be faster
than the memcached flush.
(imported from commit 48174aadad398fb7a7c917a1df765c1261b12a55)
We were previously having an issue where the tutorial could
be pre-empted if you got a few messages while you were first
logging in.
I have some reservations about this being slightly fragile, and a
better approach might be to just have a bit that we use to determine
whether or not you've already seen a tutorial. (Or potentially that
checks whether or not you've ever sent a message.)
(imported from commit f8858f64a36bcd25887b76314caff283929f340c)
The new system, called blueslip, makes errors fatal when in debug
mode and only output a message when running in production. In the
future, it could also send user errors back to us automatically.
(imported from commit 1232607c0311e885c8b5a5e8a45ffb28822426e0)
This should substantially improve the repeat-rendering time for pages
with large numbers of tweets since we don't need to go all the way to
twitter.com, which can take like a second, to render tweets properly.
To deploy this commit properly, one needs to run
./manage.py createcachetable third_party_api_results
(imported from commit 01b528e61f9dde2ee718bdec0490088907b6017e)
By splitting up all_msg_list and home_msg_list, we can properly add/remove
streams from the home view without having to jump through hoops.
(imported from commit 92767197759f7519197dfc58be951b60fa823fbb)
Messages are now selected on a MessageList, which triggers a
message_selected event that other parts of the code can listen for.
(imported from commit 1da9e4121425c0ac4461b41b7aea169072e1512b)
Previously we would select the first message in the block. Now, we
only do that if a message that is selected will not be in the
resulting narrowed view. If the selected message will be in the
narrowed view, we select that message once narrowed.
(imported from commit 4da5a3a0b597b58c2e028f1b29ac20ae3808a4d1)
- Find Python and JavaScript files the same way, by looking at
everything in Git with specific exceptions
- Check the non-browser JavaScript, with appropriate JSLint options
Fixes#19 and #387.
(imported from commit 042bd6c26aa7a7ea2209935320ef4da12344b02b)
Zev Benjamin
Tim Abbott
acrefoot
Steve Howell
Leo Franchi
Jessica McKellar
Luke Faraone
Keegan McAllister
Waseem Daher
Jeff Arnold