We weren't accounting for backlog subscriptions with non-zero message
indexes in their wires. Now, we look through all of our outgoing
subscriptions to identify relevant backlog subscriptions, and leave all
of those.
The more proper fix here is to not include message indexes in the wire
in the first place, since we don't ever reuse that anywhere. But that's
a more invasive change, so we just leave a TODO for it instead.
Previously, we were removing the relevant entry from the `synced` map
before calling `+pull-wire`, which requires an entry to still be there.
This lead to subscriptions not actually being pulled, commonly leading
to "subscribe wire not unique" errors on re-join.
In addition to fixing that, `%remove` actions now try to pull the
subscription regardless of whether they have an entry in the `synced`
map or not. `%leave` is always safe, and we might want to clean up
subscriptions that shouldn't be there anymore in the first place.
On initial load, instead of getting the first 25 links for every
collection, only get their totals and unread counts. This drastically
reduces the pageload time.
Updates the frontend to match:
- Handle initial results without link content.
- Dynamically load in collection's links, even for page 0.
Instead of auto-watching a new collection when it gets added to group
associations, require that the user explicitly choose to view it.
Of course, creation counts as opting in, so watch the collections we
create by default.
This commit pulls the spinner out of the header bar -- and
reincorporates it as a component that hooks into local state when
awaiting a new prop, or disabling an input.
In the wild, ships that were live pre-OS1 still had launch subscriptions
open to the clock on the /tile path, instead of the currently-used
/clocktile path. Additionally, launch state for the clock tile seemed
incomplete.
Here, we simply re-%add the clock to launch.
Note that launch currently does not clean up old subscriptions on
path change. For the pre-OS1 case, the old path is no longer in use,
rendering the subscription harmless. For cases where the correct
subscription was already in place, it'll print a %watch-wire-not-unique,
but doesn't do any harm besides that.
Previously, when the refresh-rate timer activated, and the thread from
the previous activation was still running, we would kill it and start
a new one. For low refresh rates, on slower machines, nodes, or network
connections, this could cause the update to never conclude.
Here we add a timeout-time to eth-watcher's config. If the refresh-rate
timer activates, and a thread exists, but hasn't been running for at
least the specified timeout-time yet, we simply take no action, and wait
for the next refresh timer.
Note that we opted for "at least timeout-time", instead of killing &
restarting directly after the specified timeout-time has passed, to
avoid having to handle an extra timer flow.
In the +on-load logic, we configure the timeout-time for existing
watchdogs as six times the refresh-rate. We want to set
azimuth-tracker's timeout-time to ~m30, and don't care much about other,
less-likely-to-be-active use cases of eth-watcher.
During the #2607 upgrade, strictly local collections got left out of the
listening set. (Because they did not have any outgoing subscriptions.)
This led to personal collections not being available on the frontend.
Here, we add upgrade logic for adding those back to our listening set again.
If a user had explicitly left a personal collection (instead of deleted it, for
whatever reason), they will have to leave it again. This case seems much more
rare than the "my collection is gone" one.
(Re)subscribing gets us a %contacts update, containing the full set of
contacts as it currently exists.
Previously, we would fully delete our local state, only to recreate it
using the data from the update.
Now, we never delete existing data, instead only creating if we don't
have it yet, and adding, removing or recreating contacts if they
changed.
In the future, we'll want an easy way to turn two contacts into an %edit
diff, to let us apply correct semantics to individual contacts, too.
Instead, %bundle and %add if we don't have the group locally yet,
or %add and %remove whatever the difference is between the local group
and the group as specified in the %path update.
In both, we make clear that the wire is always of the /@/group/^ form,
and alias the "group path" portion of the wire for clarity.
For kick, more obviously reuse the same wire, don't reconstruct it.
For watch-nack, only delete from the synced map if the source of the
watch-nack is still relevant. While we don't expect this to be relevant
considering current mode of operation, this does protect us against
strange cases.
Instead of going purely off metadata, we now track the collections we're
listening to, and allow the user to remove collections from that list.
This allows us to remove/ignore collections, without mutilating group
assocations locally.
We were taking care not to re-add something to our data store if we
already had it in there, but were still sending out an update
regardless.
With this, we only send out an update if we weren't previously aware
of the content.
