* SyncFatalError: show nag instead of hard-crashing.
## Issue
When sync fails, we crash the app.
## Ticket
Maybe closes 39 "Better handle both internal and web backend interruptions / downtime"
## Approach
I'm tackling this from the standpoint that (1) sync errors are not that fatal -- we'll just lost a few recent changes (2) network disconnection is the common cause.
## Changes
- If we are offline:
- Inform user through a nag. All other status is meaningless if we are offline.
- If we are online:
- If api is STATUS_DOWN, show the existing crash page.
- If there is a sync error, show a nag saying settings are now potentially unsynchronized, and add a button to retry sync.
- If there is a chunk error, nag to reload.
* Attempt to detect `status=DOWN`
Previous code resolves the status to either "ok" or "not", which makes the app unable to differentiate between the "degraded" (nag) and "down" (crash) states.
## Issue
The TUS client automatically removes the upload fingerprint whenever there is a 4xx error. When we try to resume later, we couldn't find the the fingerprint and ended up creating a new upload ID.
## Changes
Since we are also storing the uploadUrl ourselves, provided that to override the tus client's default behavior of restarting a new session on 4xx errors.
## Issue
- It was checking the blacklist on every render.
- Finding opportunities to improve performance
## Changes
Since the final destination will be a dead end anyways, skip the blacklist check so that livestreams can render a bit faster when there lots of mentions.
The only downside is that a claim preview for a blacklisted item would now appear (vs. being a text previously)
## Issue
- Large resolve count (albeit batched) on bootup.
## Changes
- Skip the call on bootup. The same call will happen when you click the notification bell, so it's not too late to resolve at that time.
- Added `true` to `doResolveUris` to return cached results, otherwise it will keep resolving the same channels every time we enter Notifications Page.
The stalling behavior has changed a bit, probably with the removal of CF.
The stall difference between 10MB and 50MB is not too noticable, so picking 25MB as a start.
## Issue
312 Save notification on back navigation, enable filter on mobile
## Changes
- Don't clear then filter when mounted and there are unread notifications.
- We previously clear the filter because the user could be clicking the notification bell (which is showing some number) and we ended up with a blank page because of the filter.
- Allow the filter in mobile.
- Previously, it was intentionally removed for mobile (see bd42418f). I believe it was just because we don't have the style set up for mobile. Here's my quick attempt.
## Mistake
Tried to simplify the logic between checking Channel vs Content claim, and ended up always checking against Channel. This is correct for commentron blocklists, but not blacklists where the txid is per claim.
## Changes
- Restored original logic.
- While at it, restore the usage of `selectClaimForUri` (i.e. no need to split into 4 selectors anymore), since we've updated the reducer to prevent invalidation from things like 'confirmation' and 'is_my_output'.
## Issue
One of the bottlenecks of livestream page.
The component probably needs a re-design:
- Don't perpetually mount -- only mount when activated by the user through "@". This would avoid the heavy processing entirely.
- Better way of resolving uris (too many arrays, too many loops).
- Tom also mentioned that we should not be resolving every commenter as we see encounter them in a livestream. This is currently the case because the component is always mounted.
## Changes
Until the re-design occurs, attempt to cache the heavy processing. Also, trimmed down the amount of loops.
For the case of livestreams, the comments are added incrementally via websocket. The selector returns everything, which grows as a user watches the livestream.
We could even make it a bit more efficient by passing in `maxCount` to `filterComments`, and do a `for` loop there, but decided to keep things readable by not changing the `filter` usage.
As long as the input parameters are the same, the selector will return the cached value so that we don't construct the list twice, which involves blocklist filtering.
`<div>` cannot be a descendend of `<p>`, and `{subtitle}`s sometimes need to have `<div>`s.
Just switch from `<p>` to `<div>`, and let the client decide when the actual text paragraphs are.
## Issue
Lots of time is spent mapping the state to props for this component (since there are lots of tiles).
## Changes
Using this component as a starting point, go through the selectors and make the usual cleanup/fixes:
- Move away from the `makeSelect*` model, which creates a new selector on every call instead of actually re-using the cached version.
- Do proper caching for multi-param selectors using `re-reselect`.
- Don't cache simple functions or direct access to states.
- Memo not required. `resolvingUris` is very dynamic and is a short array anyways.
- Changeg from using `indexOf` to `includes`, which is more concise.
## Issues
- 251 Dragging the floating player is super laggy
Recent changes and/or refactoring combined the effects or added dependancies into the effects, causing them to re-run excessively.
## Changes
- Restored effects to their original behavior.
- Don't perform the position check when dragging -- only do it when released.
- Do proper debouncing for the 'resize' listener -- the previous method was incorrect as a new function is created on each render.
## Changes
- doHandleSyncComplete: only call doGetAndPopulatePreferences when there is new data.
- But for that to work, we'll need to populate preferences at least once. We'll do that in doSignIn.
- We can also remove the "sync/prefs ready" mechanism that was mainly meant for Desktop.
Then came another problem: while trying to spark changes between 2 tabs, `sync/get` was saying "no change" despite the local and server hash being different. I think it is because the both `sync_hash + sync/get` combo is operating on server data, so the hash is the same. I'm guessing this is why we ended up just running doGetAndPopulatePreferences every time before PR, since this flag wasn't correct in this scenario.
- Updated `data.changed` to consider both API results and comparison with local hash.
