The previous batch of improvements to this code path in
6562ea94e4 introduced a race bug where:
- You navigate to a narrowed view; Zulip renders cached data from
`all_messages_data` that we know is current, but
`fetch_status.has_found_newest` is initialized to `false`
nonetheless.
- The bottom is visible, triggering the check for whether the view
should be marked as read.
- `fetch_status.has_found_newest` is still `false`, and so we
incorrectly refuse to mark as read.
- We finish fetching data from the server, do the background rerender
for that (with no changes), but that doesn't trigger a re-check for
whether the bottom is visible.
There's several ways to address this, but most correct to me is to not
check fetch_status in this particular code path.
The same reasoning applies to the navigate.js call sites.
This error message didn’t make sense for the check as written, and our
OpenAPI document already provides the expected format for our 200
responses.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Real requests would not validate against the previous version. There
seems to be no consistent way to determine whether a string parameter
should be coerced to an integer for validation against an allOf
schema (which works at the level of JSON objects, not strings).
See also https://github.com/python-openapi/openapi-core/issues/698.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Previously, when no parameter was passed to the get_timestamp_for_
flatpickr method, it would result in an uncaught exception. This is
breaking the "Add global time" of compose bar.
This can be avoided by doing an early return of current time to hour
in case no string is passed.
This will require customers to include an address when setting
up, or updating, the credit card information for their account.
The billing address for the card will also be saved as the
billing address for the stripe customer object.
The customer object billing address appears on the invoices
that are generated by stripe.
This was introduced in #28767 with the intention to skip scrolling
the selected message.
So, the actual bug that the PR fixed would have been just fixed
by opening the compose box early.
This matches the algorithm that we designed for the Python API, except
that we use a ratio of 2 rather than sqrt(2) in the message_fetch code
path, because it's a heavier request.
We increase the number of failures before showing a user-facing error
to roughly preserve the same time period before a user-facing error is
shown.
Previously, these endpoints just did exponential backoff, without
looking at the rate-limiting headers returned by the server, resulting
in requests that the client could have been certain would fail with an
additional rate-limiting error.
Fix this by using the maximum of the existing exponential backoff with
the value returned by the rate-limiting header.
Fixes#28807.
This should help reduce the risk of hitting rate limits when users
have a very large number of messages to fetch via this mechanism.
Inline the `messages` variable that was only used in one place while
we're touching this.
This function incorrectly and misleadingly did an immediate initial
call, despite both of its callers doing immediate calls themselves (in
one case, with a different parameter passed).
This led to unnecessary server load when reloading the app via event
system triggered reloads, since every client would call `/` twice.
I was not able to reproduce obviously badly broken behavior from these
logic bugs, but after the renaming of message_viewport helpers in the
last few commits, it's clear that this logic was trying to check if
we're actually at the start/end of the possibly message feed, not just
the rendered portion, and doing so incorrectly.
The previous logic for both scrolling down and using pagedown would
incorrectly mark an entire conversation as read when reaching the
bottom of a render window, even if there were more messages loaded or
to fetch from the server.
Fix this error in the calculation by asking the correct data
structures if we're actually at the bottom.
To avoid the navigate.js keyboard shortcut code paths circumventing
this new logic, or needing to duplicate it, they now call
process_visible, rather than its helper.
Since we always call `deactivate` from `hashchange`,
`browser_history.state.changing_hash` is always `true` and hence
`save_narrow` just retuns without doing anything.
Tim Abbott
Anders Kaseorg
roanster007
Alex Vandiver
Aman Agrawal
Lauryn Menard