This stat is shown on a few of our pages, such as our homepage,
the performance page and also a platform admin page
and is currently catched for the
default TTL of 1 week. I think there is no reason we can't make
this only cache once an hour and give slightly more up to date stats
which will update more regularly.
This mimics the approach and also the TTL choice of 1 hour that has
been added for the performance page (although there is no
particular bug here to fix, it is just nice to have slightly more up
up to date data).
Note, the API call only takes about 0.3 seconds at the moment
so it is not particularly intensive on the DB to run this more
regularly.
I came across a bug where the performance page might be visited
on say the 22nd and we expect it to show data for yesterday and
the past 7 days (so the 21st and back 7 days) but instead it
only shows it for the 20th and then back 6 days.
The cause is that we have nightly tasks that run to
calculate the number of notifications sent per service (the
ft_notification_status table)
calculate the number of notifications sent within 10 seconds
(the ft_processing_time table)
Both of these tables get filled between the hours of midnight and
2:30am. The first seems to be filled by about 12:30 every night
whereas the processing stats seems to be filled about 2am every
night.
The admin app currently caches the results of the call to the API
(https://github.com/alphagov/notifications-api/blob/master/app/performance_dashboard/rest.py#L24)
to get this data with the key
performance-stats-{start_date}-to-{end_date}.
Now the issue here is that if the performance page is visited at
00:01 on the 23rd, it will call the API for data on the 22nd and
backwards 7 days. The data it gets at 00:01 will not yet have the
data for the 22nd, it will only go as far as the 21st, because
the overnight jobs to put data in the tables have not run yet. The
admin app then caches this data, meaning that for the rest of the
day the page will be missing data for the 22nd, even though it
becomes available after approx 2am. We have seen it a few times
in production where the performance page has indeed been visited
before say 2 am, leaving the page with missing data.
In terms of solutions, there are several potential options.
1. Remove the caching that we do in redis. The downside of this is
that we will hit the API more (currently once a day for the first
visitor to the performance page). We have had 255 requests
to the performance page in the last week and when there is no value
in redis, the call to the API takes approximately 1-2 seconds.
Removing the caching would against why the caching was originally
added which was to act as a basic protection against malicious
DOS attempts.
2. Make the admin app only set the cache value if it is after
2:30am. This one feels a bit gross and snowflakey.
3. The API flushes the redis cache after it has finished its nightly
tasks. We don't have that much precedent of the API interacting
with redis, it is mostly the admin app that does but this is still
a valid option.
4. Cache in redis the data for 2 days ago to 7 days ago and then
get the data for yesterday freshly every time. This would mean
some things are still cached but the smallest bit that we can't rely
on can be gathered fresh.
5. Remove the caching we do in redis but then try and optimise the
API endpoint to return results even faster and with less
interaction with the DB. My hypothesis is that the slowest part
is where the API has to calculate how many notifications Notify
has ever sent
(https://github.com/alphagov/notifications-api/blob/master/app/performance_dashboard/rest.py#L36)
and so we could potentially try and store these totals in the DB
every night and update them nightly rather than having to query all
the data in the table to figure them out every time.
6. Reduce the expiry time of the cache key so although the bug will
still exist, it will only exist for a much shorter time.
In the end, we've gone for option 6. The user impact of
this bug sticking around for an hour is minimal, in particular
because that would be in the early hours of the morning. The solution
is also quick to implement and keeps the protection against a DOS attack.
The value of 1 hour for expiry was a fairly abitrary choice, it could
have been as little as 1 minute or as much as 6 hours.
Looking up the coordinate reference system for a given polygon’s bounds
is surprisingly expensive.
We can make things run faster by precomputing it at the time of
generating the simplified polygons, then storing it in the SQLite
database alongside the polygon data.
Transforming full resolution polygons to linear coordinates is somewhat
expensive. We can make things run faster by:
- looking at `simple_polygons` which have fewer points and are therefore
faster to transform.
- reusing polygons that have already been transformed where possible
Brings in https://github.com/alphagov/notifications-utils/pull/889/files
At the moment, we are not doing any transformation of features before
applying geometric algorithms to them. This is, in effect, assuming that
the earth is flat.
This new version of utils implements the transformation of our polygons
to a Cartesian plane. In other words, it converts them from being
defined in spherical degrees to metres.
For the admin app this means we need to convert places where the code
expects things to be measured in degrees to work in metres instead.
WTForms versions less than 3.0.0 have a security vulnerability where
arbitrary HTML can be inserted into the label of a form, allowing the
possibility of a cross-site scripting attack.
I don’t know if there’s anywhere we put user-generated content into form
labels but it’s possible we are vulnerable somewhere.
This require moving some imports because as of
https://github.com/wtforms/wtforms/pull/614/files
there is no longer a separate module for HTML 5 fields, they are now
considered core fields.
As of https://github.com/wtforms/wtforms/issues/445/files custom
implementations of `pre_validate` or `post_validate` must raise
`ValidationError` to trigger a validation message, where we were raising
`ValueError` this was no longer being caught.
As of https://github.com/wtforms/wtforms/pull/355/files `StringField`
returns `None` for empty data, not `''` but our `validate_email_address`
function only accepts strings.
We think that the API is returning incorrect data for this column.
It’s going to take a while to figure out what’s going on with the
queries in the API, so this pull request temporarily removes the column
so we’ve not giving people incorrect data.
We are starting to see lots of 100.0%s in the current table
and we think this looks suspiciously too good so think it is
beneficial to change it to be 2dp such that we get a few more
non 100.0% values.
This will put all the values in to having 2dp, however it will
also require the API to have a change to
https://github.com/alphagov/notifications-api/blob/master/app/performance_dashboard/rest.py#L81
where it is currently losing the granularity down to a single
decimal place (meaning that if we were really at 99.894% then that
would be shown on the page as 99.90% rather than 99.89%). However,
I don't think it is a blocker that we get that sorted before this
can be merged.
This means that the data in the report will match what’s on the page,
where the values are rounded to the nearest penny.
This uses the same string formatting to round the numbers which the
`big_number` component does, so it should round the numbers in the same
way.
It’s a pain having to remember to update both of them. Looks like `pip`
now supports a different syntax without the `egg` bit which means a git
dependency only needs the version number changing in one place.
This is so org level users can use this data easier for things
like determining spending per service.
We do not include sms fragments sent column and remove other sms columns
consistency.
Do not add sms fragments sent column for now until we agree on an
unambiguous name for it. The data in this column is sms billing units
multiplied by international sms weighing. My favourite for a clear
name would be 'text message credits used', but we need a naming
strategy for this.
The sentences in the key to the map have 2 forms:
1. the visual form, with an graphic pattern
alongside text saying what the pattern means
in the map
2. the semantic form, with text describing what
the pattern means and how big an area it covers
for this alert
This is achieved by making the graphic
non-semantic and having hidden words in the
sentence for each paragraph.
While the spaces between words are correct for
both forms of these sentences, we observed screen
readers speaking some groups of words as one when
they spoke the sentence. Most noticably, 'the
alert' was spoken as 'thealert'.
Swapping the class used to hide the visually
hidden words from ours to the GOVUK one seemed to
fix this. It's unclear why it does but the
`govuk-visually-hidden` class adds the following
styles to those in ours:
- width: 1px
- height: 1px
- white-space: nowrap
It also has `margin: 0` instead of `margin: -1px`.
At the moment, the icons in the key of the map
showing what each type of area drawn on the map
means change colour in high contrast mode. This
causes a mismatch between the areas on the map and
the key.
These changes remake the images in SVG to prevent
the colours changing in high contrast modes. They
also add a white background to the icons, to make
sure they match the areas they refer to in the
map.
