The feature added in 4.1 to lock the modinfo cache does not work when
using Redis, because:
* The API to acquire a cache lock is confusing, and the code did not
check that it successfully acquired a lock before going on to build
the cache anyway.
* Unlike the other types of cache lock, the Redis store did not retry
the lock for a timeout period before giving up and failing.
This change fixes both points.
The new library version should work without problems with older
php-mongodb extension versions:
- v1.15.0 (paired version, tested).
- v1.14.1 (tested)
But, in order to keep them paired, the requirements have been
raised to the extension v1.15.0. See:
https://www.mongodb.com/docs/drivers/php/#compatibility
This updates the library from v1.13.1. Fixes some php81 warnings
leading to failed/risky tests.
The new library version should work without problems with older
php-mongodb extension versions:
- v1.15.0 (paired version, tested).
- v1.14.1 (tested)
In some cases, we get an error message such as:
Wrong parameter count for Redis::zRem()
Within the delete_many function. This function requires at least one
key to be supplied, but if delete_many is called with an empty array,
we will call it with no keys.
The cache file is only ever written to in the `write_file` function,
where it does so by writing to a temp file and performing an atomic
rename of that file. When writing, the target file is never locked.
The cache file is only ever read in the `get` function, and there is no
need for an exclusive lock in that situation.
There is therefore no need to obtain any lock, shared or exclusive, to
read the cache file. Doing so only affects performance of the file sytem
as file locks must be needlessly obtained and written to disk for a read
operation which does not benefit from them.
The Redis cache store and session handler both do a 'ping()' after
connecting to Redis. This is unnecessary because the connect() call
has just checked the network connection and it's hardly likely that
the server has gone down since then.
According to my profiling, both connect() and ping() take
measurable time when talking to a separate server, i.e. a few
milliseconds. So it's not the case that connect() doesn't really
talk to the server, as I initially wondered.
If using Redis on a separate (non-localhost) server for both session
and cache store, removing these ping calls can save a millisecond
or two per request.
When deleting a cache instance, it tried to purge it, even though
purge only makes sense for an individual definition (not the whole
instance). As a result, it errored.
Applied the following changes to various testcase classes:
- Namespaced with component[\level2-API]
- Moved to level2-API subdirectory when required.
- Fixed incorrect use statements with leading backslash.
- Remove file phpdoc block
- Remove MOODLE_INTERNAL if not needed.
- Changed code to point to global scope when needed.
- Fix some relative paths and comments here and there.
- All them passing individually.
- Complete runs passing too.
Special mention to:
- When belonging to other components and being valid api:
- analytics related tests have been moved to tests/analytics subdir.
- backup & restore related tests have been moved to tests/backup subdir.
- events related tests have been moved to tests/event subdir.
- privacy related tests have been moved to tests/privacy.
- task related tests have been moved to tests/task subdir.
- Some simple renames, not including the component part anymore (not
needed now that they are namespaced):
- some xxxlib_test.php have been renamed lib_test.php
(when they where testing the corresponding lib.php).
- cache stores tests have been all renamed store_test, originally
each one had its own name (file_test, apcu_test, redis_test...)
- assign feedback tests have been all renamed feedback_test, originally
each one had its own name (file_test, editpdf_test...)
For cache types which mean this information can be obtained without a
significant performance cost (i.e. just by calling strlen and not
having to serialize something that wasn't serialized already),
this change calculates the size of data read from or written to cache
in each request and includes it in the perfdebug table at bottom of
output (when that is turned on).
This supports the following cache types:
* File store
* Redis (only if caching is enabled)
A list of times for each cache key in a TTL cache is kept in a Redis
sorted list, which can be queried efficiently to delete expired
cache items later in a scheduled task.
This change makes set and delete 2x slower (only for caches which use
TTL) but there is no impact on get performance.
It seems that the new phpcs3 checker is now controlling those
line comments that previously were ignored.
This commit just looks for all the cases and bulk-add
them when needed. The bash script (mac) used to add all them is:
while read -r line; do
arr=(${line//:/ })
if [[ -n ${arr[0]} ]] && [[ -n ${arr[1]} ]]; then
echo " file ${arr[0]}, line ${arr[1]}"
sed -i "${arr[1]}s/\$/\./" ${arr[0]}
fi
done < <(find . -name version.php | xargs ag --nomultiline '>(version|requires) *=.*//.*[^;\.]$')
All the setup/teardown/pre/post/conditions template methods
now are required to return void. This was warned with phpunit 7
and now is enforced.
At the same time, fix a few wrong function names,
provider data and param types, return statements...
version = 2021052500 release version
requires= 2021052500 same than version
Why 20210525? (25th May 2021) ?
Because master is going to be Moodle 4.0, to be released
on November 2021. And, until then, we are going to have
a couple of "intermediate" releases:
- Moodle 3.10 to be released 9th November 2020. (2020110900)
This version will be using versions from today to 2020110900
(once it's released the YYYYMMDD part stops advancing).
- Moodle 3.11 to be released 10th May 2021. (2021051000)
This version will be using versions from 3.10 release to 2021051000
(once it's released the YYYYMMDD part stops advancing).
That means that all versions from today to 2021051000 are going
to be used by those 2 "intermediate" releases (3.10 and 3.11).
And we cannot use them in master, because it's forbidden to have
any overlapping of versions between branches (or different upgrade
paths will fail).
So, get that 2021051000, let's add it a couple of weeks to cover
the on-sync period (or a 2 weeks delay max!) and, the first version
that master can "own" in exclusive (without any overlap) is, exactly,
25th May 2021, hence our 20210525.
Thanks to Luca Bösch and Matteo Scaramuccia that detected
these 2 minor, but ugly details in core version files:
1) Adding comments pointed to versions (3.x, 20190801...) is not
recommended, they easily come outdated.
2) Using plugin->release in core is no sense. Core plugins belong
to the core release.
This commit just gets rid of them. That's all.
Significant string changes:
* configcustommenuitems,core_admin - corrected example
* custommenuitems_desc,tool_mobile - fallback behaviour of custom menu
items in the app
