If you’re using this recipe to run WordPress on a virtual machine on the Google Cloud Platform (GCP) chances are high you’re experiencing problems. Here’s how to solve them.
My example is a customer’s website running on a small instance of type
e2-small. This instance only has 2G of RAM and 2 vCPUs.
Main problem of the setup is that it runs Apache webserver using
mpm_worker process management model with PHP integrated as Apache module. This way Apache spawns multiple processes with separate memory areas each running a copy of the PHP interpreter together with all the PHP code needed to executed WordPress. That is quite a lot. For the website i was analyzing Apache had a resident set size of 1,75G without much traffic. So there wasn’t much RAM left for mysql and once some requests came in it lead to mysqld being killed due to out-of-memory errors.
One could argue that running an website on such a small instance was prone for error from missing resources from the beginning. But from my experience i always try to increasing efficiency of a given runtime environment before increasing resources because problems persist and eventually increased resources will be also be maxed out.
So how to improve this setup? Just switch Apache to event based process management
mpm_event and run PHP separately as PHP-FPM using FastCGI as interface between Apache and PHP. This way the memory footprint of Apache decreases dramatically and PHP’s usage resources can be managed independently by configuration.
So here is a quick recipe:
Activate Apache mpm_event and PHP-FPM
sudo su -
systemctl stop apache2
apt install php7.4-fpm libapache2-mod-fcgid
apachectl configtest && systemctl restart apache2
Using this recipe the website i was analyzing went from having no RAM left to ~800M of RAM available. Monitoring will show how the website keeps up with traffic.
I recently implemented JWT (JSON Web Token) based auth on an embedded device using nginx with embedded Lua. Following are implementation details.
- refresh and access token are created as JWT (signed with variable JWT_SECRET)
- refresh token is delivered as cookie and stored on the client. It is valid for 24hrs only for path
- access token is delivered in responses and requests as Authentication-header.
- refresh token is stored on the device in
/media/persistency/www/ is created and owned to web-user at nginx startup by systemd service file.
- refresh of access and refresh token is done via route
/jwt/refresh by verification of the refresh token which is invalidated by using it once.
- refresh token can be verified in 2 variants:
1) by checking the signature with JWT_SECRET
2) if this fails because JWT_SECRET has changed (e.g. due to device reboot) the token value is compared to the token value stored in
/media/persistency/www/refresh_token.txt. If both are equal it is verified.
- additional validations of the tokens
- token type refresh/access is validated so both tokens can only be used for their intended purpose.
- IP of refresh token using client has to be identical to obtaining IP of refresh token which is stored in the token
Free repo for testing servers
To get updates for your testing servers running ProxMox VE you can use following repo which is not meant for production use but works well for testing purposes. Just add the following lines to
# PVE pve-no-subscription repository provided by proxmox.com,
# NOT recommended for production use
deb http://download.proxmox.com/debian/pve buster pve-no-subscription
Uncomment the enterprise repo in
/etc/apt/sources.list.d/pve-enterprise.list and run
apt-get update afterwards. You can now run
apt-get upgrade to get package upgrades.
Resizing disks of Ubuntu VMs
If your VM runs an Ubuntu cloud image you can use the following simple procedure. This probably applies to other Linux distributions as well.
Just resize the block device of a VM on the host using following command e.g. by 20G (106 is the VM ID):
qm resize 106 scsi0 +20G
This should be visible in
dmesg of that VM:
sda: detected capacity change from 45311066112 to 66785902592
Afterwards just reboot that VM and the filesystem will be expanded automatically, again visible in
EXT4-fs (sda1): resizing filesystem from 11033851 to 16276731 blocks
So you have been lured into storing attribute data as JSON in your MySQL to avoid defining a proper schema from the beginning. Well, now for some report you need a certain attribute as column. Here is how to solve this task.
We have following simplified schema:
CREATE TABLE `store` (
`storeId` int(11) NOT NULL,
`attributes` json DEFAULT NULL
With data as in following example:
We may want to extract the value of attribute
representedBy as column in our report. So to extract we use following trick
json_search(attributes, 'one', 'representedBy'),
) as char) as representedBy
How does it work:
json_search(attributes, 'one', 'representedBy') searches the first occurrence of „representedBy“ and returns the path to the key where it is stored
replace we modify the path to point to the value instead of the attributeId
json_unquote we can use that path in
json_extract. The return value is again unquoted and casted to char to have a proper string in the report
Why do we have to use that trick?
We can’t extract the attribute value by using the function
json_extract directly because
- JSON paths which are used in
json_extract for selection don’t provide selection based on keyValues – you can only select by keyName
json_extract doesn’t separate selection from evaluation – what is selected is evaluated.
- Feature request: with both features added we could extract using
json_extract(attributes, '$[*].attributeId="representedBy"', '$.value')
Recently i had to migrate some mailboxes from one imap server to the other. After a bit of research into tools for imap migration running on Linux i decided to use imapsync. It has all the relevant features and is maintained. Running it showed that it tries to do the right things by default. As i will have to do such a migration in the future i document my setup.
imapsync is a great tool but somehow doesn’t come with perlbrew preparation.
imapsync has a lot of dependencies (mostly perl libs) which are installed via the system’s package manager. But sometimes you just want to execute a migration and not have all those libs on the system afterwards. So a Docker setup setup is provided which solves this easily. But Perl has its established solution for this problem, so we want to use this: perlbrew.
So i prepared a
cpanfile which will hopefully be included in the future.
Just save as
cpanfile and run
cpanm --installdeps .
There a good examples for testing the migration of single mailboxes (https://imapsync.lamiral.info/examples/imapsync_example.sh) and doing it in bulk (https://imapsync.lamiral.info/examples/sync_loop_unix.sh). From there you can craft your own scripts.
Don’t forget to test your new mail setup using https://www.mail-tester.com/ 😉
Here you find a quick fix for a wordpress problem i found nowhere written:
I have a wordpress instance running on shared hosting which requires having group read bit set in permissions of uploaded files to be readable by the web server. But uploaded files were set to
0600 instead of
0640 and were thus not viewable on the website.
First i tried setting
FS_CHMOD_FILE but this didn’t work as it is not used for uploaded files.
Then i found that wordpress determines the permission bits of uploaded files by taking the permissions of the target directory and if needed the parent directory of it and removing the execution bits. So to solve the problem just set the permissions of
wp-content/uploads and all sub directories to
0750. This way wordpress sets file permissions for uploads to
0640 as required.
These days i have the pleasure to upgrade an application running on AWS Elastic Beanstalk from platform Multi-container Docker to new platform Docker. The former is based on Amazon Linux 1 (AL1) while the latter runs on Amazon Linux 2 (AL2). I follow this guide.
I have encountered the following advantages so far:
- links between containers are now fully bidirectional: on AL1 those links could only be defined and used in one direction. If you wanted bidirectional communication you had to implement service discovery yourself
- the limit on 10 containers per app is gone: in my tests i could easily spin up 12 containers.
- apparently this limitation came from the automatic translation of an Elastic Beanstalk app into a task definition. The limitation is documented as task definition max containers.
- on AL1 one got the following error:
frontend-staging - ERROR: Service:AmazonECS, Code:ClientException, Message:Too many containers., Class:com.amazonaws.services.ecs.model.ClientException
- i could restart single containers without affecting the other running containers: on AL1 this would cause a restart of all containers for reasons unknown to me.
- simplified logging – all services log to
/var/log/eb-docker/containers/eb-current-app/eb-stdouterr.log which can be streamed to CloudWatch.
Today i had the issue that after dumping and importing a Magento Database from one instance of MySQL into another one the order dates got transported 2 hours into the future. How come?!?
Investigation and testing showed that the source database was running with timezone „GMT“ aka „UTC“ aka +00. The target database was running in system timezone which was CEST aka „Europe/Berlin“ aka +02. Still no problem in sight as Magento anyway stores all values in timezone „GMT“!
Turns out Magento uses data type
timestamp for order dates etc. MySQL documentation states:
MySQL converts https://dev.mysql.com/doc/refman/5.7/en/datetime.html
TIMESTAMP values from the current time zone to UTC for storage, and back from UTC to the current time zone for retrieval. (This does not occur for other types such as
DATETIME.) By default, the current time zone for each connection is the server’s time. The time zone can be set on a per-connection basis. As long as the time zone setting remains constant, you get back the same value you store. If you store a
TIMESTAMP value, and then change the time zone and retrieve the value, the retrieved value is different from the value you stored. This occurs because the same time zone was not used for conversion in both directions. The current time zone is available as the value of the
time_zone system variable. For more information, see Section 5.1.13, “MySQL Server Time Zone Support”.
Exactly this happened: If you store a
TIMESTAMP value, and then change the time zone and retrieve the value, the retrieved value is different from the value you stored.
What was the solution? Well first i wanted to set the target timezone in the dump file. This way target MySQL would convert into GMT during importing and back when the shop reads values. This works except that all values get converted into summer time as it is summer now.
Alternative and better solution: Just set the timezone for each of the shop’s connections by adding the follwing in
<initStatements><![CDATA[SET NAMES utf8;SET time_zone = '+00:00';]]></initStatements>
This way MySQL doesn’t convert timestamp values in those connections and Magento gets GMT as expected from the database.
So in my view there are some interesting observations:
- Why does Magento 1 by design a double conversion of date values by storing them as timestamp in MySQL (except when the server’s system timezone is GMT)?
- Magento creates GMT dates
- storing such a value in MySQL interprets it as in the connections timezone and transforms it into GMT for storage.
- reading is done vice versa: read as GMT and convert into connections timezone
datetime would be a more fitting data type in my view
- Magento 2 consequently fixes this by setting in
$this->_connection->query("SET time_zone = '+00:00'");
- MySQL automagically converts timestamps with the correct timezone offset valid for the indivdiual timestamp in the connection’s timezone
If you want to change MySQL setting
innodb_log_file_size to improve performance the top answer on stack overflow says you have to delete the redo log in
/var/lib/mysql/ib_logfile* so that MySQL creates it with the new size. Before deleting you obviously shutdown mysql.
But a normal shutdown of MySQL is a fast shutdown which might leave changes stored in logs but not in table pages. This is normally OK as mysql works through the logs and processes all changes.
But by deleting the redo log…you might create an inconsistent database! As table pages might be marked
dirty but the changes can’t be loaded anymore from the redo log. I did it many times with fast shutdown not knowing the risk but today mysql got corrupted with the database still working but in inconsistent state.
How to prevent this?
Old answer was: switch off fast shutdown by
SET GLOBAL innodb_fast_shutdown = 0
After shutdown you can now safely delete the redo log. But this is only relevant up to MySQL 5.6.8.
So what is the correct way in modern mysql to change
- shutdown mysql
- start mysql
- documentation says: “If
InnoDB detects that the
innodb_log_file_size differs from the redo log file size, it writes a log checkpoint, closes and removes the old log files, creates new log files at the requested size, and opens the new log files.“
Of course we had backups so nothing happened except for the work.
- MySQL is really robust: it starts and allows creating a dump even with
- corrupted databases
- corrupted mysql database (missing tables etc)
- read the documentation
- read all answers on stack overflow carefully – not just the top one 😉
Just as Chrome has its jumping the shark moment* Microsoft moves to Chromium engine…they don’t have any luck with browsers since 1995 😉
I’m migrating to Firefox…
(*) So what’s wrong with Chrome?
- removing Adblocker API
- flagging slow sites
- hiding parts of the current URL in the address bar – protocol + „www.“ are no longer shown. This can be fixed:
- goto chrome://flags
- Search UI Hide.
- Disable these 3.
- this might be controversial: promoting encryption (HTTPS) in an aggressive way as this is per se a good measure for websites with data entry but is also a burden for website owners.
Many turn to centralized services like AMP by Google and CloudFlare to encrypt their website and make it faster. So these measures weaken an independent web and strengthen big corporations like Google.