Thursday, April 6, 2017

About weasyprint, pango and cairo

I explored why we have these warnings (#1462):

WARNING: @font-face support needs Pango
WARNING: There are known rendering problems with Cairo <= 1.14.0

You can run the following to see them:

$ python -c "from weasyprint.text import pango; print(pango.pango_version())"
WARNING: There are known rendering problems with Cairo <= 1.14.0
WARNING: @font-face support needs Pango >= 1.38

It is true that the warnings do not appear on my machine anymore. On my machine I have:

$ python -c "from weasyprint.text import pango; print(pango.pango_version())"

I tried whether it is a solution to run “pip install -U cffi cairocffi” on that machine:

$ pip install -U cffi
Collecting cffi
  Downloading cffi-1.10.0-cp27-cp27mu-manylinux1_x86_64.whl (392kB)
    100% |████████████████████████████████| 399kB 1.3MB/s
Requirement already up-to-date: pycparser in ./env/lib/python2.7/site-packages (from cffi)
Installing collected packages: cffi
  Found existing installation: cffi 1.9.1
    Uninstalling cffi-1.9.1:
      Successfully uninstalled cffi-1.9.1
Successfully installed cffi-1.10.0

Nope, that doesn’t change it. On SR (jessie) I tried this:

$ sudo apt update
$ sudo apt upgrade

Which also didn’t help. So Debian Jessie is not enough, but Ubuntu Xenial is.

Switch from MyISAM to InnoDB

I ran, verified that they are currently using MyISAM (see Which engine am I using?), then uncommented the OPTIONS from their DATABASES and then restored the snapshot.

This first restore took 2 hours and a half.

I think that this can be reduced drastically by temporarily disabling certain constraint checks. TODO: find out how to do this.

Optimizations for Jane

While waiting for the restore to finish, I worked on #1685.

  • I reorganized the WishTypes choicelist, adding new types “Agenda item” and “Resolution”.

  • “Activities” is better than “Events” (as verbose_name of courses.Course)

  • lino_xl.lib.deploy.desktop.DeploymentsByTicket now shows only active milestones. To implement this, I added a new feature to the model API: add_param_filter.

  • I started a new tested document deploy : Deployment management.

I released these changes to Jane.

Quick optimizations for weleup

When the restore had finally finished, I tried whether #1649 is fixed, but stumbled into an AttributeError “‘NoneType’ object has no attribute ‘mail_mode’ in lino_xl.lib.coachings.mixins” (und Frau Marthà Musterfrau hat eine Begleitung 27.03.2017 DWEF ohne Begleiter).

I fixed this bug, then pulled the newest version. Which caused another, more serious problem: cal.Room is now ContactRelated. So I have to restore my snapshot again! Another 2,5 hours? No, I preferred to find out how to make the restore quicker.

At 6:00 I started a second restore with the new version, but this time with FOREIGN_KEY_CHECKS disabled.

  • 12 additional minutes because I renamed EventStates to EntryStates.

While this restore is still running, I tried on Jane how much time I would gain. Ouch, a restore of Jane takes 4 minutes independently whether FOREIGN_KEY_CHECKS is disable dor not:

real        3m55.933s
user        2m18.828s
sys 0m3.348s

So my trick doesn’t work. After reading this, I changed the Python wrpper for mysql from mysql-python to mysqlclient:

$ pip uninstall mysql-python
$ pip install mysqlclient

No difference.

I started to worry, because at 6:25 their cron job starts a snapshot. So I disabled their /etc/cron.d/lino_backup just in case…

But then I was lucky. Despite my prognose the second restore went significantly faster:

real        42m20.987s
user        22m8.216s
sys         0m25.992s

I guess that actually I had bad luck the first time. Their nightly backup was running and had caused the restore to take 2,5 hours.

TODO: find out whether we must/can optimize the speed of a database restore. 40 minutes is actually okay when you have a whole night. It seems that InnoDB is slower than MyISAM for appending new rows.

HO with Tonis

Tonis and I talked about (1) deployment strategy and (2) planned the coming sprint for Jane. We did a release together because we were exploring #1689 (which is still not fixed).

In many projects the master branch is stable and each team member develops in separate branches and one of them merges the branches into master. In our current deployment strategy, Luc is using master as development branch because he is old and lazy, or rather because he currently maintains all our production sites and because these are configured to pull master.