On scalability of MySQL

Anyone who says that MySQL is not scalable has no idea.  Facebook is one of the examples for a large deployment of MySQL:

How big is Facebook’s Internet infrastructure? Facebook VP of Technology Jeff Rothschild provided some details in a panel at the recent MySQL user conference. Rothschild says Facebook is now running 10,000 servers, including 1,800 MySQL servers that are overseen by just two database administrators.

Facebook recently surpassed 500,000,000 users – half a billion!

The saga of a lost connection

I came across an interesting problem at work a couple of days ago.  One of our CakePHP-based applications has a scheduled task (cronjob) that runs every minute and imports mail from a number of IMAP mailboxes into the database.  All of a sudden the script broke.  In a very particular way.  Not a single line of code was change in the script itself, and there were no significant changes in the rest of the application that the script is using.  But one day, for some reason, it started to import emails from only half of the mailboxes, completely ignoring the rest.

A lengthy midnight drunk debugging session helped to realize the problem.  Apparently, in the list of mailboxes to check there were a few old mailboxes, which were not used anymore for any email.  The script was checking them but they were always empty.  Server administrator removed those mailboxes during a scheduled maintenance window.  Now, the script couldn’t connect to the mailbox anymore.  No problem, there was error handling for this case.  But was error handling didn’t take into account is the combination of time it takes timeout on the non-existing mailbox check and the database connection timeout.

It so happened, that these few old mailboxes were right in the middle of the list of all mailboxes.  The script was importing mail from the first few mailboxes just fine.  But then while it was taking a long time to timeout on a few non-existing mailboxes, the database connection got closed due to the inactivity timeout.  This wasn’t handled properly.  The script was only checking for the existing database connection at the beginning, when it was opening the connection – not later on.  As the result, nothing could have been imported from the rest of the mailboxes.

As always, once you know the problem, you also know how to fix it or work around it.  But troubleshooting a problem like this is tricky.  Seeing the queries working for one mailbox and not working for another is disorienting at least.  Having no recent changes to the application doesn’t help either.  And the problem being not in the code itself, but in the resources is also not very obvious.

So, here you go.  If you ever have a weird problem like that – check that your resources (database connection, network connection, file handler) are still available to you.

CakePHP + GraphViz = making sense of a numerous models

NOTE: THIS IS VERY MUCH OUTDATED! Read about the update or go directly to GitHub repository for the new version.

 

I have a task at hand.  I have to re-introduce myself to a rather large codebase.  It’s a project that migrated to CakePHP a couple of years ago and haven’t seen since.  There was a whole team of people working on it sense then, and now I need to make sense of all those changes that were done and help reorganize and refactor them a bit.  When I looked into CakePHP’s models/ folder, I was surprised to find 50+ models there.  Each and every one of them links to other models.  And documentation is practically non-existing.  How do I go about it?  I hack up a little script to help me out.

There is a really elegant and beautiful tool for graphing things – GraphViz.  If you haven’t heard of it, you need to drop whatever is that you are doing and familiarize yourself with GraphViz.  Right now.  Right.  This. Second.  You are missing out a whole universe until you do so.  I’ll wait.

Now that you are back, I just want to mention a very slick tool, which is a part of GraphViz package – dot.  It is a simple language in which you can describe graphs.  Sort of like “A goes to B, which goes to C”.   You specify your graph in a very human readable format in a text file, and then dot will transform that text file into an image format of your choice (PNG, JPEG, GIF, etc).  The primary beauty of this is that those text files can be generated automatically by using pretty much any programming language.

So here is what I did.  I assumed the following:

  1. Project documentation should be in app/docs/ folder.  That’s where I’ll put the script and that’s where it will generate the dot configuration, dot later will generate the graph of all my models and their relationships.
  2. Main project application folder is app/.  Models are stored in app/models/ folder.
  3. Project can have a number of plugins, which can have their own models, which I still want to know about.  Plugins are in app/plugins/ and if plugin xyz has models, they are stored in app/plugins/xyz/models.
  4. My project is under version control.  Specifically I use Subversion, but it’s easy to adjust the script to support other systems.
  5. I can get current project revision by running a command in shell.  For Subversion that is /usr/bin/svnversion.

I probably assumed a whole bunch of other things, but you can get an idea of how simple the setup is from the above ones.

Here is how I generate a graph of all models and their dependencies:

cd app/docs/
php -f graph.php > graph.dot
dot -Tpng graph.dot > graph.png

Obviously, I can’t show you the full graph from that system (it’s not open source, it’s not mine, and it will drive you insane in a matter of seconds), but here is how a small part of that image looks like.

There is a different colour for each type of model relationship ($belongsTo, $hasMany, and $hasAndBelongsToMany). Each model folder is in a separate sub-graph cluster. There is a legend graph on the image. The current time stamp and version control revision are also imposed on the image for easier referencing.

And here is the source for the graph.php script. Feel free to modify any way you like. If you spot any major bugs or better ways of doing things, please let me know in the comments.

Continue reading “CakePHP + GraphViz = making sense of a numerous models”

Oracle and PHP – the deadly mix

WI’ve spent most of the last week getting into, around, and out of the issues related to interoperability of Oracle and PHP.  Before you start laughing, cursing, and blaming, Oracle wasn’t my choice of the database for this specific project.  It’s just the company already had it installed and working for the background, and there needed to be some integration with the front, which is of course MySQL and PHP based.

First thing I do, obviously, is visit PHP.net to check for the prefix of the functions that I need for Oracle.  Through out my experience with PHP, that’s about the only thing I need to know to start working with the new database.  Oh, and the PHP module installed to provide those functions. Oracle interface for PHP is called is called OCI8.  All you need to do now is install the oci8 module.

Here comes the first trouble.  oci8 is not provided as a pre-compiled package for Fedora Linux.  There is an alternative yum repository – Remi, which has oci8 RPMs, but first of all, the oci8 module is compiled against somewhat outdated Oracle headers (version 10.2.0.4 instead of the latest 11.1.0.1), and it also needs to replace your native PHP and MySQL packages.  I tried that, and it sort of worked, but I wasn’t happy.  So I got my Fedora packages back and decided that I need to compile oci8 myself.

In order to compile oci8, one needs to download Oracle InstantClient (basic package) and some header files (devel package).  These can be downloaded from the Oracle web site, for free, minus the time for the registration.  The little trick here is that during oci8 compilation process, the includes are searched from locations which do not include the one from Oracle RPM.  I did a simple symlink of the includes folder to where Oracle headers were, and compilation went on just fine.  (Hint: otherwise you’ll get a whole lot of Zend related messages and a fatal error).  Gladly, I only had to do this path correction on the Fedora 9 machine.  My production server with Red Hat Enterprise Linux 5 compiled oci8 without any problems all by itself.

Update: more detailed instructions on the actual installation can be found here and here.

Now that oci8 installed and configured, I spent some time figuring the correct way to specify the DSN.   Oracle uses some weirdly name file (tnsnames.ora) in some weird location, but luckily there is a way to go around it.  More so, I recommend that you remove tnsnames.ora file altogether, since it can add to your troubles.  For example, if you mix spaces and tabs as whitespaces in that file, you are screwed.  So, just get rid of it.  The way you specify DSN is directly in the PHP script, and you use the syntax like so:  “//hostname.or.ip:port/dbname“.  Intuitive, I know.

Once you’ll get connected to the server, you have a whole bag of surprises waiting for you.  That is if you are too used to working with MySQL.  First is the syntax.  Oracle is using PL/SQL, so you wipe the dust of from that really old Pascal textbook that you have somewhere.  “begin :result := some.procedure.call(:param1, :param2); end;” – that sort of thing.  Secondly, you’ll be happy to know that prepared queries are supported.  So your workflow will slightly change.  Perl programmers will feel more at home here.  oci_bind_by_name() and oci_execute() are your friends here.  Oh, and while you are at, get familiar with the types of the parameters, because they are important.  And don’t forget that you’ll have to bind each and every variable in the query, or get a fatal error. And since you are learning something here, get ready for the oracle errors.  The most frequent one you’ll get would be something like “Failed to retreive the error message for ORA-12345”, where 12345 would be a number of the error.  So you’ll google for ORA-12345 and ORA-54321 and ORA-XYZZZ a lot.  But than you’ll have a wrapper library and you’ll be OK.

Update: as was noted in the comments, PL/SQL is just an option, not a requirement.  Also, most of the headaches of the above paragraph could be avoided by using one of the PHP frameworks.  I personally haven’t yet tried the framework yet, since I’d like to see things working directly first.  Especially since we are not in the test mode only.

The bigger surprise is still waiting for you though.  You are very likely to discover that OCI8 implementation for PHP is very slow.  And I do mean extremely very slow.  I couldn’t believe that it could be slow, so I went into the source code and OMG!  It is really slow.  The slow part is around fetch_all() against fetch_row().  Basically, it’s always row by row and never all, even if you tell it how many rows you need fetched.

In my case, I have the server a bit far away, and there is a possibility to get many rows back.  So even for a simple query with 140 rows in results I was getting 20 seconds execution time.  Oracle was serving results fast, the network was OK, machines on both sides were powerful and all, but it was still taking 20 seconds or more.

I am still trying to find the solution to this issue, but so far it seems that the current way I do it will be the way to do it.  And the way I do it now is the following.  Never ever run direct SQL queries.  Everything goes through a stored procedure.  The results are returned all in a single row.  And that single row has the BLOB (CLOB actually) with all results in one single XML.  Fetching works good enough to get it, and then parsing is done with one of the billion XML parsers for PHP.

In my case MiniXML worked pretty good until bigger results started coming in.  That’s when I learned an important lesson.  MiniXML parses XML with a regular expression.  PHP has a couple of settings in the configuration file that limits the size of the memory and recursion during regex parsing – pcre.backtrack_limit and pcre.recursion_limit.  If you really want to kill your server, set these to -1 (instead of default 100000) and try a regex against a 1 MB XML file.  Enjoy, cause it won’t be long before everything goes down. I didn’t feel like changing from MiniXML so we just implemented some limits in the queries and stored procedures on the Oracle side, and add a few checks in PHP fail rather than crash the system.

So, to some it up, here is my experience with Oracle and PHP from the last week:

  • I had to register on Oracle web site to download packages
  • I had to re-learn my long forgotten compilation skills
  • I had to go read some C
  • I had to step on the “re-inventing the wheel” path more than once
  • I am parsing XML when working with the database
  • I had a head ache more than twice
  • I didn’t have much fun
  • After all, it works.  Sort of.

One last point in this saga is about Googling.  Ask me any question, and I do mean any question, about MySQL.  Heck, even PostgreSQL.  And the answer is just there, on the first page of Google results.  In any human or programming language.  For any operating system.  You’ll be sorted out and working in less then a minute.   Then, try asking even the simplest of the simplest questions about Oracle and PHP.  Sometimes you’ll find something.  Some other times, you won’t.  The overall feeling I have is that not a lot of people are using Oracle with PHP, and those of them who do are in their majority not very happy.

Now I’ve joined the army.

InnoDB is better for SugarCRM

After SugarCRM was deployed, we were experiencing some lock ups.  Not frequent or dangerous, but annoying.  About once a week or every ten days or so, SugarCRM would lock up and won’t answer any queries at all.  Not even the login was possible.  A brief investigation showed that somehow it was locking up the MySQL database – about 15 processes (using “show full processlist”) in Locked state, with no data being sent back or forth.  All locked queries were rather complex, with several JOINs.  The load on the system was somewhat high, since we have about a few dozen operators working on it at the same time.

A similarly brief Google search suggested (see here and here) and explained converting MySQL tables from InnoDB to InnoDB.  A test has been performed and everything went OK.  Our SugarCRM database is about 600 MBytes and it was converted from MyISAM to InnoDB in under 20 minutes.  The best part is that it takes even less to convert back to MyISAM, in case you change your mind.

It’s been a few days now since we did the conversion and it looks OK.  Also, the CRM itself feels a bit faster.