Blog of Leonid Mamchenkov

We’ve been doing some interesting things at work, as always, with yet more people and Linux boxes.  And of the side effects of mixing people, Linux boxes, and several locations is this need for some sort of centralized logging.  Luckily we have either syslog-ng or rsyslog daemons installed on each machine, so the only two issues seemed to be reconfiguration of syslog services for remote logging and setup of some log reading/searching tool for everyone to enjoy.

As for log reading and searching, there seems to be no end of tools.  We picked php-syslog-ng, which has web interface, MySQL back-end, access control, and more.  There were a few minor issues during setup and configuration, but overall it seemed to be OK.  I also patched the source code a bit in a few places, just to make it work nicer with our setup and our needs  (both numerical and symbolic priorities, preference for include masks over excludes, and full functionality with disabled caching).  In case you are interested, here is a patch against php-syslog-ng 2.9.8f tarball.

Once everything was up and running and we started looking through logs from all our hosts in the same place, there was one thing that surprised me a lot.  Either I don’t understand the syslog facilities and priorites fully (and I don’t claim that I do), or there is just too many software authors who don’t care much.  Most of our logs are coming in at priority critical.  Even if there isn’t much critical about them.  Emergency is also used way too much.  And there is hardly anything at debug or info or notice levels.  (RT, SpamAssassin, and many other applications seem to be using critical as their default log level).  Luckily, that  almost always is trivial to fix using either the configuration files or applications’ source code directly.

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.

We spent a couple of weeks now at work, setting up our Nagios installation.  This is one of those things that one can never finish.  Monitoring is like security - no matter how good it is, it can always be better.

One thing that I particularly enjoyed figuring out was how to monitor dynamic stuff.  Hosts and services are the easy part - they are always there and should be there.  If they aren’t - let me know.  But how to monitor dynamic values, that change based on the time of day or day of week?  How to configure the monitoring so that we don’t need to update the limits every other week?

For example, consider the number of the user registrations through a web form.  If we are to measure this number over any sensible (monitoring-wise) period of time, such as one hour, we’d see that it fluctuates a lot during the day.  Furthermore, this number fluctuates differently depending on the day of the week.  We have three separate pick hours during the day, and we have a great decrease in numbers over the weekend.  Plus the amount of user registrations is linked to all the advertising that company does, so this week might different from the last week and from the next week.  How can we measure it so that a notiification is sent when the number is abnormal?  What is abnormal?

The solution turned out to be much simpler than I originally thought it would.  It is sufficient to get a few samples of the data in same hour last week and a week before.  If current value is more than twice the maximum or less than twice the minimum from the sample data, then we should be notified. This, in fact, works pretty well.  The only time when we get a lot of false positives is when the values in the sample data are small.  With values under 10, it’s very easy to jump over or under the limit.  When the sample values are higher, there is more space between the boundaries and the system works as expected.

We’ll get some more sample data now and we’ll be adjusting the formulat accordingly.  But as I said, even as it is, it’s pretty good.

Once again, today is the last Friday of July, which means that today is Sysadmin’s Day! Congratulations to all system, network, database, etc administrators out there. Today is your day. I wish you all receive some appreciation from your users and management. Don’t stay in the office for too long. Go out and celebrate!

My boss has left the company. I had to delete his account and clean some groups and other access lists. It felt very weird. It felt like a beginning of the end. Although it’s not.

Not to mention that removing one of the global administrators, who had access virtually to every corner of the network, is not a trivial job…