Category: Revelation

Finally, something worked right!

The last thing I’m going to be moving from my linode VPS to my colo box is my Navaid.com waypoint generators. I’ve started doing some work on that – originally I was going to export the MySQL database from the linode, and massage them and import them into PostgreSQL on the colo box. But when I first started doing that, I found no end of trouble – the version of MySQL in Debian Sarge doesn’t have the “compatibility” mode in the dump command, plus I discovered that when I’d originally moved from PostgreSQL to MySQL I’d converted all the boolean fields to tinyint(1) or something, and I’d like to change that. Plus there were fields that were set to “not null default 0” which should really have allowed nulls and the like. Basic clean-up stuff, but time consuming. So I’d decided to bring the database over as MySQL, and maybe bring up the site in MySQL and write a conversion script later on so I could convert to PostgreSQL later.

Thursday evening while experimenting with some of the scripts navaid scripts on the colo site, I discovered some bad values. Investigation proved that the FAA had changed the data format for the “APT” airport record in the last load, and I hadn’t adapted my load script.

So Friday evening I went back to the “real” navaid site and fixed the load script, and ran the two scripts to reload the data. I started the run about 8pm. But it was going *really* slow. So while I was waiting, I copied the changes over to the same scripts on the colo site and ran them there. The two scripts took less than a hour to run, which pleased me immensely, and I was able to verify on that site that the damage was fixed. But the scripts on the real site were still running. I waited and finally went to bed. When I got up this morning, the first script was still running. As a matter of fact, it finally finished and went on to the second script at about 2pm. It’s slowly grinding through the second one.

But I didn’t want my “real” site to be down for this length of time. So I hit on an idea – I enabled remote connections to the MySQL database on my colo box, and made a test version of the generation script on the real site with a slightly different “DBI->connect” line as the only change. And it worked, and it worked amazingly quickly. So I changed the whole site over, and restarted Apache, and so now the web services are running on the “real” site, but the database they are hitting is on the colo box. This will make the ultimate migration easier, and it means that navaid.com’s users are already getting a bit of a speed advantage from this move.

The only hitch, and it’s a small one, is there is a script that runs once a night to expire old saved sessions. It uses subqueries, which the MySQL on my colo box doesn’t support. I’m going to have to re-write that as a script that uses a left inner join to find the ones to delete, and then deletes them one at a time. The reason why this worked on my VPS box and not on my colo box is that on the VPS, I’m connecting to a MySQL server provided by the ISP, not my own. So while Debian Sarge installs MySQL 4.0.24, the server I’m connecting to is newer than that.

I thought I was the mightly debugging king…

…but I just handed my debugging crown to him.

Kris and I have been banging our heads on our desks because of problems we’re having with our JTreeTables. A JTreeTable is a class that we found on a Sun Java forum that combines the attributes of the JTable with a JTree – basically giving you JTree behaviour with columnar (table) data. It’s really handy. But frequently, and often in cases I could easily reproduce, the damn thing wasn’t updating correctly. Kris and I both made sure that our updates were properly protected by synchronization locks, and the events were being fired in the event loop, lessons we’ve both learned by hard experience. But it was still acting strangely. Kris spent a lot of time reading forum posts, running the debugger down deep in Java library code, and basically working this problem from all angles for days upon days.

Yesterday he found the problem. And the problem was in my code. When you fire an event, you need to give it an array of Objects that starts at the root node of the tree, and follows down through the tree to the node that actually changed. But of course, when you actually change a node, you’re already at the node that changed, and it’s pretty easy to trace up from node.parent() to node.parent() until you reach the top, so that’s what I do. And then I attempt to reverse the order of what I’ve got to make the required array. But it appears that I fundamentally misunderstood the Stack class, because pushing objects on the Stack and then doing a “toArray” on it doesn’t reverse the order, as I’d thought. So the view was getting a totally messed up event, and that was messing everything else up.

Kris changed my Stack.push into ArrayList.add(0, node), and everything works now. And I never thought about doing it that way because I thought List.add(0, Object) would replace the object at position 0, not push them all up.

And Kris’s small change (after big effort) closes three bug reports assigned to me, and a bunch that were assigned to him.

In my defence, I’d actually come up with the Stack thing while Vicki was driving us to Pittsburgh. So perhaps it wasn’t my best work.

Time well wasted

This weekend, I’ve accomplished two major things of the four or so that I wanted to get done.

Yesterday, I moved my picture gallery from my home machine http://xcski.com/gallery/ to my colo machine at http://gallery.xcski.com/. That was surprisingly easy once I found the part of the Gallery FAQ that showed what I was doing wrong. The biggest glitch is that when I first brought it up, I had the “Square thumbnails” option turned on, and while I turned it off and told Gallery to regenerate all the thumbnails, most of them are still square. Nothing wrong with square thumbnails, but it means that it had to trim the pictures, so for example airplane pictures tend to show the middle part of the fuselage instead of the whole thing, and some full length portraits of people cut off the heads and feet. I may have to try regenerating thumbnails again.

This morning, I woke up to find my newish external USB drive that I use for backing up my colo box is dead, and so last night’s backups didn’t work. When this has happened with my other external USB drive, usually powering it down and up again fixes it, but this time it didn’t. So I did a “mkfs.ext3” on it, and started the nightly backups again. And mid-way through, the logs started filling up with errors saying that the drive was in the process of being unplugged(?!). I rebooted the server for the first time in 66 days, and that seems to have fixed it. Hopefully tonight’s backup (which will end up being huge because the old Sunday night backup is gone) will work.

Today, I’ve mostly been concentrating on trying to restore the archives to my mailman mailing lists. When I moved my mailing lists from my home server to my Virtual Private Server at linode.com, disk space was an issue to I trimmed the archives down to just the last two years. I kept the full archives on my home server just in case. And now that the lists are on my colo box, disk space isn’t an issue any more and so I tried to restore them.

My first attempt, using vim and a split screen with the old archive (which goes up to mid 2005) and the current archive (which starts in January 2005) and attempting to cut the pre-2005 stuff out of the old one into the new one didn’t work very well. I quickly bogged the machine down in extensive swapping.

So then I cut the old archive down to stop at the beginning of 2005 using

sed -n '0,/^From.*2005/p' < old.mbox > old.mbox_to_2005
head -n -1 old.mbox_to_2005 > old.mbox

and then catted the old and the current one together.

That’s when I discovered that mailman’s arch program, that regenerates archives from the mbox files, is a huge memory hog and also has a couple of bugs. First couple of times I tried to run it, it processed a few thousand messages and then died with a message about an empty module name. When I realized it was dying on the same message both times, I discovered that back in 2000 one of the mailing list users had a weird-ass bug in their mailer that was sending email with the header

Content-Type: TEXT/PLAIN; charset=".chrsc"

Evidently some sort of misconfigured character set. I used sed to change that to “us-ascii” and arch seemed a lot happier. At least until it happily consumed all the ram and most of the swap on the system. Everything dragged down to slower than a very slow thing.

I found some awk code to split up mbox files into smaller chunks, and set it to run on this huge unified archive, and then ran arch on the chunks. That mostly worked, and only slowed the server down to a not-very-slow thing, except that arch did the wrong thing on any line that started with “From ” that wasn’t the start of a mail message. I didn’t discover this until it had been running for quite some time, so I had to start again.

I used my extensive knowledge of awk (in other words, I cargo culted something) to make the mbox splitter also change any of these bogus “From ” lines into “>From “. After another hour or so of running I discovered a small bug in my splitter that meant it worked for early archives and not for later archives, probably due to a mailman upgrade or when I switched from sendmail to postfix.

So I fixed that bug, and started again. It’s been running for over an hour now, and seems to be working fine. Well, except for 5 or 6 messages that came in January 2000 that had the year set to “100”, and one place where somebody actually quoted a mailbox header without putting a “>” at the beginning. Minor inconveniences.

I guess tomorrow morning I’ll have to check that the archive is correct, and the nightly backups worked. But for now, I’m going to bed.