Musings of an Anonymous Geek

I have a small EC2 instance running with a 25GB EBS volume attached. It has a database on it that I need to manipulate by doing things like dropping indexes and creating new ones. This is on rather large (multi-GB, millions of rows) tables. After running one DROP INDEX operation that ran all day without finishing, I killed it and tried to see what was going on. Here’s the results of the first 10 minutes of testing:

-bash-3.2# dd if=/dev/zero of=/vol/128.txt bs=128k count=1000
1000+0 records in
1000+0 records out
131072000 bytes (131 MB) copied, 0.818328 seconds, 160 MB/s

This looks great. I’d love to get 160MB/s all the time. But wait! There’s more!

-bash-3.2# dd if=/dev/zero of=/vol/128.txt bs=128k count=100000
dd: writing `/vol/128.txt': No space left on device
86729+0 records in
86728+0 records out
11367641088 bytes (11 GB) copied, 268.191 seconds, 42.4 MB/s

Ok, well… that’s completely miserable. Let’s try something in between.

-bash-3.2# dd if=/dev/zero of=/vol/128.txt bs=128k count=10000
10000+0 records in
10000+0 records out
1310720000 bytes (1.3 GB) copied, 15.4684 seconds, 84.7 MB/s

So the performance gets cut in half when the number of 128k blocks is increased 10x. This kinda sucks. I’ll keep plugging along, but if anyone has hints or clues, let me know. If this is the way it’s going to be, then this is no place to run a production, IO-intensive (100,000s and maybe millions of writes per day, on top of reads) database.

Thought I’d throw this out to the lazyweb for advice. I’ve been running OS X as my primary desktop OS since I got my macbook pro around May 2006 or so. I installed Parallels as soon as I found it, and it was great. I still have it installed. However, I’ve found two issues with Parallels in the past two years:

1. There’s no kind of easy upgrade path for existing parallels users that I can find. Their web site is a mess, (like most software vendors, oddly) and there’s no upgrade license that I could find.

2. It’s non-trivial to install of versions of linux that are released after the version of parallels you’re trying to install to, for whatever reason.

I remember having issues with installing an Ubuntu version, and a CentOS version, both newer than the parallels I had installed. Then I upgraded parallels, and they installed fine.

So today I tried to do the impossible: install CentOS 5.2 on a build of Parallels from March 2007. I get a kernel panic, which I remembered a hack for (add “linux kernel agp=off” to the boot: line), but even that didn’t have any effect. Guess I just pushed the limit too far this time.

So, since there’s no upgrade license to make it cheaper for me to stay with Parallels, I have no particular reason to stay loyal to them. However, I cannot find a single mention of Linux on VMWare Fusion’s web site (?!), so….

What’s everyone doing about virtualization on the OS X (Leopard) Desktop?

I got to actually speak to Brian Aker for maybe a total of 5 minutes after his micro-presentation about Drizzle, which took place at the Sun booth at OSCON 2008. I was a bit nervous to ask what questions I had out loud, because the things I had wondered about were things I really didn’t see too much discussion about out in the intarweb. I’m happy to report that, if Brian Aker is to be considered any kind of authority (hint: he is), my ideas are not completely ridiculous, so maybe I’ll start talking a bit more about them.

UPDATE: lest anyone get the wrong idea, Brian Aker did, in fact, state that views are not on the short list of priority items for Drizzle, but he did say that views are one of the features he finds most useful, and that they’d probably be higher on any future priority list than, say, stored procedures. So, take my notes below about views with a grain of salt. It’s not necessarily “coming”.

My three ideas were these:

1. Materialized views: my experience with views in MySQL is that they just plain old don’t scale well compared to other database systems I’ve used. I used Sybase in 2000 and views scaled better for me then than MySQL views do now, and I’m using them in mostly the same way (which is to say that I’m not using them to do evil things - I’m using them in the way most of the database community agrees they should be used). In the past, I thought materialized views were “nice to have”, but now that I’m working with much larger data sets, without a need for my reporting to be 100% real-time, materialized views would be great. To be honest, I win either way with Drizzle in all likelihood, as Brian Aker has proclaimed that views in Drizzle will not look like views in MySQL. He confirmed that materialized views would be a great thing to have a closer look at, and I was happy with that reply.
2. “Query Fragments”: I didn’t know they were called query fragments. What I explained to Brian Aker was that I wanted to harvest subsets of cached result sets. So, for example, if I do a date range query (I do that a lot), and the result set is cached, and then my app does another query which is identical save for that the date range is a subset of the one in the cached result set, I’d like to grab that data from the cache. My actual question to Brian was “can this be built in such a way that this would be a reliable, trustworthy result coming from some middle tier component?” And that’s when he told me about query fragments. He said that this idea was not at all crazy, and was also worthy of further discussion.
3. Vertical data-padding. Well… that’s what I’m calling it. Here’s the logic: I have lots of temporal data. Tons of it. My queries are largely things like “for this user, show me all of the foo’s bar’d — group by day — where day is between these two days”. MySQL is good at these kinds of queries (assuming you index well and, basically, that you’ve read “High Performance MySQL” a couple of times… per edition), but there’s something missing. When I get the result set back, any of the days for which no foo’s were bar’d, I don’t get a record. This is perfectly within the realm of reasonable behavior, but I’ve never known MySQL to let things like being reasonable get in the way of helping their users, so my suggestion was that MySQL, in order to do the comparisons for the “WHERE” clause, *MUST* know what dates fall within the dates in, say, a “BETWEEN” statement. It would then seem logical to have some way to tell MySQL “If one of those dates has no value, return a NULL” (or 0, or an empty string, or something). I don’t know the real name for this proposed feature, but I call it “vertical data-padding” because you’re padding columnar data, which, in my mind, is visualized vertically. Just like when you do a “GROUP CONCAT” or something, I would refer to that as horizontal data padding. I explained to Brian that one way I’ve seen this handled is to have a lookup table of static dates that gets joined to the main data table. You do a left outer join with the date lookup table on the left, and you get back a row for every date whether there’s data in the right-hand table or not. This works when ‘n’ is small (like everything else), but it’s hurrendous when you have, say, 10 million rows to deal with. Then you’re in what I call the “No Bueno Zone”. Brian seemed interested in the problem, and I’ll be discussing that with him further when his life settles down a bit (he’s been at OSCON, and he’s still settling in at Sun).

I want to thank Brian Aker for his enthusiastic attitude toward helping others, and for all of the work he pours into all of this stuff. I also want to say that, for me, Drizzle is really exciting, not so much because the feature set is more or less cherry picked to map onto what I do for a living, but also because it represents an opportunity to get ideas in the door before a lot of legacy cruft makes it impossible to implement these somewhat idealistic features without rocking the boat for the millions of users already launching it into production.

Launching a Startup in 3 Hours was a great talk given by Andrew Hyde (of techstars.org) and Gavin Doughtie (of Google). Both of the speakers are heavily involved in the recent trend of doing “Startup Weekends”, and techstars.org is an organization that hosts startup weekends all around the US (and I think internationally as well - Andrew mentioned one in Germany if I heard correctly).

The first half of the talk was about the general concept of a startup weekend, the problems it avoids (”we’ve been working for 9 months and haven’t launched anything”), the problems it brings up (”If you’re not using Java, you’re an idiot, so count me out!!”), and lots of details about how to organize, how to assign roles, and some common tools they use (like Basecamp and whatever your IM of choice is). There was also talk of legal issues, how (basically) to think about forming the company with the people involved, and decisions that need to be made at a business level aside from just the coding.

The second half of the talk wasn’t a talk at all. Instead, people who had ideas stood up, presented their idea in a couple of sentences, and once the ideas were out there, we were told to break into groups and get to work! So people would get up and move over to the person whose idea they liked, and they’d start brainstorming. I decided to head out after about 30 minutes of observing and talking with people about ideas, but when I left, there were probably 6-8 groups of people engrossed in conversations, and the energy level was very high. Overall, it was a really exciting experience!

I think I have pictures of most of the basic parts of the conference at my OSCON Flickr set, and I thoroughly enjoyed day 1 of the conference. Of course, while *day* 1 is over, *night* 1 has yet to even begin. There are lots of BoF sessions, and maybe even more smaller meetups going on, as smaller groups take to discussing things over dinner and a beer or three.

I have to say, that I occasionally pop into irc channels for conferences I’m not even at and follow up on that because I’m involved a bit in conference planning as part of my work with Python Magazine (I’m helping to organize the PyWorks conference in November). This conference seems to have a pretty happy audience, if IRC chatter is any indication (and it usually is). Sure, there are a couple of weak spots in the wireless network, there are some fuzzy projectors, and there was a little confusion regarding breakfast this morning, but the important bits have been well-covered by the OSCON organizers and the “boots on the ground” here on site. Kudos to them all.

This afternoon I hopped to a couple of different talks: one on Memcached and MySQL, and the other on A/B testing. Both contained good content. Of course, I’m a systems guy primarily, so I sort of wanted more of an overview of memcached from the point of view of an admin who is deploying it rather than a developer implementing their code around it. I still got plenty of value out of that talk, and this *is* really more of an open source *developer’s* conference, so the expectations of 99% of the people in the room were met, I’m sure.

A/B testing is just not an exciting topic, and I would imagine that peoples’ bosses made them go to that talk whether they liked it or not. Not to say the talk wasn’t good - the parts I saw (I came in after the break) were good, and I learned from it, and that was the goal. If you’re a QA/QC person, I’m sure the talk was riveting, and there were a lot of good ideas and things I’d never considered flying by in the slides.

Overall, Day 1 is a win. I’ll cover more about this evening’s events in the pre-breakfast hours tomorrow. Stay tuned!

I’ve posted a picture of the BoF board for day 1. Click on it to see bigger sizes. The full size image (maybe smaller) is perfectly suitable for reading at your leisure. I’ll update this if/when I see significant changes to it:

The morning session I attended was “Python in 3 hours”, which provided excellent coverage of Python for folks who may never have seen it before. Steve Holden, being more qualified than most to present this material, did a stellar job of not only presenting the material, but also addressing questions coming from the attendees.

Steve’s talk is fantastic, but I really wish there was an introductory talk for folks who were coming in with the question “Why Python?”. Steve’s talk assumes interest in the language, and that’s completely fair, but I brought a buddy who I think was looking more for the “why” than the “how”. Of course I try to evangelise the language where I can, but to be honest, I’m not as good a salesperson for Python as I wish I was. I use Python for all of my sysadmin tasks these days and love it, but I already hated the other languages I was using, so I guess I just wasn’t a very hard sell.

I got an early start. Too early. But I’m from the west coast, so my body thinks I slept in. I wandered around a bit, took a few pics which you can see at my Flickr OSCON set, and I discovered a couple of things that might be of interest:

• The starbucks in the conference center charges over $2 for a small cup of joe. There’s a starbucks right across the street (you can see it from the breakfast area - seriously, it’s 5 seconds away), and they charge less than $2 for a medium (grande). That’s less than I pay at home.
• The ATM outside the starbucks charges $3 for cash. I’ll report back when I find a cheaper one, but most places seem to take plastic here.
• Every computer involved in this conference, from registration to the video screens that dot the common areas, are running Windows XP. Just sayin’.
• The light rail system is free to go just about anywhere except for the airport, so there’s no excuse not to get out and see Portland and take in the food and beer and stuff.
• For beer-lovers, not only is there the Oregon Brewers Festival starting at the tail end of this conference, but there’s apparently another festival that we missed *last* weekend!! Keep that in mind when you’re planning to come to OSCON next year.

If you’re attending the 10th Annual Open Source Convention, I’ve compiled just a few tips for you on this, “day 0″ of the event:

• Don’t check bags. Everything is slower if you check bags, and if you’re packing more than three shirts, you’re crazy, because if history is any indicator, you’re going to be bombarded with shirts over the course of the week. One maximum size (22″ x 14″ + 9″) suitcase, and a bookbag with a laptop pocket is all I brought, and I’m confident I’ll have all I need. I’ll report back if things change
• Request a room away from the ice machine. They can be loud. This year my room is the last room at the end of a long hallway. Ahhhhhhhh….
• Don’t bring toiletries of any kind: you can’t bring a lot of them on board, and I’d rather just avoid it altogether and buy stuff when I get to my destination. Don’t use the Hotel store though - there’s a Dollar store about 2 blocks from the Lloyd Center Doubletree Hotel (on the back end of the Lloyd Center mall), and they probably have everything you’ll need. If not, walk another block north to the Safeway, and you can get anything, though I didn’t find any travel-sized stuff.
• Show up to registration early: I’m leaving shortly for registration. Registration moves pretty quickly even if you go on Monday morning, but on Sunday night (from 5-7pm) there’s a nice, jovial, laid-back mood around the registration areas.
• When you’re in Portland, know that you’re in an area that is something of a mecca for beer. Even if you don’t like beer, I urge you to join friends and at least have a look at the beers available. You’re in an area where even the hotel bar has an ok beer selection. Saying you don’t like beer is like saying you don’t like food. If “beer” to you means Coors Light (or similar), you have no idea what beer is - but that’s ok, because you’re now in a place that can grant you a PhD in beer snobbery in a matter of a weekend. Really. Take advantage of it!! (a hint: many people who “don’t like beer” really just don’t like the bitterness that comes from hops. Ask a bartender for a sample of their finest wheat beer. I’ll bet you’ll be hooked).
• Don’t stay in your hotel room if you can help it. Engage. Look at the whiteboard that is probably in the registration area as I type this. Find the conference web site, irc channels, wikis, and everything else that you can. 75% of the value of coming to OSCON is finding and meeting people you’ll be in contact with well after you leave. It’s a commercial conference, yes — but it’s a community atmosphere.
• Plan your day. You can try to plan everything you’re going to attend before you get here, but it probably won’t work very well, because you’ll inevitably hear someone talking about something else and decide to attend that instead. What might work better is if you try to plan the night before — but not after the parties — probably sometime between the last session of the day and dinner. At least have an idea what you’re doing the next day, because parsing the program on-the-fly is, imho, difficult, especially when ten people you know walk by and say hi and stuff.
• Try to plan lunch in the city. This can be a little difficult, but you can hop on the light rail for free as soon as the conference breaks for lunch, and be downtown in no time. Last time I attended, I only made it out for two lunches downtown, and I’m kind of a foodie, so I would’ve liked to sample more of the local faire. Try to keep away from the chains (you can get that at home) and be adventurous!!

I’m working on a new project that will be open sourced if I can ever get it to be generically useful. It’s called “sarviz”, and it’s a visualization tool for output from the “sar” UNIX system reporting utility. I know tools like this exist, but please read on, as I’m looking to do something a bit different from what I’ve seen.

A quick, simple explanation of sar

System administrators typically run sar as a cron job, and each day sar will generate a report that lists the values of various system counters for a specified time interval throughout the day. So you end up with a text file that lists, for example, the cpu iowait value every 10 minutes throughout the day. There are maybe a dozen different categories of counters enabled by default, and more that aren’t (like disk-related counters). Anyway, you wind up with a text file that looks something like this:

23:30:01          CPU     %user     %nice   %system   %iowait    %steal     %idle
23:40:02          all      0.32      0.00      0.32      6.57      0.49     92.29
23:40:02            0      0.32      0.00      0.32      6.57      0.49     92.29
23:50:01          all      0.74      0.00      0.82      7.14      0.55     90.76
23:50:01            0      0.74      0.00      0.82      7.14      0.55     90.76
Average:          all      0.82      0.00      0.72     13.54      0.78     84.14
Average:            0      0.82      0.00      0.72     13.54      0.78     84.14

This is just a small part of one section of the file (this box has only one cpu, which is why the ‘all’ and ‘0′ numbers are the same, btw). The whole file on one server, running with default configurations, is 4000 lines long.

There’s a ton of great information in here, but… it all looks like the above. There’s no graphical output to be had. This is bad, because it would be nice to use this (historical) monitoring output for things like capacity planning, problem tracking, etc. You would, of course, want to couple this type of monitoring with something else that’ll do real-time monitoring, alerts, dependencies, escalation, etc.

So I want to write an application that’ll generate graphs of all of this stuff. Furthermore, I thought it would be cool to do something like what planetplanet does, which is to say that I want sarviz to run as a cron job, parse all of this stuff, and generate static html files, with an index.html that’ll make it really easy to browse this information either by host, by date, by resource… whatever. Later on I can add features to actually do even more useful stuff like longer-term trending of resource usage (by aggregating across various ’sar’ output files), and more.

Sar is not alone

Sar comes with some friends, and it turns out they can be extremely useful. The best one for my purposes here is called ’sadf’, and it is used to basically format the sar output to make it more useful for programmatic processing. It can output the information in CSV format, or make it ready for insertion into a relational database, but what I’m currently using for sarviz (and it’s early, so this could change) is the XML output capability. With XML output, I won’t have to deal with parsing out column headers, scanning an entire file for information from a single sar run, dealing with the blank lines sar uses by default to make it easier to read on a console, etc. So with sadf I can get output that looks like this:

<timestamp date=”2008-06-15″ time=”07:10:01″ interval=”600″>
<processes per=”second” proc=”0.93″/>
<context-switch per=”second” cswch=”221.50″/>
<cpu-load>
<cpu number=”all” user=”1.77″ nice=”0.00″ system=”0.56″ iowait=”0.04″ steal=”0.08″ idle=”97.55″/>
<cpu number=”0″ user=”1.77″ nice=”0.00″ system=”0.56″ iowait=”0.04″ steal=”0.08″ idle=”97.55″/>
</cpu-load> ….

This is a bit nicer to deal with, and I was excited to use Python’s (now built-in) ElementTree module to do something from scratch after having dealt with it being somewhat abstracted in the Python tools for the GData API (which I used to write a command line client for Google Spreadsheets, for example).

Doing Simple Things with ElementTree

Well, as it turns out, I had kind of a hard time getting started doing what I thought were simple things with ElementTree, so I want to post a few examples of how I did them so that I and others have something to refer to online.

The first thing to know about ElementTree is that there are Element objects, and ElementTree objects. ElementTree objects are made up of a hierarchical collection of Element objects, and Element objects are the things you can actually get attributes from that you’re likely to want. For whatever reason, I was a little confused starting out, because I wanted to get an ElementTree object and then ask that object to “scan the tree and give me all of the “time” attributes of the “timestamp” elements in the tree. You might be able to do this with a one-liner, but I never found a document that said how.

So here’s how to load in an XML file, parse it, and return all of the timestamp elements in that tree (or, rather, this is how I did it, which seems reasonable):

strudel:sa jonesy$ python
Python 2.5.1 (r251:54863, Jan 17 2008, 19:35:17)
[GCC 4.0.1 (Apple Inc. build 5465)] on darwin
Type “help”, “copyright”, “credits” or “license” for more information.
>>> from xml.etree import ElementTree as ET
>>> tree = ET.parse(”sa15.xml”)
>>> for ts in tree.findall(”host/statistics/timestamp”):
…        isotime = ts.attrib["date"]+”T”+ts.attrib["time"]
…        print isotime

2008-06-16T05:00:01
2008-06-16T05:10:01
2008-06-16T05:20:01
2008-06-16T05:30:01
2008-06-16T05:40:01
2008-06-16T05:50:01
2008-06-16T06:00:01
2008-06-16T06:10:01
2008-06-16T06:20:01
2008-06-16T06:30:01
2008-06-16T06:40:01
2008-06-16T06:50:01
….

So, I imported the ElementTree module, fed my xml file to a method called “parse()”, and that gives me an ElementTree object. In that tree, I then ask for the timestamp elements which are under the root element at “host/statistics/timestamp”. You can then see that I create an ISO8601-formatted timestamp by asking for the “date” and “time” attributes of the timestamp element, and put a “T” between them. I would’ve used something like “T”.join, but there are other attributes in that element, and I only needed two, so I took the easy way out here instead of creating a list first and then doing the join on the list.

Of course, my real interest in the timestamps isn’t to print them, but to get the statistics for each sar run (represented by a timestamp, since sar records statistics for regular time intervals). So now let’s grab the 1-, 5-, and 15-minute load averages according to sar. I want all of this printed on one line along with the timestamp, because this output is going to be graphed using Timeplot, and that’s how Timeplot wants the data. Here goes:


>>>for ts in tree.findall("host/statistics/timestamp"):
...        isotime = ts.attrib["date"] + “T” + ts.attrib["time"]
…        for q in ts.findall(”queue”):
…             qstat = [isotime, q.attrib["ldavg-1"], q.attrib["ldavg-5"], q.attrib["ldavg-15"]]
…             print “,”.join(qstat)

2008-06-16T05:10:01,0.05,0.12,0.09
2008-06-16T05:20:01,0.03,0.06,0.07
2008-06-16T05:30:01,0.02,0.02,0.03
2008-06-16T05:40:01,0.02,0.06,0.03
2008-06-16T05:50:01,0.03,0.06,0.03
2008-06-16T06:00:01,0.04,0.03,0.00
2008-06-16T06:10:01,0.02,0.06,0.03
2008-06-16T06:20:01,0.06,0.10,0.04
2008-06-16T06:30:01,0.13,0.11,0.06
2008-06-16T06:40:01,0.16,0.12,0.08
2008-06-16T06:50:01,0.04,0.06,0.06

The thing to note here, in case it escaped your eyeball, is that the second call to ‘findall’ feeds an argument relative to the ‘ts’ object rather than the ‘tree’ object.

This data is ready for Timeplot, and now it’s just a matter of somehow generating the files with the appropriate HTML and JavaScript in them to present the information. I have absolutely no clue how to easily use dynamic variables from Python to easily generate static HTML and JavaScript, so what I have in that area of my code is not something I want to share, out of sheer embarrasment. If someone has done that, let me know. PlanetPlanet does not output JavaScript, best I can tell, but it does output HTML, so I’ll be checking that part of the code out (probably uses BeautifulSoup I guess?). Input on that is hereby solicited!