What's it all about, Alfie?

Today was a stressful day - very. I'm trying to get the CouchDB and Salesforce endpoints to scale, having essentially no luck, and then going into a meeting this afternoon about how we can't hold back anything for the new Pilot Launch on Monday.

Like the Orange Juice commercial, the fun just kept coming as we were in a weekly planning meeting and I found out that we need to have even more features in the code before tomorrow evening.

13 hours.

We have 13 hours to stabilize this code, get some sense of scalability, fix a few outstanding features, and then del with these few more, and all in 13 hours.

I think it's a mistake, and it's driven by nothing more than the desire by the project manager to look good to his boss(es). It's all about face. His. He isn't asking us for dates, he's telling us dates, and then when things come up - as they often do, it's "impossible" to move the dates. All for face.

So I'm getting a little tired of this today. I need to just go home.

Thankfully, it's Angelina's Birthday, so I have that to look forward to tonight. Good.

This morning I noticed that Google Chrome dev 25.0.1323.1 was out and the release notes have returned to a more spartan style. Yes, people can read SVN logs, but that's not the point - really, is it? If you take the time to make a blog post about the release, you should be able to make release notes that say more than "Read the SVN logs".

Still… progress is nice to see continuing.

The current plan was to have Production running on the old EC2 boxes and have UAT run out of our data center until we were sure things were OK, and then switch Production over as well. This seemed like a good plan, but there were issues with this and management wanted to run the essential production data, and UAT data, in EC2 and then run the non-essential production data and UAT data in SNC1. This means that there would be multiple boxes running the same code hitting the same data sources and sinks, but covering different regions.

Sounds reasonable, and even safe. So let's do that.

The issues with getting a different config for UAT in one datacenter is that all we really have is hostname versus hostname -f, and I had to use that every place I could. The wrinkle came in when looking at the configuration, as that isn't looking at the machine name - just the environment setting - 'hat' or 'production'. Not so easy.

But I worked on this all day. It was not easy. And then I was ready to test.

It wasn't pretty.

The problem comes in that we aren't synchronizing the work between data centers, and this makes the later processing steps (prioritization) fail because you only have part of the picture. There's no easy way around it.

The next problem was continued Couch errors. Yes, using my cacheing endpoint helped, but we'd still get problems now and then. No easy solution in our code.

So in the end, almost a wasted day. Almost.

Not a great feeling.

Today I've been having real troubles trying to get our system scaled up to new hardware in the datacenter. Moving from hosts in Amazon EC2 to machines in our own datacenter are a big step in the right direction, but going from iffy bandwidth in EC2 to solid switches in the datacenter and 2 cores to 24 are something we simply have to do in order to scale up to handling the global data load that we're going to need.

The problems I've run into today are all about loading. I've already added a queueing system to the CouchDB interface, in order to minimize the connections to the Couch server so as not to overload it - there were times when the Couch server would simply shutdown it's socket listener, and therefore refuse all updates sent from the process. Not good.

There's also a lot of problems today with Salesforce. I don't think they expected the kind of loads we're delivering. This morning, at 3:00 am, Salesforce called the support guys at the shop and told them that a process was bringing one of their sandbox clusters to it's knees, and that, it turns out, is but one of four boxes we need to bring online. They're having a hard time handling this much - I can't imagine what's going to happen when we try to really ramp it up.

I'm starting to have real concerns about both these endpoints of the project. I know there's a lot of data getting moved around, and while we're able to handle it, it's these endpoints that are having the hardest time. I've talked with the project manager and the technical manager about this, and I think we need to start thinking about potential bail-out scenarios.

It's certainly possible to read from Salesforce. We're planning on re-doing the complete demand system, so there shouldn't be an issue there. Persistence? Go back to MySQL or PostgreSQL and store it all in tables. The data is getting pretty nicely finalized, so a nice schema should be able to be made. Save it all in a SQL database, make a simple service that reads/caches this data and offers it up to the clients, and the pages already built on the existing data sources are pretty easily modified.

Odd to think that I was looking at MySQL before Couch popped up. Funny thing is, I have a strong feeling that Salesforce can come up with hardware that makes the grade, but it's the bugs in Couch that worry me the most. You just gotta wonder if it's in the Erlang code, or on the boxes, or what.

So many unknowns, but it's clear that we can't scale to one nice box - there's no way we're going to make it work globally without some serious changes.

I've been fighting a problem with our CouchDB installation using the CouchRest interface for ruby/jruby. What I'm seeing is that when I have a lot of updates to CouchDB from the REST API in CouchRest, we start to get socket connection errors to the CouchDB server. I've gone through a lot of different configurations of open file handles on both boxes, and nothing seems to really 'fix' the problem.

So what I wanted to do was to make a queueing database class. Currently, we have something that abstracts away the CouchDB-centric features, and adds in necessary metadata so that we can more easily track all the data held in CouchDB. This is a nice start, in that I only really needed to add to the bulk of the code was a simple flush method:

  Database.flush

where I initially started with it being implemented as:

  def flush
    nil
  end

At this point, I was free to do just about anything with the Database class - as long as I didn't change the public API. What I came up with is this:

What I really like about this is that I can use it either way - as a cacheing/flushing endpoint, or as a straight pass-through with thread-local database connections. This means that I can start with a simple config file setting:

  data_to: 'database'
  writers: 3

which will give me three writer threads, as specified in the start method argument. Then, when I fix the CouchDB issues, I can switch to the simpler, thread-local connection storage with the code:

  Database.start(config.writers) if config.writers > 0

The call to flush is a no-op if we didn't start anything, so there's no harm in always calling it at the end of the process. Pretty nice. I've easily verified that this gets me what I need, and it's just a matter of how 'throttling' I want to be with the writing of the data to CouchDB. But I tell you this… I'm closer to MySQL than ever before - just because of this. There's no reason in the world to put up with this kind of software if it can't do the job.

When faced with the idea that CouchDB might be running low on sockets (file handles), I decided to do a little digging into the CentOS 5 configuration, and see what I could do. Turns out it wasn't all that hard, and I got immediate results. The first thing is to up the hard and soft limits on the file handles in /etc/security/limits.conf for the couchdb user:

  couchdb  soft  nofile  32768
  couchdb  hard  nofile  60000

then simply restart CouchDB. Look at the process limits by:

  $ cat /proc/1234/limits

where 1234 is the pid of the erlang 'beam' process, and see what it says. You are also going to need to tell Erlang about the number of file handles in the /usr/local/etc/default/couchdb file:

  export ERL_MAX_PORTS=32768
  export ERL_FLAGS="+A 8"

The second little bit there is to allow 8 threads in Erlang to process disk updates. This multi-threaded filesystem access isn't on be default for Erlang, so if you're on a big box with lots of cores, it makes a lot of sense to turn it on.

With these, you have CouchDB running about as well as you can get. The rest is up to you.

While trying to figure out a problem with the scaling, I was getting a bunch of BindExceptions and they seemed to point to this point in the jruby-1.7.0 code:

  try {
    // This is a bit convoluted because (1) SocketChannel.bind
    // is only in jdk 7 and (2) Socket.getChannel() seems to
    // return null in some cases
    channel = SocketChannel.open();
    Socket socket = channel.socket();
 
    if (localHost != null) {
      socket.bind( new InetSocketAddress(
                            InetAddress.getByName(localHost),
                            localPort) );
    }
 
    try {
      // Do this nonblocking so we can be interrupted
      channel.configureBlocking(false);
      channel.connect( new InetSocketAddress(
                            InetAddress.getByName(remoteHost),
                            remotePort) );
      context.getThread().select(channel, this, SelectionKey.OP_CONNECT);
      channel.finishConnect();
 
      // only try to set blocking back if we succeeded to finish connecting
      channel.configureBlocking(true);

If I was getting bind exceptions, then it had to be coming from line 104 - after all, that's the call to bind(). The problem was, this really needed to be preceded with a call to set SO_REUSEADDR to true so that we didn't have these kinds of issues.

Something like this:

  try {
    // This is a bit convoluted because (1) SocketChannel.bind
    // is only in jdk 7 and (2) Socket.getChannel() seems to
    // return null in some cases
    channel = SocketChannel.open();
    Socket socket = channel.socket();
 
    if (localHost != null) {
      socket.setReuseAddress(true);
      socket.bind( new InetSocketAddress(
                            InetAddress.getByName(localHost),
                            localPort) );
    }
 
    try {
      // Do this nonblocking so we can be interrupted
      channel.configureBlocking(false);
      channel.connect( new InetSocketAddress(
                            InetAddress.getByName(remoteHost),
                            remotePort) );
      context.getThread().select(channel, this, SelectionKey.OP_CONNECT);
      channel.finishConnect();
 
      // only try to set blocking back if we succeeded to finish connecting
      channel.configureBlocking(true);

Simple change, but it'd make a huge difference to the operation of the socket. So I submitted a request to the JRuby team.

After a little bit, I realized that I needed to try and build this all myself, and see if I couldn't just prove the problem for myself. The difficulty was - How to build JRuby?

Turns out, it's not all that hard. First, fork the code on GitHub so you can issue Pull Requests with the changed code. Then, follow the directions to build everything about JRuby:

  $ git clone git://github.com/drbobbeaty/jruby.git
  $ cd jruby
  $ ant
  $ ant jar-complete
  $ ant dist

It should all build just fine. Then check everything in, and push it up to GitHub. Next, you need to tell rvm to build a new version of ruby from that GitHub repo:

  $ rvm get head
  $ rvm reinstall jruby-head --url git://github.com/

And we're almost done… Update the Gemfile to have the paths to the jruby-jars file where you built it, and then update the .rvmrc. I was able to get things running and even packaging a jar file for our deployment.

Hacking on the language. That's a first for me. Quite a fun experience.

What I noticed was that it wasn't the bind on line 104 - it was the connect on line 113. Java returned the BindException as a general Can't get the endpoint, and I found this out by putting in the catch statements and logging the host and port of the exception. This was very illuminating, and I'm very glad I did it.

I'm going to even send a pull request to see if they'll take it. It's a very useful debugging tool.

As part of my working in scaling up the code to a global system, I ran into some very odd problems in jruby-1.7.0. I'm not exactly sure if they are in jruby or in the JDK 1.7.0_05 implementation for CentOS 5, but I'm willing to bet they are in jruby, as I think the JDK has been hammered on a lot on these issues. So here's what I found.

It's all about using Java Executors in jruby-1.7.0. I'm creating a series of threads with the newFixedThreadPool() method, and then running through an array of things to send out to a REST API (Salesforce), and then waiting for it all to finish up.

  require 'java'
  java_include 'java.util.concurrent.Executors'
 
  executors = Executors.new_fixed_thread_pool(5)
  updates.each do |u|
    send_update(u)
  end

What I'm seeing is that when the threads are done processing, they don't all seem to "clear" the executor. For some reason, they aren't seen as "done" by the executor, but when looking at the REST API service, I know they completed. And it's always in the last batch of tasks.

This doesn't ever seem to happen on the Amazon EC2 hardware - only the nice, new, fast boxes in the datacenter.

So what I decided to do was to add a special timeout to the shutdown of the executor (start at line 36). This says that if we know how long any action should take, then if we get to the end of the processing queue in the executor, and we have waited long enough, then it's OK to forcibly shut down the executors and know that ready-of-not, it should have been done.

It's not ideal, and in most good cases, it shouldn't happen. But I'm getting a lot of problems with Salesforce and CouchDB as a part of this scaling, and I really have no idea what's going on inside either of those systems. Better to add this and be safe.

This afternoon I finally got around to fixing up one of the problems I noticed this morning in my battles with production, but didn't want to attack at the time - and that was converting the class I was having problems with to use class methods (static methods), more in keeping with the other, similar, classes in the codebase. It's a reassigned, and in that mode it's really quite stateless, so there's no reason to "create one", and then let it work on the data. We can simply just have something that works on the data. But the original author didn't see it that way, and passed in things that the other, similar, functional blocks had internalized.

This made it a little odd in the sense that one part of the system did it as a "factory", and the other as an "instance", but that's easy enough to change - now that I have the time. I simply went in and cleaned everything up, correcting all the references and making sure it all worked as it should.

It wasn't too hard, and in the end it was a lot more like the other code. In fact, I wouldn't be surprised if we could abstract a lot of both of these into a superclass that has all the common stuff, and the only point of these individual worker classes is the specifics they deal with.

But that's for another day...

When I got in this morning I saw that once again, I was going to be fighting production again, but this time, it's really from the additional processes that we're running - not the main pipeline process that we've been working on for several months.

No… this was somewhat new stuff, and in that I'm not terribly upset, but it is annoying when you see code like this:

  hot_lead.downcase[0..1]

because I know there's going to come a time that hot_lead is nil, and that'll blow up. It's so easy to say:

  hot_lead.downcase[0..1] unless hot_lead.nil?

or something like that. So easy… but the Code Monkeys don't think it's necessary.

Sigh…

In addition to some code problems there were legitimate timeouts with Salesforce, and those had to really be felt with. Again, far too optimistic in this new code, assuming everything would work, and in the real world, we know that's just not the case. So I quickly pulled in the timeout retry handling code from another class and converted it to work with the code having the timeouts. This meant converting the pulled-in methods from class methods to instance methods, but I'll get back to this later in the day and clean this up.

With the timeouts under control, I needed to fix a few issues with the summary script that parses the log files and summarizes them nicely. I had a few issues there, but they weren't as big as the problems I'd felt with up to that point in time. Annoying, yes, but not horribly so.

In the end, I got production run, but it was a touch frantic at times to make sure it got done as quickly as possible.