Matt Connolly's Blog

my brain dumps here…

Category Archives: SmartOS

Using ZFS Snapshots on Time Machine backups.

I use time machine because it’s an awesome backup program. However, I don’t really trust hard drives that much, and I happen to be a bit of a file system geek, so I backup my laptop an iMac to another machine that stores the data on ZFS.

I first did this using Netatalk on OpenSolaris, then OpenIndiana, and now on SmartOS. Netatalk is an open source project for running AFP (Apple Filesharing Protocol) services on unix operatings systems. It has great support for new features in the protocol required for Time Machine. As far as I’m aware, all embedded NAS devices use this software.

Sometimes, Time Machine “eats itself”. A backup will fail with a message like “Verification failed”, and you’ll need to make a new one. I’ve never managed to recover the disk from this point using Disk Utility.

My setup is RaidZ of 3 x 2TB drives, giving a total of 4TB of storage space (and 2TB redundancy). In the four years I’ve been running this, I have had 3 drives go bad and replace them. They’re cheap drives, but I’ve never lost data due to a bad disk and having to replace it. I’ve also seen silent data corruptions, and know that ZFS has corrected them for me.

Starting a new backup is a pain, so what do I do?

ZFS Snapshots

I have a script, which looks like this:

ZFS=zones/MacBackup/MattBookPro
SERVER=vault.local
if [ -n "$1" ]; then
  SUFFIX=_"$1"
fi
SNAPSHOT=`date "+%Y%m%d_%H%M"`$SUFFIX
echo "Creating zfs snapshot: $SNAPSHOT"
ssh -x $SERVER zfs snapshot $ZFS@$SNAPSHOT

This uses the zfs snapshot command to create a snapshot of the backup. There’s another one for my iMac backup. I run this script manually for the ZFS file system (directory) for each backup. I’m working on an automatic solution that listens to system logs to know when the backup has completed and the volume is unmounted, but it’s not finished yet (like many things). Running the script takes about a second.

Purging snapshots

My current list of snapshots looks like this:

matt@vault:~$ zfs list -r -t all zones/MacBackup/MattBookPro
NAME                                      USED  AVAIL  REFER  MOUNTPOINT
zones/MacBackup/MattBookPro               574G   435G   349G  /MacBackup/MattBookPro
...snip...
zones/MacBackup/MattBookPro@20131124_1344 627M      -   351G  -
zones/MacBackup/MattBookPro@20131205_0813 251M      -   349G  -
zones/MacBackup/MattBookPro@20131212_0643 0         -   349G  -

The used at the top shows the space used by this file system and all of its snapshots. The used column for the snapshot shows how much space is used by that snapshot on its own.

Purging old snapshots is a manual process for now. One day I’ll get around to keeping a snapshots on a rule like time machine’s hourly, daily, weekly rules.

Rolling back

So when Time Machine goes bad, it’s as simple as rolling back to the latest snapshot, which was a known good state.

My steps are:

  1. shut down netatalk service
  2. zfs rollback
  3. delete netatalk inode database files
  4. restart netatalk service
  5. rescan directory to recreate inode numbers (using netatalks “dbd -r ” command.)

This process is a little more involved, but still much faster than making a whole new backup.

The main reason for this is that HFS uses an “inode” number to uniquely identify each file on a volume. This is one trick that Mac Aliases use to track a file even if it changes name and moves to another directory. This concept doesn’t exist in other file systems, and so Netatalk has to maintain a database of which numbers to use for which files. There’s some rules, like inode numbers can’t be reused and they must not change for a given file.

Unfortunately, ZFS rollback, like any other operation on the server that changes files without netatalk knowing, ends up with files that have no inode number. The bigger problem seems to be deleting files and leaving their inodes in that database. This tends to make Time Machine quite unhappy about using that network share. So after a rollback, I have a rule that I nuke netatalk’s database and recreate it.

This violates the rule that inode numbers shouldn’t change (unless they magically come out the same, which I highly doubt), but this hasn’t seemed to cause a problem for me. Imagine plugging a new computer into a time machine volume, it has no knowledge of what the inode numbers were, so it just starts using them as is. It’s more likely to be an issue for Netatalk scanning a directory and seeing inodes for files that are no longer there.

Recreating the netatalk inode database can take an hour or two, but it’s local to the server and much faster than a complete network backup which also looses your history.

Conclusion

This used to happen a lot. Say once every 3-4 months when I first started doing it. This may have been due to bugs in Time Machine, bugs in Netatalk or incompatibilities between them. It certainly wasn’t due to data corruptions.

Pros:

  • Time Machine, yay!
  • ZFS durability and integrity.
  • ZFS snapshots allow point in time recovery of my backup volume.
  • ZFS on disk compression to save backup space!
  • Netatalk uses standard AFP protocol, so time machine volume can be accessed from your restore partition or a new mac – no extra software required on the mac!

Cons:

  • Effort – complexity to manage, install & configure netatalk, etc.
  • Rollback time.
  • Network backups are slow.

As time has gone on, both Time Machine and Netatalk have improved substantially. And I’ve added an SSD cache to the server, and its is swimmingly fast and reliable. And thanks to ZFS, durable and free of corruptions. I think I’ve had this happen only twice in the last year, and both times was on Mountain Lion. I haven’t had to do a single rollback since starting to use Mavericks beta back around June.

Where to from here?

I’d still like to see a faster solution, and I have a plan: a network block device.

This would, however, require some software to be installed on the mac, so it may not be as easy to use in a disaster recover scenario.

ZFS has a feature called a “volume”. When you create one, it appears to the system (that’s running zfs) as another block device, just like a physical hard disk, or file. A file system can be created on this volume which can then be mounted locally. I use this for the disks in virtual machines, and can snapshot them and rollback just as if they were a file system tree of files.

There’s an existing server module that’s been around for a while: http://nbd.sourceforge.net

If this volume could be mounted across the network on a mac, the volume could be formatted as HFS+ and Time Machine could backup to it using local disk mode, skipping all the slow sparse image file system work. And there’s a lot of work. My time machine backup of a Mac with a 256GB disk creates a whopping 57206 files in the bands directory of the sparseimage. It’s a lot of work to traverse these files, even locally on the server.

This is my next best solution to actually using ZFS on mac. Whatever “reasons” Apple has for ditching them are not good enough simply because we don’t know what they are. ZFS is a complex beast. Apple is good at simplifying things. It could be the perfect solution.

Advertisements

Time Machine Backups and silent data corruptions

I’ve recently heard many folk talking about Time Machine backup strategies. To do it well, you really do need to backup your backup, as Time Machine can “eat itself”, especially doing network backups.

Regardless of whether your Time Machine backup is to a locally attached disk or a network drive, when you make a backup of your backup, you want to make sure it’s valid, otherwise you’re propagating a corrupt backup.

So how do you know if your backup is corrupt? You could read it from beginning to end. But this would only protect you from data corruptions that can be detected by the drive itself. Disk verify, fsck, and others go further and validate the file system structures, but still not your actual data.

There are “silent corruptions”, which is where the data you wrote to the disk comes back corrupted (different data, not a read error). “That never happens”, you might say, but how would you know?

I have two servers running SmartOS using data stored on ZFS. I ran a data scrub on them, and both reported checksum errors. This is exactly the silent data corruption scenario.

ZFS features full checksumming of all data when stored, and if your data is in a RAIDZ or mirror configuration, it will also self-heal. This means that instead of returning an error, ZFS will go fetch the data from a good drive and also make another clean copy of that block so that its durability matches your setup.

Here’s the specifics of my corruptions:

On a XEON system with ECC RAM, the affected drive is a Seagate 1TB Barracuda 7200rpm, ST31000524AS, approximately 1 year old.

  pool: zones
 state: ONLINE
status: One or more devices has experienced an unrecoverable error.  An
        attempt was made to correct the error.  Applications are unaffected.
action: Determine if the device needs to be replaced, and clear the errors
        using 'zpool clear' or replace the device with 'zpool replace'.
   see: http://illumos.org/msg/ZFS-8000-9P
   
  scan: resilvered 72.4M in 0h48m with 0 errors on Mon Nov 18 13:28:16 2013
config:

        NAME          STATE     READ WRITE CKSUM
        zones         ONLINE       0     0     0
          mirror-0    ONLINE       0     0     0
            c1t1d0s0  ONLINE       0     0     0
            c1t0d0s0  ONLINE   2.61K  366k   635
            c1t4d0s1  ONLINE       0     0     0
        logs
          c1t2d0s0    ONLINE       0     0     0
        cache
          c1t2d0s1    ONLINE       0     0     0

errors: No known data errors

On a Celeron system with non-ECC RAM, the affected drive is a Samsung 2TB low power drive, approximately 2 years old.

  pool: zones
 state: ONLINE
status: One or more devices has experienced an unrecoverable error.  An
        attempt was made to correct the error.  Applications are unaffected.
action: Determine if the device needs to be replaced, and clear the errors
        using 'zpool clear' or replace the device with 'zpool replace'.
   see: http://illumos.org/msg/ZFS-8000-9P
  scan: scrub repaired 8K in 12h51m with 0 errors on Thu Nov 21 00:44:25 2013
config:

        NAME          STATE     READ WRITE CKSUM
        zones         ONLINE       0     0     0
          raidz1-0    ONLINE       0     0     0
            c0t1d0    ONLINE       0     0     0
            c0t3d0    ONLINE       0     0     0
            c0t2d0p2  ONLINE       0     0     2
        logs
          c0t0d0s0    ONLINE       0     0     0
        cache
          c0t0d0s1    ONLINE       0     0     0

errors: No known data errors

Any errors are scary, but the checksum errors even more so.

I had previously seen thousands of checksum errors on a Western Digital Green drive. I stopped using it and threw it in the bin.

I have other drives that are HFS formatted. I have no way of knowing if they have any corrupted blocks.

So unless your data is being checksummed, you are not protected from data corruption, and making a backup of a backup could easily be propagating data corruptions.

I dream of a day when we can have ZFS natively on Mac. And if it can’t be done for whatever ‘reasons’, at least give us the features from ZFS that we can use to protect our data.

Network latency in SmartOS virtual machines

Today I decided to explore network latency in SmartOS virtual machines. Using the rbczmq ruby gem for ZeroMQ, I made two very simple scripts: a server that replies “hello” and a benchmark script that times how long it takes to send and receive 5000 messages after establishing the connection.

The server code is:

require 'rbczmq'
ctx = ZMQ::Context.new
sock = ctx.socket ZMQ::REP
sock.bind("tcp://0.0.0.0:5555")
loop do
  sock.recv
  sock.send "reply"
end

The benchmark code is:

require 'rbczmq'
require 'benchmark'

ctx = ZMQ::Context.new
sock = ctx.socket ZMQ::REQ
sock.connect(ARGV[0])

# establish the connection
sock.send "hello"
sock.recv

# run 5000 cycles of send request, receive reply.
puts Benchmark.measure {
  5000.times {
    sock.send "hello"
    sock.recv
  }
}

The test machines are:

* Mac laptop – server & benchmarking
* SmartOS1 (SmartOS virtual machine/zone) server & benchmarking
* SmartOS2 (SmartOS virtual machine/zone) benchmarking
* Linux1 (Ubuntu Linux in KVM virtual machine) server & benchmarking
* Linux2 (Ubuntu Linux in KVM virtual machine) benchmarking

The results are:

Source      Dest        Connection      Time          Req-Rep/Sec
------      ----        ----------      ----          --------
Mac         Linux1      1Gig Ethernet   5.038577      992.3
Mac         SmartOS1    1Gig Ethernet   4.972102      1005.6
Linux2      Linux1      Virtual         1.696516      2947.2
SmartOS2    Linux1      Virtual         1.153557      4334.4
Linux2      SmartOS1    Virtual         0.952066      5251.8
Linux1      Linux1      localhost       0.836955      5974.0
Mac         Mac         localhost       0.781815      6395.4
SmartOS2    SmartOS1    Virtual         0.470290      10631.7
SmartOS1    SmartOS1    localhost       0.374373      13355.7

localhost tests use 127.0.0.1

SmartOS has an impressive network stack. Request-reply times from one SmartOS machine to another are over 3 times faster than when using Linux under KVM (on the same host). This mightn’t make much of a difference to web requests coming from slow mobile device connections, but if your web server is making many requests to internal services (database, cache, etc) this could make a noticeable difference.

ZeroMQ logging for ruby apps

I’ve been thinking for a while about using ZeroMQ for logging. This is especially useful with trends towards micro-services and scaling apps to multiple cloud server instances.

So I put thoughts into action and added a logger class to the rbczmq gem that logs to a ZeroMQ socket from an object that looks just like a normal ruby logger: https://github.com/mattconnolly/rbczmq/blob/master/lib/zmq/logger.rb

There’s not much to it, because, well, there’s not much to it. Here’s a simple app that writes log messages:

Log Writer:

require 'rbczmq'
require_relative './logger'
require 'benchmark'
ctx = ZMQ::Context.new
socket = ctx.socket(ZMQ::PUSH)
socket.connect('tcp://localhost:7777')
logger = ZMQ::Logger.new(socket)
puts Benchmark.measure {
 10000.times do |x|
 logger.debug "Hello world, #{x}"
 end
}

With benchmark results such as:

  0.400000   0.220000   0.620000 (  0.418493)

Log Reader:

And reading is even easier:

require 'rbczmq'
ctx = ZMQ::Context.new
socket = ctx.socket(ZMQ::PULL)
socket.bind('tcp://*:7777')
loop do
 msg = socket.recv
 puts msg
end

Voila. Multiple apps can connect to the same log reader. Log messages will be “fair queued” between the sources. In a test run on my 2010 MacBook Pro, I can send about 13000 log messages a second. I needed to run three of the log writers above in parallel before I maxed out the 4 cores and it slowed down. Each process used about 12 MB RAM. Lightweight and fast.

Log Broadcasting:

If we then need to broadcast these log messages for multiple readers, we could easily do this:

require 'rbczmq'
ctx = ZMQ::Context.new
socket = ctx.socket(ZMQ::PULL)
socket.bind('tcp://*:7777')
publish = ctx.socket(ZMQ::PUB)
publish.bind('tcp://*:7778')
loop do
 msg = socket.recv
 publish.send(msg)
end

Then we have many log sources connected to many log readers. And the log readers can also subscribe to a filtered stream of messages, so one reader could do something special with error messages, for example.

Building Ruby 2 in SmartOS

Installing Ruby 2 in SmartOS!

Ruby 2 has been out for a while, so let’s get it going in SmartOS!!

I’m working with the SmartMachine (base64 1.9.0).

Part 1: Requirements.

The SmartMachine has a whole bunch of useful packages already installed. Here are the additional packages needed to build Ruby 2.0.0 in SmartOS:

# pkgin install build-essential gcc47 gcc47-libs libyaml 

Part 2: Configure hacks

There are a few issues with Ruby 2.0.0, but fortunately all of them have a command line workaround.

1. Ruby bug #5384 https://bugs.ruby-lang.org/issues/5384

Actually an upstream bug in (Open)Solaris/Illumos/OpenIndiana/SmartOS. Workaround is to add `ac_cv_func_dl_iterate_phdr=no` to the configure line.

See also: https://www.illumos.org/issues/1587

2. Ruby bug #8071 https://bugs.ruby-lang.org/issues/8071

Non-portable code in ruby’s configure scipt. Easy workaround, prepend the configure command with a shell that can handle the current state of the configure script, ie: `bash`. (A fix has already been submitted, ans should be in the next ruby patch.)

3. Ruby bug #8268 https://bugs.ruby-lang.org/issues/8268

“-ggdb3″ C flags issue (logged with Ruby, but not necessarily actually a bug in ruby. Please help if you can on this one!)

Workaround 1: Add `debugflags=”-ggdb”` to the configure line. Caveat: It appears this will add gdb debug info to built ruby binaries, which may not be desired.

Workaround 2: Add `CFLAGS=”-R -fPIC”`. This introduces a make error. Missing function ‘signbit’. Boo.

Workaround 3: Add `CFLAGS=”-R -fPIC” rb_cv_have_signbit=no`. Works!

So with these three taken care of, the following command line will configure ruby-2.0.0-p0 to compile on SmartOS (in 64 bit):

$ bash ./configure --prefix=$PREFIX --with-opt-dir=/opt/local --enable-shared ac_cv_func_dl_iterate_phdr=no CFLAGS="-R -fPIC" rb_cv_have_signbit=no
$ make && make install

And there we have it. Ruby 2.0.0-p0 building in SmartOS. Next challenge, making use of those built in DTrace probes…

Building netatalk in SmartOS

I’m looking at switching my home backup server from OpenIndiana to SmartOS. (there’s a few reasons, and that’s another post).

One of the main functions of my box is to be a Time Machine backup for my macs (my laptop and my wife’s iMac). I found this excellent post about building netatalk 3.0.1 in SmartOS, but it skipped a few of the dependencies, and did the patch after configure, which means if you change you reconfigure netatalk, then you need to reapply the patch.

Based on that article, I came up with a patch for netatalk, and here’s a gist of it: https://gist.github.com/mattconnolly/5230461

Prerequisites:

SmartOS already has most of the useful bits installed, but these are the ones I needed to install to allow netatalk to build:

$ sudo pkgin install gcc47 gmake libgcrypt

Build netatalk:

Download the latest stable netatalk. The netatalk home page has a handy link on the left.

$ cd netatalk-3.0.2
$ curl 'https://gist.github.com/mattconnolly/5230461/raw/27c02a276e7c2ec851766025a706b24e8e3db377/netatalk-3.0.2-smartos.patch' > netatalk-smartos.patch
$ patch -p1 < netatalk-smartos.patch
$ ./configure --with-bdb=/opt/local --with-init-style=solaris --with-init-dir=/var/svc/manifest/network/ --prefix=/opt/local
$ make
$ sudo make install

With the prefix of ‘/opt/local’ netatalk’s configuration file will be at ‘/opt/local/etc/afp.conf’

Enjoy.

[UPDATE]

There is a very recent commit in the netatalk source for an `init-dir` option to configure which means that in the future this patch won’t be necessary, and adding `--with-init-dir=/var/svc/manifest/network/` will do the job. Thanks HAT!

[UPDATE 2]

Netatalk 3.0.3 was just released, which includes the –init-dir option, so the patch is no longer necessary. Code above is updated.