It seems to me that, in fact, your original idea was, in fact, the correct one - rsync probably would have been the best way to do this (and separately, a truck full of disks probably would have been the other best way).
First, rsync took too long probably because you used just one thread and didn't optimize your command-line options - most of the performance problems with rsync with large filesystem trees comes from using one command to run everything, something like:
rsync -av /source/giant/tree /dest/giant/tree
And the process of crawling, checksumming, storing is not only generally slow, but incredibly inefficient on today's modern multicore processors.
Much better to break it up into many threads, something like:
That alone probably would have dramatically sped things up, BUT you do still have your speed of light issues.
This is where Amazon import/export comes in - do a one-time tar/rsync of your data to an external 9TB array, ship it to Amazon, have them import it to S3, load it onto your local Amazon machines.
You now have two copies of your data - one on s3, and one on your amazon machine.
Then you use your optimized rsync to run and bring it up to a relatively consistent state - i.e. it runs for 8 hours to sync up, now you're 8 hours behind.
Then you take a brief downtime and run the optimized rsync one more time, and now you have two fully consistent filesystems.
No need for drbd and all the rest of this - just rsync and an external array.
I've used this method to duplicate terabytes and terabytes of data around, and 10s of millions of small files. It works, and is a lot fewer moving parts than drbd
Whenever someone gripes that rsync/scp is slow, it's usually because they didn't bother to look into proper solutions for their problem. rsync will barely fill the pipe in many/most cases. Using GridFTP or bbcp is usually preferred.
Which is how I've done it in the past - I'm sure these days there's utilities that will do it for you, but I had a bunch of perl code that would fork off N threads out of a queue and as one exited successfully, kick off another worker.
The issue with xargs back in the day was that you might need to run several hundred rsync processes, and suddenly launching 500+ processes in parallel made your server very very sad. So you needed some basic job queueing system.
Use at most max-args arguments per command line. Fewer than max-args arguments will be used if the size (see the -s option) is exceeded, unless the -x option is given, in which case xargs will exit.
...
--max-procs=max-procs
-P max-procs
Run up to max-procs processes at a time; the default is 1. If max-procs is 0, xargs will run as many processes as possible at a time. Use the -n option with -P; otherwise chances are that only one exec will be done.
This was also...14?-ish years ago. Sometimes on Solaris 2.6 or 8 boxes. I am pretty sure xargs didn't have that flag back then (which is why I said, "back in the day").
Yes, it likely would - most disks will top out on read i/o rate, especially on somewhat random access.
rsync in general is quite optimized and usually the limiting factor to a data transfer is the network or disk rather than CPU speed (unless crypto is involved, for example, over a ssh connection)
1. Ask someone else who has already done what you're thinking of doing. They have already made all the mistakes you might make and have figured out a way that works.
2. Assume that whatever you think will work will fail in an unexpected way with probably-catastrophic results. Test everything before you try something new.
Stories like these remind me of Andy Tannenbaum's statement: Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway.
I've been involved with multiple large data center migrations over the course of the last 25 years. Every single time, there is a discussion over the best way to transfer the data. Every single time, we choose the same option: Copy all the data to external hard drives. A tech puts them in a carry on bag, heads to the airport, and flies to the new data center.
Honestly, this seems to me like a case study in getting carried away by cloud. There's a very strong case for cloud hosting when you have a clear need for elasticity and an distributed application to leverage it.
Here, you have your entire business in one non-distributed Amazon instance. Amazon does not provide excellent service, availability, flexibility, or value for this model. It is in every way inferior to what you would get from colo or managed dedicated server hosting. Hosting your whole business on a single anonymous Amazon cloud instance that you can't walk up to and touch is engineering malpractice.
We're pretty aware of the advantages and drawback of the cloud. We moved from AWS to a private server and then back to AWS. You're right that cloud hosting makes more sense if you need scaling. The move described was to a single server. But now that we have the data on AWS we can expand. Probably the first step will be a separate fileserver and appserver. After that we can add multiple appservers and consider availability zone failover.
LVM can only shrink devices by full extents, but LVM is just a wrapper around device-mapper (plus metadata management). By creating device-mapper devices directly with the linear target, you could have gotten sector (512 byte) granularity for offset and size.
I don't quite understand. Why did the entire truck dance have to be redone when there was a perfectly nice stale copy of the data sitting there? Couldn't the second truck dance have used rsync and been done in under a day?
I wonder how GlusterFS would perform in this case. Replicated volumes with one or more local nodes for the most likely failover requirement, with async-georeplication in case the building burned down.
First, rsync took too long probably because you used just one thread and didn't optimize your command-line options - most of the performance problems with rsync with large filesystem trees comes from using one command to run everything, something like:
rsync -av /source/giant/tree /dest/giant/tree
And the process of crawling, checksumming, storing is not only generally slow, but incredibly inefficient on today's modern multicore processors.
Much better to break it up into many threads, something like:
rsync -av /source/giant/tree/subdir1 /dest/giant/tree/subdir1
rsync -av /source/giant/tree/subdir2 /dest/giant/tree/subdir2
rsync -av /source/giant/tree/subdir3 /dest/giant/tree/subdir3
That alone probably would have dramatically sped things up, BUT you do still have your speed of light issues.
This is where Amazon import/export comes in - do a one-time tar/rsync of your data to an external 9TB array, ship it to Amazon, have them import it to S3, load it onto your local Amazon machines.
You now have two copies of your data - one on s3, and one on your amazon machine.
Then you use your optimized rsync to run and bring it up to a relatively consistent state - i.e. it runs for 8 hours to sync up, now you're 8 hours behind.
Then you take a brief downtime and run the optimized rsync one more time, and now you have two fully consistent filesystems.
No need for drbd and all the rest of this - just rsync and an external array.
I've used this method to duplicate terabytes and terabytes of data around, and 10s of millions of small files. It works, and is a lot fewer moving parts than drbd