I had the same question as you, so I worked out how to do it.
First, I did this from the Ubuntu 32-bit EBS-backed ami from the US-East region, other OS's or images may work differently. However, I suspect that you should be ok, as long as you are using an ext* filesystem. It might work on other filesystems, but you'll have to figure out how to resize them on your own.
The steps are basically:
Attach two volumes to a running instance, the first based on the snapshot you want to shrink, and the second a blank volume having the new size you want to shrink to.
Check the file system of the first volume and repair any errors.
Shrink the file system on the first volume so it is only as big as it needs to be to hold the data.
Copy the file system from the first volume to the second.
Expand the file system on the second volume to it's maximum size.
Make sure everything looks good by checking the second volume for errors.
Take a snapshot of the second volume.
Create a machine image based on the snapshot of the second volume you just took.
You first need to get some information from the ami you want to shrink. In particular, you need the kernel ID and ramdisk ID, if any (the image I shrunk didn't have a ramdisk). All this information should be available from the aws management console ,in the AMI window.
The kernel ID looks like kia-xxxxxxxx, and the snapshot ID looks like snap-xxxxxxxx, and ramdisk IDs look like RIA-xxxxxxxx.
Next, launch a linux instance. I launched a Ubuntu instance. You can use a t1.micro instance if you like. It doesn't take much power to do these next steps.
After the machine is running, attach the snapshot you wrote down from the first step. In my case, I attached it to /dev/sdf
Then, create a new volume, having the size you want. In my case, I created a 5GB volume, as that's the size I wanted to shrink it to. Don't create this new volume from a snapshot. We need a new blank volume. Next, attach it to the running instance, in my case I attached it as /dev/sdg
Next, ssh into the machine but don't mount the attached volumes.
At this point, I erred on the side of paranoia, and I opted to check the file system on the large volume, just to make sure there were no errors. If you are confident that there are none, you can skip this step:
$ sudo e2fsck -f /dev/sdf
Next, I resized the file system on the large volume so that it was only as big as the data on the disk:
$ sudo resize2fs -M -p /dev/sdf
The -M shrinks it, and the -p prints the progress.
The resize2fs should tell you how large the shrunkin filesystem is. In my case, it gave me the size in 4K blocks.
We now copy the shrunkin file system to the new disk. We're going to copy the data in 16MB chunks, so we need to figure out how many 16MB chunks we need to copy. This is where that shrunken file system size comes in handey.
In my case, the shrunk file system was just over 1 GB, because I had installed a lot of other programs on the basic Ubuntu system before taking a snapshot. I probably could have gotten away with just copying the size of the file system rounded up to the nearest 16MB, but I wanted to play it safe.
So, 128 times 16MB chunks = 2GB:
$ sudo dd if=/dev/sdf ibs=16M of=/dev/sdg obs=16M count=128
I copied in blocks of 16MB because with EBS, you pay for each read and write, so I wanted to minimize the number of them as much as possible. I don't know if doing it this way did so, but it probably didn't hurt.
We then need to resize the file system we just copied to the new volume so that it uses all the available space on the volume.
$ sudo resize2fs -p /dev/sdg
Finally, check it, to make sure everything is well:
$ sudo e2fsck -f /dev/sdg
That's all we need to do in this machine, though it couldn't hurt to mount the new volume, just as a test. However, this step is almost certainly optional, as e2fsck should have caught any problems.
We now need to snapshot the new volume, and create an AMI based on it. We're done with the machine, so you can terminate it if you like.
Make sure the small volume is unmounted if you mounted it, and then take a snapshot of it. Again, you can do this in the management console.
The final step requires the commandline ec2 tools.
EDIT:
Since this answer was posted the AWS console allows you to simply
right click a snapshot and select Create Image from Snapshot. You will
still need to select the appropriate Kernel ID. If it does not appear
on the list make sure you've selected the appropriate architecture.
We use the ec2-register application to register an AMI based on the snapshot you just took, so write down the snap-xxxxxxxx value from the snapshot you just took.
You should then use a command like:
ec2-register -C cert.pem -K sk.pem -n The_Name_of_Your_New_Image
-d Your_Description_of_This_New_AMI --kernel aki-xxxxxxxx
-b "/dev/sda1=snap-xxxxxxxx" --root-device-name /dev/sda1
You of course need to replace the kernel ID with the one you wrote down at the beginning and the snapshot ID with the one you created in the previous step. You also need to point it at your secret key (called sk.pem) above, and your x509 cert (called cert.pem). You can of course choose whatever you want for the name and description.
Hope this helps.
Best Answer
You're talking about Elastic Block Store (EBS) vs Instance Store (IS).
EBS / Network Attached Storage
EBS is network attached storage, but that storage can be general purpose SSD, provisioned IOPS ssd, throughput optimized hard disks, or cold hard disks. Docs here.
Instance Storage The instance store has disks attached to servers. Some are SATA SSD (I2), some are NVM SSD (I3), some are hard disks (Dx).
Instance Types
The current generation general purpose instances instances (M4 general purpose/R4 high RAM/C4 high CPU) don't have instance stores. Some of the older ones do, like C3.
T2 instances are a great general purpose instance type for varying workloads, such as public web servers or servers that are high load during the day and low load overnight. They use a credit system for access to the CPU. T2 doesn't have an instance store, it's EBS only.
There are however the I (I/O Optimized) and D (dense storage optimized) instances. They have storage directly attached to the instances. The i3 instances have nvm ssd that can do 100,000 IOPS or more. This is useful for huge databases you want to run yourself, other specialised workloads.
Answers to your Questions
To answer your questions:
If you use an instance store (I / D / F series instances) then yes, you lose your root drive on stop or terminate, but not on restart. If you choose EBS SSD then no, it's persistent disk. In either case you can take an image of the disk for backup, called a snapshot, which lets you restore quickly.
This is fine for caches, and distributed databases where the data exists elsewhere, or if you keep VERY good backups. Sometimes your I/O needs can't be satisfied any other way. For example a MongoDB cluster of three instances can serve an extremely high workload, and the reliability is ok since if one instance stops the other two pick it up, and may have capacity to replicate to another mirror when it comes online.
In general you should start with T instances for varying workloads like web servers, or M for constant workloads.
Pricing Table
The pricing table says "EBS Only", "SSD", or "HDD". "SSD" and "HDD" are instance store.