Also, many applications will already try avoiding duplicate data to varying degrees which makes dedup even less beneficial (backup utilities, VCSes, etc).
If you are dealing with a lot of text, though (like in a mail archive, or with lots of source code) compression really shines. It has way less memory overhead, pretty low CPU overhead, and it can significantly increase disk throughput.
You'll notice I was careful to say "under many common workloads" in my other post, it was because of all this. For most people, most of the time, there would just be no real benefit and a decent performance hit with dedup.
Edit: to clarify
New mails are typically appended to the end of the archive, but the rest of it remains more or less the same across versions.
Also, a well-written block-level deduplication algorithm will work regardless of boundary shifts. I designed and wrote the proprietary blocklevel algorithm used in Genie Timeline, and can tell you that it saves terabytes of data in corporations for just PSTs alone. And it doesn't rely on block boundaries, capable of detecting slided or shifted data.
In other words, ZFS operates on blocks at a block level - it can't make any assumptions about the contents other than "this block is the same as this block" or "this block is not the same as this block." It can't (based on my understanding) realize that "this block is almost the same as this block" and shift things accordingly, because then it would be messing with application data that might be corrupted by the changes.
If the application is making those shifts on the other hand, it would work just fine because the application could be smart enough to shuffle and unshuffle things in the most efficient way it knows how, but the file system doesn't have the luxury.
If I'm still missing something, please let me know. Anyway, it'd be pretty straightforward to properly test this so maybe I'll do that.