I have to say - I've been bouncing around the concept for a while - whether going for an AFA is 'worth it'. 
I actually think that Filers ... are less so than some options. 
For writes, between WAFL, RAID-DP and NVRAM .... it's actually not all that often you need fast sustained throughput.

For reads... well, yes, the read latency is always a factor. But it remains a question - how much true random read do you do? 
Between 'stats show -p flexscale-access' and 'priv set diag; stats show wafl:wafl' which will show you where the reads are coming from.

For most of my filer heads, I'm getting pretty high read cache hit rates - 90%+ (between 100G of RAM and 2TB of PAM) leads me to conclude that I'd be wasting my money going all flash. 

So it's really a workload question. If you've a lot of write IO, then more controllers and spindles will be as good. (short stroking if necessary). 

The place where all flash really delivers is for true random-read workloads. If you have one of those you'll be paying the seek latency for most IOs, at which point having flash will pay dividends. You'll be able to see this, because your read cache hit rate will be really low. And your service users are complaining about read latency being too high. 

I've not seen the need yet, but then I don't have any workloads like that. For most, a bit of cache (RAM + PAM) - and maybe for extremes going for hybrid aggregates - gives me the better cost-benefit ratio.