So, with MyISAM I got similar results. The machine is not benchmark-tuned, so not only is it slow, but results can deviate a lot, on different reasons. But the trend still holds:

Same 64 errors in each case on the small tmpdir.

Based on IRC discussion, I ran the tests from a different angle.
The table is of the same structure as above, but MyISAM, to rule out TokuDB specifics.
All runs below are on 2,000,000 rows dataset.
The rows are inserted in two different ways.
INSERT1:
insert t1 select seq, repeat(seq, 1000/length(seq)), date'2001-02-03' + interval seq second, seq % 20, seq, seq % 1000, 60 + seq % 25, seq % 100, repeat(seq, 1000/length(seq)), seq % 1000 from seq_1_to_2000000
INSERT2:
insert into t3 select seq, concat('http://my.url',seq), now(), seq, seq, seq, seq, seq, concat('script',seq), seq from seq_1_to_2000000

So, while INSERTs are similar, they produce essentially different data layouts.

The tests were run on tmpdirs with 3 different free space volume: 16 Gb, 8 Gb, 300 Mb.
On 16 Gb, in both cases ANALYZE passes. At most it takes ~53% of tmpdir space, so it's slightly above 8 Gb.
On 8 Gb and 300 Mb, it fails with the same infamous 64 disk space errors.
Execution time:

I believe the difference on 8 Gb can be ignored, it might well be a random fluctuation. The difference on 300 Mb is noticeable though. – both the horizontal and vertical difference. It is not random, I double-checked.

Motivated by CSC#7897, I've tried to see how much space can ANALYZE use. (So far my understanding was that max. temp dir usage would be rougly equal to size of the table w/o indexes).

Let's try this:

create table t1 (pk int primary key, col1 varchar(100)) charset=utf8;

insert into t1 select A.a+B.a*1000, concat('val-', A.a+B.a*1000) from one_k A, one_k B;

The table has 1M rows. Rows in the VARCHAR(100) column occupy less than
300 bytes (which is typical).

The table on disk:

 -rw-rw---- 1 psergey psergey  44M Aug  1 23:09 t1.ibd

Now,

analyze table t1 persistent for all;

  Breakpoint 4, unique_write_to_file_with_count (key=0x7ffebf2ce0a0 "\005", count=1, unique=0x7ffebf047ef0) at /home/psergey/dev2/10.0/sql/uniques.cc:53

(gdb) print key

  $7 = (uchar *) 0x7ffebf2ce0a0 "\005"

(gdb)  

(gdb) p unique->size

  $8 = 302

It's writing 300 bytes, the unpacked length.

In total, we get:

  Breakpoint 3, Unique::flush (this=0x7ffebf047ef0) at /home/psergey/dev2/10.0/sql/uniques.cc:379

  $46 = 281,734,200

  Breakpoint 3, Unique::flush (this=0x7ffebf047ef0) at /home/psergey/dev2/10.0/sql/uniques.cc:379

  $48 = 297,386,100

...

  Breakpoint 3, Unique::flush (this=0x7ffebf047a30) at /home/psergey/dev2/10.0/sql/uniques.cc:379

  $47 = 4,194,304

300MBit = 286 M, i.e. ANALYZE may require much more space than is occupied by the table (44M)

If we can't quickly fix ANALYZE to take less diskspace, we could have it refuse to run if there is a risk of running out of disk space. The thing is, current way ANALYZE uses diskspace is inefficient, but it is very predictable.

POSIX.1-2001 provides a function to check free disk space, statvfs. TokuDB uses is. It seems, my_XXXX portability layer doesn't have a function to check for free disk space.

The "EITS could eat all tmpdir space" is not fixed, the "hang" is.