Please look this position from a my own game which I played by correspondence several years ago:

Taking again this game I'm a little bit amazed to notice that now, like then, all engines are useless, including Rybka: they don't recognize a typical fortress, the evaluation as like as the self-playing don't give clear indications.

Analysis by Rybka 2.2n2 mp 32-bit :
depth=25
1. -+  (-4.47): 57...Da5 58.Td7 Rh6 59.Rg2 Da4 60.Tc7 De4+ 61.Rf1 Db1+ 62.Re2 Da2+
2. -+  (-4.47): 57...Da2 58.Td7 Db1+ 59.Rh2 Da1 60.Rg2 Da4 61.Te7 Da6 62.Rh2 Da2 63.Rg3 Da5 64.Rf3
3. -+  (-4.47): 57...Da6 58.Tc7 Da1+ 59.Rg2 Da4
4. -+  (-4.47): 57...Da1+ 58.Rg2 Da2 59.Td7 h5 60.Tc7 hxg4 61.hxg4 Dd5+ 62.Rg3 De4 63.f3 Df4+
5. -+  (-4.47): 57...Da3 58.Rg2 Da6 59.Td7 Da5 60.Te7 Da1 61.Rh2 Da2 62.Rg3 Da5 63.Rg2 Da1 64.Rh2
6. -+  (-4.47): 57...Dd1+ 58.Rh2 Da1 59.Td7 Da2 60.Rg3 Da3+ 61.Rg2 Da4 62.Te7 Da6 63.Rh2 Da2
7. -+  (-4.47): 57...h6 58.Rh2 h5 59.gxh5+
8. -+  (-4.47): 57...h5 58.gxh5+
9. -+  (-4.47): 57...Rh6 58.Rh2 Da2 59.Rg1 Da1+ 60.Rg2 Rg6 61.Td7 Da4 62.Te7 Da6 63.Rh2 Da2 64.Rg3
10. µ  (-1.39): 57...f5 58.gxf5+

This is only a my example, I don't dare to think all possible real endgames could to present to an ambitious player who entrust on engines.
It's was more useful to me some almost nineteenth-century studies by Nikolay Grigoriev (of 1917, to be precise, particularly the following games from "EndGame Study database 2000").

[+1300.12b4a7]
[+1300.12b4b7]
[+1300.12e5b7]
[+1300.12h1d7]
[+1300.12c5b7]
[+1300.12c5d7]
[+1300.12d4a7]

I discovered that very probably I missed a winning path after 57...h6 with idea f5 fxg5 h5 h4 (I had only few days of time to conclude the game and sure to win I played 57...h5, but after 58. gxh5+ Kxh5 59. Rf7 f5 60. Rxf5 Qxa7 61. Rf8 Qa1+ 62. Kg2 Qe5 63. Rf3 Qd5 64. Kg3 Kg6 65. Re3 Qh1 66. Rc3 Kf5, I agreed the draw)
However, also with a powerful engine I can't to be sure that I could have won: I'm not really a master.....
Vaaaaas....sigh