In the LoRaWAN 1.0 specification the duty cycle for joins in EU is indeed 0.1%. So if the initial joins fails for whatever reason, it will take a long time before it sends another attempt.
The LoRaWAN 1.0.1 update has relaxed this limitation to 1% join duty cycle in the first hour.
The reason for missing some join (accepts) is probably because the TTN backend can respond in the RX2 window and then uses data rate SF9. This is not compliant, and I just opened an issue for it. Before the join, the backend should respond with the LoRaWAN default values (=SF12 for RX2), and then it can change from the defaults by setting the new configuration parameters in the join accept message. After the join accept, it can then use the TTN values (=SF9 for RX2).
Also note that LMiC based end-node could miss a join accept because of incorrect RX slot timing, if your code does other stuff than just LMiC.