not cleanly closing ssl sockets on phased_restart? #1002
Description
We're running apache bench against a running puma server bound to ssl sockets. When we perform a puma phased restart, we get a batch of SSL errors in apache bench each time a worker goes down. With verbose ab flags, a healthy connection includes messages like
SSL/TLS State [connect] SSLv3 flush data
read from 0x7f1e770b74e0 [0x7f1e770aab33] (5 bytes => -1 (0xFFFFFFFFFFFFFFFF))
read from 0x7f1e770b74e0 [0x7f1e770aab33] (5 bytes => 5 (0x5))
0000 - 16 03 03 00 aa .....
Where a failed connection looks like
SSL/TLS State [connect] SSLv3 flush data
read from 0x7f1e770b77a0 [0x7f1e770aab33] (5 bytes => -1 (0xFFFFFFFFFFFFFFFF))
read from 0x7f1e770b77a0 [0x7f1e770aab33] (5 bytes => 0 (0x0))
SSL handshake failed (5).
I believe that this is the reason that our AWS ELB returns 504 errors each time we perform a phased restart. cc @timabdulla, who reported similar symptoms in #957