Connector doesn't drop failed connections for a long time (probably, never)

[drewdzzz]

According to the logs of test Client, some time after the failed connection is destroyed Connector still tries to use it. We should investigate it.

Test logs

*** TEST test_sigpipe started... ***
ERROR: Send failed: Connection reset by peer (10)
ERROR: Connection got an error: Failed to send request: Connection reset by peer
*** TEST test_sigpipe: done
*** TEST test_dead_connection_wait started... ***
ERROR: Peer shutdown (10)
ERROR: Connection got an error: Failed to receive response: Operation now in progress
ERROR: Connection got an error: Failed to receive response: Operation now in progress
ERROR: Connection got an error: Failed to receive response: Operation now in progress
*** TEST test_dead_connection_wait: done
*** TEST response_decoding started... ***
   *** TEST CASE decode data with non-matching format ***
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
ERROR: Recv from dead stream (10)
   *** TEST CASE decode data to rvalue object ***
   *** TEST CASE decode data to object with an mpp tag ***
*** TEST response_decoding: done

[drewdzzz]

Well, manual call of client.close(conn) helps.
However, why couldn't we just close the connection when it is destructed?

UPD: when EpollNetProvider is used, close must not be called on a dead connection. So what's the universal way to destroy a dead connection? That's unusable - we should revise the whole handling of dead connections.